_id
stringlengths 36
36
| text
stringlengths 200
328k
| label
stringclasses 5
values |
|---|---|---|
0018c021-8f7d-42d4-ad56-a6d9a78cecbf | We train a single shadow model using the methods of Salem et al. [1]}, building upon the methods of Rahman et al. [2]}.
The shadow model used in this work is a VGG inspired model [3]}.
Our shadow model is then used to train the attack model, which accepts a vector of probabilities from a model output and outputs the probability that the image which produced that probability vector was in the training data set for that model.
The shadow model is used to create the dataset which trains the attack model using the process described in section REF .
| m |
33003698-ca22-4d91-b334-115cdc71b060 | The attack model is a 2-layer fully-connected neural network.
Given a 10-dimensional input vector, the network learns to predict the probability that the model which output the vector was also trained on that sample.
The model uses the tanh activation in the hidden layers since the more common ReLU activation had worse attack accuracy.
| m |
00d4a3f4-8784-464b-81cb-ecae23e956f0 | Although it is not the focus of this work, the probabilities for test-set data were spot-checked.
We found that models trained with the dropout-induced spike-and-slab prior had generally less confident predictions on previously unseen data than the models without dropout for both the differentially-private and non-differentially private case.
This is the essential conclusion of prior work, and our experiment here confirms their finding.
<TABLE> | r |
691272e7-053c-4ad1-bb93-529a8c79af06 | We summarize the results of the attack experiment in Table REF .
From the results, we observe that there is a meaningful improvement in defense against the attack by using differential privacy.
In particular, the models trained with DPSGD and not dropout was better at defending against model inference on both datasets, with an attack efficacy rate of 53.75% and 56.74% versus the worst case of dropout and no differential privacy: 60.16% and 64.64%.
There is, however, a cost associated with using differentially private training in terms of test accuracy, as the baseline LeNet-5 model achieved the best test set accuracy on both datasets and the differentially private models were meaningfully worse on CIFAR-10.
| r |
b490a2fe-af9d-4437-9070-a7527b009dff | The effect of dropout can be seen most clearly on the CIFAR-10 dataset, where both with and without differential privacy, the attack was more successful, yielding an attack efficacy rate of 60.16% and 60.01% compared with the vanilla and non-dropout DP models achieving only 54.76% and 53.57% attack efficacy respectivel.
To determine if this was a significant effect, we establish the null hypothesis that models with and without dropout are equally vulnerable to membership inference attacks.
Using an independent sample t-test across the models given the null hypothesis found that there was significant difference between models with and without dropout \(p = 0.015429\) .
<FIGURE> | r |
04916c94-238c-4ed4-81e9-2ff6880409e3 | Figure REF shows the area under the curve plot for the attack efficacy on the CIFAR-10 dataset.
We find that the models with dropout have nearly the same AUC in all cases for CIFAR-10, and the differentially private model without dropout has a marginally better curve than the vanilla model.
| r |
e00f305d-5768-4038-8ec8-631506da2c13 | In Figure REF , we observe that the green line depicting the AUC for the non-differentially-private model with dropout is substantially higher than the other lines, demonstrating the high efficacy that the attack had on that dataset and model combination.
The differentially private models are much closer together for the MNIST dataset than they are on the CIFAR-10 dataset, suggesting that the dataset is a meaningful factor in attack efficacy.
We note that in both Figure REF and Figure REF , the membership inference attack performs better than the baseline 0.5 AUC that random guessing would offer in all cases.
<FIGURE> | r |
27879eee-c5a1-4dc7-99fc-7eb83b1eaf9a | As an informal metric, we consider the ratio of attack efficacy to test set accuracy in Table REF .
We observe that by this metric, the effects of dropout are more apparent, as is the effect of differential privacy.
We address the ramifications of this in Section .
<TABLE> | r |
35535937-e6d4-4b25-8bd0-aa6da7467391 | We find that likely due to the valuable information provided by moderating the confidence of predictions, dropout as implemented in Gal and Ghahramani [1]} also increases the efficacy of membership inference attacks.
Since uncertainty quantification seeks to tell us a priori whether or not a sample is familiar in the way that model inference seeks to derive the information a posteriori, it's reasonable to assume that a model that quantifies aleatoric uncertainty will give high uncertainty for new points, and thus be easier to derive information about points which have already been seen.
This means that there is an implicit tradeoff between improving our ability to avoid membership inference attacks and our ability to quantify model uncertainty.
| d |
aee129d4-daff-49a0-9f6a-68b7ea6abc60 | One non-privacy drawback of using dropout before every layer is the creation of up to \(d\) additional hyperparameters to tune, where \(d\) is the depth of the network.
This can lead to significant training burden when optimizing the level of dropout to use and is likely infeasible for larger networks.
Despite this finding, in cases where the privacy of training set data is not needed, the models trained with dropout had much more reasonable probabilities when making predictions on previously unseen data.
In cases where there is a sensitivity to highly confident predictions, it stands to reason that there is significant value in this approach.
| d |
a9e44ca4-1a0d-4c90-84f8-6f44b836808e | As suggested in Section REF , hyperparameter tuning to find the optimal ratio of attack efficacy to test set accuracy is another fruitful opportunity for future work, as highly accurate models which are also highly resistant to attack is ideal.
This could be done by using a grid search to find the values for the differential privacy mechanism that minimize attack efficacy while also searching for learning rate, early stopping, and dropout or normalization rates that maximize accuracy.
Other model architectures may also have some inherently more or less interesting properties as it relates to privacy.
| d |
dbef5df4-fa19-4e9f-b88d-dce3f458e354 | Our example here dealt only with relatively small and well-studied computer vision datasets, so there is significant opportunity for future work.
As we observed, the dataset itself had an effect on the ability to resist membership inference attacks, with the higher-complexity CIFAR-10 dataset having a better best-case resistance and MNIST having a worse worst-case.
Work on natural language datasets and models such as LSTM and Transformers could provide additional insights into the usefulness of dropout and differential privacy as a defense against membership inference attacks in the general case.
Comparing differential privacy plus dropout against differential privacy plus other Bayesian neural network methods is also a fruitful avenue for follow-up, since the natural question is whether there is something unique about dropout that changes the efficacy of differential privacy against membership inference attacks.
| d |
585a3b03-4826-4523-852d-29fdd01bb27b | Transformers have demonstrated state-of-the-art performance in several tasks, but the larger their size gets, the more difficult it is to use them in resource constrained settings [1]}. Recent works in neural architecture search (NAS) [2]}, [3]}, [4]}, [5]}, [6]}, [7]} have focused on identifying computationally efficient sub-Transformers that are easier to deploy on edge devices. However, existing works on NAS only focus on the subspace of denseTerminologies: (1) Dense architectures refer to fully activated networks for every input. (2) Sparse architectures refer to sparsely activated ones with conditional computation per input. (3) Optimal architectures refer to Pareto-optimal ones with the best trade-off in task performance vs. computational constraint. Transformer architectures, where all the network weights are activated for every input.
| i |
ac39081a-1cbc-4a43-a809-191e9ffcddee | In contrast to the above dense models, sparsely activated ones like the Mixture-of-Experts [1]} perform conditional computation in which only a subset of the weights of the network are activated per input. Selective compute allows us to design neural networks with a large number of model parameters, without significant increase in the computational cost. With increased capacity, these sparse models have demonstrated state-of-the-art performance in natural language tasks such as neural machine translation (NMT)[2]}, [3]}, [4]}.
| i |
8b999492-1e56-4547-9963-fd5a3d3e704e | Incorporating MoE architectures in the search space requires one to make several design choices. (a) Expert placement: Identifying the Transformer layers for introducing expert sub-networks. (b) Number of experts: How many experts to introduce in different layers? (c) Expert FFN size: What should be the feedforward network (FFN) size for each expert? Given the large search space of potential architectures and the exorbitant computational cost of training and evaluating them – existing approaches manually design MoE architectures with a highly-restricted homogeneous space. For instance, they use the same number of experts of the same capacity in different layers and make ad-hoc decisions like introducing experts in every other layer [1]}, [2]}, [3]}, [4]}, [5]} or every four layers [6]}.
| i |
92deea1a-8eca-4084-a80f-95850379cb1c | These design choices are not necessarily optimal. The decoder should be lighter than the encoder for auto-regressive NMT tasks due to the cumulative latency of generating tokens one at a time [1]}, [2]}. This impacts the design choice for the number of decoder layers and the number of experts to use in each. For instance, the loss of capacity with decoder layer reduction can be compensated by adding experts on the remaining ones. On the encoder side, a vanilla placement of the maximum allowable experts in each layer results in increased latency from expert communication and activation, although theoretical FLOPs can remain unaffected. These suggest that the optimal MoE's could be heterogeneous when resources like latency or FLOPs are constrained. In a recent review on sparsely activated models, [3]} note that the optimal hyperparameters depend on application and resource specifications – where a systematic simulation of the compute, memory and communication cost can aid practitioners to quickly determine optimal settings without costly trial-and-error launches. AutoMoE provides such a framework to identify optimal hyper-parameter configurations for sparse models under computational constraints.
| i |
83f9988e-9908-453c-a66c-b3c10052ceba | The above observations are depicted in Table REF , which shows demonstrative examples of manually designed architectures vs. those found by our AutoMoE framework from the search space. We compare these architectures against various computational metrics (e.g., latency, FLOPs, active MoE parameters), architectural configurations and task performance.
<TABLE> | i |
1adb0796-565b-4502-b995-f14cc2e8322e |
We introduce a heterogeneous search space for Transformers consisting of variable number, FFN size and placement of experts in both encoders and decoders; variable number of layers, attention heads and intermediate FFN dimension of standard Transformer modules.
We extend Supernet training to this new search space which combines all possible sparse architectures into a single graph and jointly trains them via weight-sharing, yielding a reduced amortized training cost.
We use an evolutionary algorithm to search for optimal sparse architecture from Supernet with the best possible performance on a downstream task (e.g., BLEU score for NMT tasks) satisfying a user-specified computational constraint.
Experiments on several NMT benchmarks demonstrate that AutoMoE-searched sparse models obtain (i) \(3\times \) FLOPs reduction over manually designed dense Transformers and (ii) \(23\%\) FLOPs reduction over state-of-the-art NAS-generated dense sub-Transformers with comparable BLEU scores.
| i |
cec8a109-dd5d-42ab-910a-281dc1069086 | Sparse expert models:
Mixture-of-expert models have a rich literature in machine learning dating back to the
early 90s [1]}.
Sparsely activated expert models, where only a small subset of experts are active at any given time, have received significant attention with works such as [2]}, Switch Transformers [3]}, GShard [4]}, BASE [5]}, Hash [6]}, GLaM [7]}, Stochastic Experts [8]}, Gating Dropout [9]} and ST-MoE [10]}. Some crucial differences in these works include choice of expert routing function, expert placement technique, stability/performance enhancement techniques and nature of the task (pre-training vs. fine-tuning). Some challenges in building sparse expert models include: (i) lack of diversity in expert design (expert layer selection, number of experts, expert size, etc.), (ii) training instability, (iii) poor out-of-distribution generalization, (iv) cross-task adaptation of pre-trained models, (v) communication bottleneck, (vi) high memory and (vii) expert load balancing issue, to name a few. A comprehensive review of recent sparse expert models can be found at [11]}.
| i |
0a17aa3d-b3d6-4eee-bbef-5d7021b8bfd0 | Expert design: Prior work on designing sparsely activated expert models have largely relied on ad-hoc manual choices in terms of expert layer selection, number of experts and their sizes. Existing approaches mostly use manual design, where they add experts on (i) alternate layers [1]}, [2]}, [3]}, [4]}, [5]}, (ii) every four layers [6]}, or (iii) final few layers [7]}. The resulting sparse models have homogeneous expert layers, i.e., same number of experts of the same size in all expert layers. These choices are generally agnostic to the computational constraints (e.g., latency, memory) of the hardware in which the sparse expert model has to be deployed.
| i |
64f7bc66-5656-4388-94fd-ba219103c5e4 | Switch Transformers: The variant of sparse expert model used in this work is Switch Transformers [1]}, mainly chosen due to its popularity and high performance. Switch Transformers replace every Feed-Forward Network (FFN) layer with an expert layer consisting of a collection of experts (independent FFN networks). Each expert layer is preceded by a parameterized routing network that is trained to route each token to top-\(k\) experts in the expert layer. In this work, we adapt Switch Transformers to an encoder-decoder model which is trained from scratch on the machine translation task using top-1 routing.
| i |
41c189fe-44fb-4c86-ab55-f5cdba612751 | Neural Architecture Search (NAS): Given a search space of architectures and efficiency constraints (e.g., model size, latency), NAS typically aims to identify the optimal architecture that maximizes the task performance, while satisfying the efficiency constraints. The main challenges in building a NAS framework include: (i) constructing a search space that covers diverse architectures for the task, (ii) building a fast performance predictor and low latency estimator for candidate architecture evaluation, and (iii) designing a search algorithm to find Pareto-optimal architectures. NAS has been recently used for natural language understanding tasks to build efficient BERT [1]} and GPT [2]} based pre-trained language models [3]}, [4]}, [5]}, [6]}, [7]}, [8]}, [9]}, [10]} as well as for machine translation tasks [11]}, [12]}. Hardware aware transformers (HAT) [12]} is a state-of-the-art NAS framework for machine translation that uses hardware latency as feedback for optimization.
| i |
9c9735ac-3602-41ff-8ea7-11ac20e85783 | However, all of the above NAS works consider a search space with densely activated Transformer models; and primarily search over typical Transformer architectural hyper-parameters like number of layers, attention heads and hidden size. In contrast, we propose the first NAS framework that considers sparsely activated Transformer models like the Mixture-of-Experts; and subsumes all prior densely activated Transformer models as a special case (i.e. one expert per layer). This allows us to increase the search space and make it more diverse by considering heterogeneous architectures; as well as organically address some longstanding design choices for MoE's like how many experts? which layers to place the experts? what should be the expert size? and so on.
| i |
7b9345f5-7e77-4a7d-91f4-372291985b00 | We evaluate AutoMoE on standard machine translation benchmarks: WMT'14 En-De, WMT'14 En-Fr and WMT'19 En-De with dataset statistics in Table REF . We use pre-processed datasets and evaluation setup from [1]}. We report BLEU score [2]} as a performance metric with beam of size 5 and a length penalty of \(0.6\) (for WMT).
| m |
ce3981f4-9a5f-459d-8166-eae776ecc9c9 | Baselines and AutoMoE variations.
We compare AutoMoE against both manually designed and NAS-searched architectures. For the manual baseline, we consider: (a) densely activated Transformers [1]} with no experts; (b) sparsely activated MoE with homogeneous experts (i.e. same number and FFN size) placed in every other layer [2]}, [3]}, [4]}, [5]}, [6]}.
For the NAS baselines, we consider (c) HAT [7]}, which is a Supernet-based state-of-the-art NAS framework for identifying efficient densely activated sub-Transformers for neural machine translation (same task setting as ours); and (d) Evolved Transformer [8]} which is one of the earlier works on finding efficient dense sub-Transformers with evolution-based architecture search.
| m |
6220627f-c0a8-47ca-afef-ec908b75ed36 | Training configurations. All the baselines and AutoMoE variants including the Supernet and final model are trained with the same setting for fair comparison, following HAT. All the models are trained for \(40K\) steps, with a warmup of \(10K\) steps from \(10^{-7}\) to \(10^{-3}\) and use cosine annealing to \(10^{-7}\) for the rest of the steps. All models are trained using fairseq toolkit [1]} with an effective batch size of \(524K\) tokens on 16 V100 GPUs.
| m |
0644edb7-cefd-43ff-bea7-b087e6c48b8a | Evolutionary search setup.
For performance estimation, we monitor the validation loss of the subnet on the NMT task.
We compute latency by measuring the time taken to perform translation from a source sentence to a target sentence with same desired input/output length (30 for WMT) and original beam settings (see Section ) on the target device (NVIDIA V100 GPU). We measure latency 300 times for gold (to report final metrics) and 100 times for partially gold (during evolutionary search) respectively; discard the top and bottom 10% (outlier latency) and compute mean of the rest. Hyper-parameter settings for evolutionary search include: 15 as iterations, 125 as population size, 25 as parents' size, 50 as mutation population size with mutation probability of \(0.3\) and 50 as crossover population size. Unless otherwise stated, the latency constraint for all the experiments is set to \(200ms\) .
| m |
7996b27d-b890-41dd-9477-9c914c401ad5 | Decoder layers vs. FLOPs.
Figure REF (a) shows the average FLOPs for all AutoMoE variants with different decoder layers as obtained during our search (varying from 3 to 6) and baseline models. Notice that the FLOPs increase with increase in decoder layers, given the auto-regressive nature of NMT tasks which require generating tokens sequentially. In contrast to manually designed Transformers with 6 decoder layers (both dense and sparsely activated MoE variants), AutoMoE- and HAT-searched architectures reduce the number of decoder layers with a resulting decrease in both FLOPs and latency. This is also evident in Figure REF (e) which shows that decoder latency dominates the total inference latency for all the models by more than \(90\%\) .
<FIGURE> | m |
ebb68890-af43-4f43-b354-4d62265ed828 | Expert distribution in encoder vs. decoder.
Figure REF (b) plots the number of encoder experts as ratio of total experts for AutoMoE-generated sub-Transformers. We observe that AutoMoE assigns a significant number of experts to the encoder as compared to the decoder. As a result, encoders have much higher capacity (i.e., encoder parameters as a proportion of overall parameters) than decoders. This correlates with the earlier observation that models with higher encoder layers compared to decoder layers enjoy better latency-performance trade-off [1]}. Our findings from AutoMoE designed architectures indicate that the number of layers and experts are two knobs that jointly help in modulating encoder capacity and decoder latency to design efficient models.
| m |
b244128c-1a24-4f3c-be20-316616e78c44 | Expert distribution in different layers.
Figures REF (c) and (d) show the percentage of experts allocated to different layers for encoders and decoders – averaged over several sampled architectures. Notice that the middle encoder layers (\(3^{rd}, 5^{th}\) ) are allocated the maximum number of experts, while the first layer receives the least. The trend reverses for decoder, with the first layer receiving most experts with gradual reduction in expert allocation. This is also consistent with keeping decoders light by dropping layers to reduce latency; while compensating for the reduced capacity with increased experts in the first few layers.
| m |
9401db06-d494-4bba-b999-26290c11a230 | Latency vs. FLOPs as constraint for search.
Table REF presents the impact of latency and FLOPs as computational constraints on the performance-efficiency trade-off. Constraining FLOPs results in models that fully exhaust the FLOPs budget for 3 GFLOPs and 4 GFLOPs; while leading to higher latency. On the other hand, constraining the latency tends to underutilize the budget and leads to relatively superior FLOPs and latency, thereby providing a stricter control.
<TABLE> | m |
c9dc0ce7-2ac3-4386-93a5-9c3f12efaa3d | Pareto-optimal AutoMoE designed architectures. Table REF in Appendix shows the sparsely activated expert architectures designed by two variants of AutoMoE (`std-expert': expert FFN size same in each layer and variable across; `fract-expert': fully heterogeneous expert size) for different datasets with the best trade-off in BLEU vs. latency. On aggregate \(69\%\) of the experts are allocated to the encoder compared to the decoder. Meanwhile, \(70\%\) of the expert layers in `fract-expert' architectures have 2 or more experts, out of which more than \(75\%\) of the expert layers have varying capacities (i.e., experts with different FFN intermediate size).
<TABLE> | m |
cede9ff2-8f97-43d4-bb4e-13f2913ba126 | Search space variations.
Table REF presents the impact of different search space choices on the efficiency and performance trade-off. The first variation is to make `#Encoder Layers' an elastic search dimension. Note that both HAT and AutoMoE consider the number of encoder layers to be fixed (refer to Table REF ). We observe that varying encoder layers degrades the model efficiency in terms of latency and FLOPs, re-inforcing our prior observations on the importance of encoder capacity and depth.
| m |
1bede450-9d9a-4505-8945-9ed5277fda80 | In the second variation (see second major row), we fix the expert architecture (with 2 experts manually placed uniformly) in the search space and only search for the standard Transformer hyper-parameters. Observe that AutoMoE-designed experts have better FLOPs than such manually designed ones.
| m |
ffe57388-fccd-4e35-9216-16f0e5b10a05 | The last variation introduces identity or dummy experts (i.e., expert with 0 intermediate FFN size, equivalent to identity operation). This explores the idea that we can skip the computation for some of the tokens based on context rather than always forcing them through an FFN. We observe identity experts to marginally hurt the performance but significantly reduce FLOPs (see last major row).
| m |
142a3c88-beec-44a2-8ce1-ca3881d0a6b6 | AutoMoE is the first framework to explore the space of sparsely activated Mixture-of-Experts (MoE) models for neural architecture search (NAS). AutoMoE identifies efficient sparsely-activated sub-Transformers with reduction in FLOPs and latency over both manually designed and NAS-searched architectures with parity in BLEU score on benchmark machine translation (MT) tasks. AutoMoE explores a fully heterogeneous search space with variable number of experts, their size and placement choices in different layers for encoders and decoders; alongside other standard Transformer architectural hyper-parameters.
| d |
bb65ebde-657b-4b2f-854e-7e585b09a161 | Given our focus on finding efficient MoE models under computational constraints, AutoMoE search space and evaluation has been restricted in scale to big-sized Transformer models for benchmark MT tasks. A natural extension of this work is to explore the limits of MoE models like SwitchTransformers [1]} and GShard [2]} that are significantly larger containing billions to trillions of parameters; as well as designing sparse and transferable efficient expert models [3]} for diverse types of tasks like reasoning, summarization and understanding.
| d |
371e352a-00fb-428a-abc5-12c81029f6f3 | There is a jumble of symbols, concepts, systems, and protocols that make engineering and science sound very vague and incomprehensive. The question we raise is: Can we provide effective access and understanding of these concepts and introduce them into an undergraduate curriculum?
| i |
1e72c121-30f5-45b6-8867-22743ac78cc9 | Digital game is one of the most efficient tools to motivate youngsters: they feel easy, keen, confident, and fun in dealing with games, which makes them enjoy thriving to go through challenges. Taking advantage of this, the digital game-based learning (DGBL) is known to make learning more attractive, motivating, and personalized from the learner's viewpoint [1]}[2]}. Some difficulties in carrying out typical forms of inquiry-based learning in a class have been pointed out [3]}-[4]}: to summarize, (i) the complexity in the contexts and (ii) the students' low engagement level. However, when delivered in the form of a digital game-based activity, greater chances of being more joyful and efficient have been found: specifically, (i) self-discovery and construction of one's own reality [5]}-[6]}; (ii) increase in students' inquiry experience [7]}[8]}; (iii) engagement of students in applying knowledge to real-world contexts [9]}-[10]}; and (iv) collaborative learning where two or more learners interact and engage to learn together [11]}-[12]}.
| i |
94abfd1d-6828-4245-9130-c584441733c7 | In further studies, it has been found that the level of students' engagement influences their inquiry learning effectiveness [1]}[2]}-[3]}. It motivates the need for precise evaluation of students' feedback on management of the proposed gamified learning platform (as shall be described in Section ).
| i |
85d3a95a-9afb-4eae-8e45-e5c5fb1b963d | We identify blockchain applied in wireless communications as the new content to explore the effectiveness of DGBL in an undergraduate curriculum. The key rationale is that this rapidly evolving field of technology offers great opportunities for students to experience multidisciplinary topics involving elements of mathematics, (i.e., abstract algebra and number theory), cryptography, computer networks, etc. The necessity for systematic comprehension of such an interplay among various, dissimilar areas makes a compelling case in favor of infusing new learning materials and strategies to enhance an undergraduate education program. In fact, the wireless technology is experiencing a variety of completely new problem domains including spectrum sharing [1]} and vehicle-to-everything (V2X) networking [2]}[3]}. Nonetheless, at the blockchain consensus 2017, scholars have concluded that academics are not keeping pace with blockchain change [4]}.
| i |
77106c11-63b2-445f-b476-f01f3ae3cbed | Given the significance of blockchain expertise in new undergraduates, strong educational support is vital and current educational curricula should reflect cutting-edge trends and needs in this sector. Specifically, students are challenged more than ever to be creative to confront contemporary issues related to blockchain. Hence, educational environments should cultivate students that are equipped with a set of tools to formulate, solve, and properly tackle multidisciplinary problems.
| i |
b9433580-83fd-4255-8d7f-e14a6e53f857 | This research aims to (i) build a DGBL platform, (ii) incorporate into an ECE course, and (iii) evaluate the pedagogical efficacy. The project had been motivated from three problem statements, which corresponded to three action plans. As reported in this paper, we have made the progress through A1, while A2 and A3 are left clearly planned. To elaborate our progress in A1, we have built a DGBL framework that is similar to a RPG where multiple players can explore different maps with different levels of wireless connectivity. By doing so, students are expected to learn how the connectivity affects the performance of a wireless network and a blockchain system overlaid on the network. In conclusion, the success of this research will cultivate students that are equipped with skillsets to deal with a wide variety of problems raised in the rapidly evolving ecosystems of blockchain and wireless communications.
| d |
ffa0c5ed-78b9-4a50-b52a-88234d6478e2 | The vast majority of recently developed online courses on Artificial Intelligence (AI), Natural Language Processing (NLP) included, are oriented towards English-speaking audiences. In non-English speaking countries, such courses' audience is unfortunately quite limited, mainly due to the language barrier. Students, who are not fluent in English, find it difficult to cope with language issues and study simultaneously. Thus the students face serious learning difficulties and lack of motivation to complete the online course. While creating new online courses in languages other than English seems redundant and unprofitable, there are multiple reasons to support it. First, students may find it easier to comprehend new concepts and problems in their native language. Secondly, it may be easier to build a strong online learning community if students can express themselves fluently. Finally, and more specifically to NLP, an NLP course aimed at building practical skills should include language-specific tools and applications. Knowing how to use tools for English is essential to understand the core principles of the NLP pipeline. However, it is of little use if the students work on real-life applications in the non-English industry.
| i |
031302c3-95b8-48d1-b6a9-6885031067f0 | In this paper, we present an overview of an online course aimed at Russian-speaking students. This course was developed and run for the first time in 2020, achieving positive feedback. Our course is a part of the HSE university's online specialization on AI and is built upon previous courses in the specialization, which introduced core concepts in calculus, probability theory, and programming in Python. Outside of the specialization, the course can be used for additional training of students majoring in computer science or software engineering and others who fulfill prerequisites.
| i |
e6d21f52-2fa5-43fa-800e-83cd4e52f29c |
We present the syllabus of a recent wide-scope massive open online course on NLP, aimed at a broad audience;
We describe methodological choices made for teaching NLP to non-English speaking students;
In this course, we combine recent deep learning trends with other best practices, such as topic modeling.
| i |
ef5aa6f0-8935-4cd7-af9c-3f2e81c296ac | The remainder of the paper is organized as follows: Section introduces methodological choices made for the course design. Section presents the course structure and topics in more details. Section lists home works. Section describes the hosting platform and its functionality.
| i |
3e3f6ed8-e6fb-428d-8049-e7d8d429233e | This paper introduced and described a new massive open online course on Natural Language Processing targeted at Russian-speaking students. This twelve-week course was designed and recorded during 2020 and launched by the end of the year. In the lectures and practical session, we managed to document a paradigm shift caused by the discovery and widespread use of pre-trained Transformer-based language models. We inherited the best of two worlds, showing how to utilize both static word embeddings in a more traditional machine learning setup and contextualized word embeddings in the most recent fashion. The course's theoretical outcome is understanding and knowing core concepts and problem formulations, while the practical outcome covers knowing how to use tools to process text in Russian and English.
| d |
ef9fd9a2-418e-4abd-92cb-067a260f828a | Early feedback we got from the students is positive. As every week was devoted to a new topic, they did not find it difficult to keep being engaged. The ways we introduce the core problem formulations and showcase different tools to process texts in Russian earned approval. What is more, the presented course is used now as supplementary material in a few off-line educational programs to the best of our knowledge.
| d |
79183dcb-ae79-47c9-ab7b-9aead54c9f6c | Further improvements and adjustments, which could be made for the course, include new home works related to machine translation or mono-lingual sequence-to-sequence tasks and the development of additional materials in written form to support mathematical calculations, avoided in the video lecture for the sake of time.
| d |
2e4b4870-c788-4975-816a-e88d25cb8c41 | Few-Shot Learning (FSL) is to recognize novel visual concepts with several examples based on the acquired knowledge [1]}, [2]}, [3]}. It is also taken as a practice paradigm for meta-learning [4]}, [5]}, which learns the method to transfer old knowledge on novel tasks. However, vanilla FSL methods are limited by that old and novel tasks are from the same domain. Obviously, this setting is far from the demands of real-world applications. Thus, Cross-Domain Few-Shot Learning (CD-FSL) is proposed, in which acquired knowledge and novel tasks come from the source domain and the target domain respectively. Accordingly, the domain shift from the source domain to the target one is the main challenge in CD-FSL.
<FIGURE> | i |
e236d1eb-1c13-4a48-8fdc-37525da59d55 | There are different embedding distributions in the source and target domain, as shown in Fig. 1. The domain shift results in that the knowledge of domain A can not work well on domain B. A typical method to address this problem is aligning the two distributions by adversarial training [1]}, domain alignment [2]} and knowledge distillation [3]}. They usually choose some samples from the target domain as an auxiliary set during training. Obviously, these approaches violate the CD-FSL setting, on which the target domain is absolutely unknown during the training on the source domain. Therefore, FT [4]} explores to realize feature transformation without access to the target domain, which employs affine transforms to simulate various feature distributions under different domains. This transformation is not directed to the target domain, which performs unstably on the different target domains. How to employ limited labeled samples to provide a directive metric space for the target source is a research focus in CD-FSL.
| i |
3fa62b94-efa7-4f32-962a-cb244d4dc95f | We observe that directly measuring the distance of the target and the source domain is difficult by limited labeled target samples. By contrast, the tasks of the same domain are highly related, so that these tasks can be decomposed to guide the metric space. This guidance can be formed by employing the labeled samples and self-labeled samples to construct multiple task-oriented metric spaces, which updates parameters to adapt for the target domain quickly, as shown in Fig.1 (c).
| i |
996dfb13-1352-45b4-9df8-75025098867b | To construct task-oriented and robust metric spaces, it is important to generate diverse and real samples. The current setting of CD-FSL is transferring universal knowledge (miniImageNet) to fine knowledge. Since the target domains are usually fine-grained datasets, the test samples often have similar backgrounds, as shown in Fig. 2. The share of backgrounds could provide task-oriented and abundant metric spaces for task-decomposition. Thus, we introduce Weakly Supervised Object Localization (WSOL) to find the object regions.
<FIGURE> | i |
503e62b2-182e-4635-99e7-ea429c3cbcae | WSOL aims at obtaining the localization of objects only with the image-level labels, which has broad applications in video understanding [1]}, medical imaging [2]} and remote sensing imagery analysis [3]}. Most WSOL methods extract activated regions of feature maps, which is called Class Activation Map (CAM) [4]}. These CAM-based approaches have achieved promising results on WSOL. Considering the capability of coverage, we apply the simplest CAM approach to find the discriminative regions and exchange the regions to realize task expansion.
| i |
576a8510-1c25-441e-b0e2-936ba4e30188 | To this end, we design inner tasks nesting in few-shot tasks (outer tasks), of which inner tasks learn task-oriented metric spaces by iterations. The inner tasks contain 2 stages, i.e., task-expansion and task-decomposition, which are formed by a task-expansion-decomposition framework. We name it Self-Taught (ST) approach. The illustration of ST is shown in Fig. 3. Specifically, in the Task-Expansion Module (TE-Module), weakly supervised object localization and self-supervised technologies are employed to enrich task-related samples by exchanging and rotating the discriminative regions. The TE-Module generates a more abundant task set. Then the Task-Composition Module (TD-Module) samples tasks from the set produced by TE-Module to finish the inner task, which contains the task of few-shot recognition and rotation classification. After several iterations of inner tasks, the parameters of networks are updated to the optimal metric space for the outer task, i.e., the actual few-shot task. Our ST approach is capable of fully utilizing limited labeled samples to construct target-oriented metric spaces by the task-expansion-decomposition framework.
| i |
796b432c-c7ea-40db-a263-677a2e9cfea5 | 1. We propose a task-expansion-decomposition framework, called Self-Taught (ST), for cross-domain few-shot learning, which alleviates the problem of undirected guidance caused by the unknown target domain. Our ST approach can work on many mainstream metric-based FSL methods, which is a plug-and-play framework.
| i |
e0edd823-974b-4cda-8df7-d1133f8032df | 2. We incorporate weakly supervised object localization and self-supervised technologies to construct target-oriented metric spaces, which aims at focusing on discriminative region representations. To the best of our knowledge, this is the first work to introduce weakly supervised object localization to cross-domain few-shot learning.
| i |
62beb074-cdf3-4930-be0d-19aa63de8a82 | We conduct experiments to prove the effectiveness of our proposed ST approach on three popular metric-based methods: MatchingNet [1]}, ProtoNet [2]}, and RelationNet [3]}. Under cross-domain settings, the source domain is miniImageNet [4]}, and the target domains contain 8 fine-grained datasets.
<TABLE><TABLE><TABLE> | m |
c5ac2adf-b0fd-4ebb-b0e9-bc902f7cbcd9 | This paper has introduced a Self-Taught (ST) approach to alleviate the problem of non-directed guidance for target tasks in cross-domain few-shot learning. Our ST approach can nest in various metric-based approaches and is a task-expansion-decomposition framework. It first employs weakly supervised object localization and self-supervision technologies to generate fine-grained level samples. Then they decompose into subtasks to finish the task of few-shot recognition and rotation classification. It not only constructs task-oriented metric spaces but also focuses on the discriminative regions. The experimental results show the impressive effectiveness of our ST approach on eight target domains.
| d |
a280e32e-b466-4172-a31c-60ce49fbb574 | Traditionally, it is desired that a route in a network uses lightly loaded links, to reduce blocking probability.
A usual way to implement this goal is to assign a weight to each link, such that it reflects the traffic load on the link. Having
these weights, we can look for a minimum weight path. The latter can be found by classic methods, such as Dijkstra's algorithm.
| i |
579c748a-722a-4b8e-8c4b-e7fc1cb8f025 | The link weight can be derived from various practical parameters that are related to the load.
Examples are delay, queue length in packet switching networks,
blocking probability or the number of occupied circuits in circuit switched networks, etc.
| i |
b8fb37ce-44ef-48ea-82cc-e106ce0b6547 | While the above approach aims at locally avoiding heavily loaded links, it only takes into account the congestion of individual links
(this is what we refer to as local view). It is not sensitive to congestion in a whole subnetwork, which would be a more globally oriented view.
For example, it does not distinguish a link, which is overloaded only in itself, from another one, which is overloaded together with its whole
neighborhood, being part of a congested subnetwork.
| i |
a21251c6-a0df-4cea-bc36-ad936220ebd3 | Our goal is to study routing strategies for global congestion avoidance, that is, finding routes that avoid
congested subnetworks, not only congested links. Specifically, we address the issues outlined in the next sections.
| i |
3f193783-63da-4c08-8b4d-aa39ecab9df0 | The goal of generic object tracking is to localize an unknown object in a video sequence given
its position in the first frame. The most popular tracking modality is color vision in the terms of RGB image frames as the input. Furthermore, the problem of
visual object tracking (VOT) can be divided into short-term and long-term tracking which are evaluated
differently [1]}, [2]}. The short-term evaluation protocol focuses on
the tracker itself by measuring its accuracy and robustness. If a tracker loses the target, it affects to the robustness metric, but the tracker is re-initialized and then evaluation continues. In the long-term (LT) protocol the tracker is not re-initialized and thus the LT trackers need special procedures for detecting whether the target is present or not and re-detection in the target-not-present operation mode.
The two latest VOT challenges [1]}, [2]} include additional tracks for RGBT (RGB plus Thermal infrared) and RGBD (RGB plus Depth) object tracking. Interestingly, there are no trackers specialized on
thermal features or depth features, but the top performing RGBT and RGBD trackers all use RGB features
learned by the leading deep tracker architectures, MDNet [5]}, ATOM [6]} and DiMP [7]}.
The additional modality T or D is used only as a “sidekick” to help in various special cases
such as occlusion detection. Therefore it remains unclear what are the potential applications of
RGBD and RGBT tracking and whether T and D channels have their own powerful features.
<FIGURE> | i |
c3e9181c-5084-48cd-8a05-6b01a0775334 | In this work we focus on RGBD tracking.
RGBD tracking has been investigated in a number of recent works
(e.g., [1]}, [2]}, [3]}), but these use
conventional RGB and depth features and are inferior to the deep RGB tracker based methods in the VOT 2019 and 2020 RGBD
challenges [4]}, [5]}. The top performing
trackers in VOT 2020 RGBD challenge, ATCAIS, DDiMP and CLGS_D, are based on the
recent deep RGB trackers, ATOM [6]}, DiMP [7]} and MDNet [8]},
and use depth only to detect occlusion, target disappearance and target re-detection. There are no “depth trackers” that are trained with depth sequences. The main reason is the lack of suitable training data. For example, the three existing RGBD datasets, PTB [9]}, STC [10]} and CDTB [11]}, contain only 100+36+80 sequences. The target and attribute annotations are available only for the STC and CDTB datasets leaving only 116 RGBD tracking sequences for research purposes. At the same time, the existing RGB tracking datasets, LaSOT[12]},
TrackingNet [13]} and GOT-10k [14]}, contain 1.55K+30K+10K sequences, i.e. there is more than two orders of magnitude difference in the amount of data. To unveil the power of depth in RGBD tracking, we need larger and more diverse RGBD or depth-only tracking datasets.
| i |
4f503262-772b-43a4-b9cd-e805f429250a | Results and findings from extensive experiments with the SotA RGB and RGBD trackers on DepthTrack.
These findings, including the fusion of RGB and depth features, will facilitate future works on collecting better RGBD tracking datasets and developing better RGBD trackers.
| i |
0abb3848-1f03-4657-a207-35f90e2cb0d9 | A new RGBD baseline, DeT, that is trained with depth tracking data and obtains better RGBD tracking performance than the existing SotA trackers.
DepthTrack RGBD sequences, annotation meta data and evaluation code are made compatible with the VOT 2020 Python Toolkit to make it easy to evaluate existing and new trackers with DepthTrack.
Related Work
RGBD tracking datasets.
There are only three publicly available datasets for
RGBD object tracking:
1) Princeton Tracking Benchmark (PTB) [1]},
2) Spatio-Temporal Consistency dataset (STC) [2]}
and
3) Color and Depth Tracking Benchmark (CDTB) [3]}.
The statistics of these datasets are compared to the proposed DepthTrack in
Fig. REF .
PTB contains 100 RGBD video sequences of rigid and nonrigid objects recorded with Kinect in indoors. The dataset diversity is rather limited in the number of scenes and attributes and approximately 14% of sequences have RGB and D syncing errors and 8% are miss-aligned.
STC addresses the drawbacks of PTB. The dataset is recorded by Asus Xtion RGBD sensor and contains mainly indoor sequences and a small number of
low-light outside sequences. The dataset is smaller than PTB, containing only 36 sequences, but contains annotations of thirteen attributes. STC addresses short-term tracking.
CDTB is the most recent dataset and is utilized in the VOT-RGBD 2019 and
2020 challenges [4]}, [5]}.
It contains 80 sequences and 101,956 frames in total. The sequences are
recorded both indoors and outdoors and the authors use evaluation
protocols from the VOT challenge.
RGBD tracking algorithms.
Until very recently the RGBD trackers were based on
engineered features and used various ad hoc methods to
combine RGB and D [6]}, [7]}, [8]},
for example, DAL[9]} embedded the depth information into RGB deep features through the reformulation of deep discriminative correlation filter.
Furthermore, the best performing trackers in VOT-RGBD2019 and VOT-RGBD2020
challenges are extensions of well-known deep RGB trackers [4]}, [5]}. For example,
the three winning RGBD trackers in VOT-RGBD2020 are ATCAIS, DDiMP and CLGS_D, which are extensions of the deep RGB trackers
ATOM [12]}, DiMP [13]} and MDNet [14]},
respectively. The main tracking cue for these trackers is RGB and
depth is used only for long-term tracking tasks. In this work,
we report results for the first RGBD tracker whose depth branch
is trained with genuine depth data and that provides superior results.
The DepthTrack Dataset
<FIGURE><FIGURE>Tracking sequences.
CDTB dataset [3]} was captured with multiple
active (Kinect-type) and passive (stereo) depth sensors and RGB
cameras, but in the VOT evaluations the acquisition devices had only marginal
effect to the results while the sequence content was the main factor. Therefore
DepthTrack was captured with a single mid-price but high quality RGBD sensor:
Intel Realsense 415.
The RGB and depth frames were stored using the same
\(640 \times 360\) resolution and the frame rate of 30 fps.
RGB images were stored as 24-bit JPEG with low compression rate
and the depth frames as 16-bit PNG.
The Intel sensor provides the RGB and depth frames temporally synchronized.
In our data collection we particularly focused on content
diversity.
The overall properties of DepthTrack are compared to three available datasets in Fig. REF and a more detailed comparison on different target types between DepthTrack
and CDTB is shown in
Fig. REF .
It is notable that there are three object types that dominate the CDTB sequences: “box”, “human” and “robot”. At the same time,
the high diversity of DepthTrack provides unique targets in almost every test sequence.
Another important factor is presence of humans in
CDTB sequences and them moving simple rigid targets objects. To reduce
the “human operator bias” in DepthTrack we included a large number of
objects that are only indirectly manipulated by hand, by a barely visible rope.
DepthTrack is split to 150 training sequences and
50 test sequences so that they do not contain common scenes
and objects. The training set contains 218,201 video frames
and the test set 76,390. As shown in
Fig. REF almost
every test sequence has its own unique target type. Only “basketball”
(2), “soccer” (2) and “pigeon” (3) appear multiple times in the test set.
Data annotation.
Each DepthTrack frame is annotated with the target bounding box location and multiple scene attributes that help to analyze the results.
The axis-aligned bounding box annotation is adopted from the VOT-RGBD challenges
and we followed the VOT annotation protocol and employed the Aibu1 annotation tool.
1https://github.com/votchallenge/aibu
In the protocol the axis-aligned bounding box should tightly fit the target object in each frame to avoid the background pixels.
To allow detailed scene-type level analysis of tracking results we annotated 15 per-frame attributes.
Besides the 13 attributes used in CDTB (VOT-RGBD 2019 and 2020), we introduced two new attributes: background clutter (BC) and camera motion (CM). BC denotes scenes where the target and background share the same color or texture and CM denotes cases where camera movement
leads to substantial target distance (depth) change. BC frames are expected to be difficult for RGB-only tracking and CM frames for D-only tracking.
For the detailed description of each attribute, please refer to our supplementary material.
The total time to annotate the DepthTrack data was more than 150 hours.
Performance metrics.
DepthTrack is a long-term RGBD tracking dataset where the trackers must detect
when targets are not visible and again re-detect them when they become visible
again. Our evaluation protocol is adopted from [3]}.
A single pass of each sequence is made and for each frame of the sequence
trackers must provide a target visibility confidence score and
bounding box coordinates. The bounding box coordinates are
used to evaluate their precision with respect to the ground truth bounding
boxes. Precision is measured by the bounding box overlap ratio. Confidence
scores are used to evaluate whether the tracker is able to recall the frames
where the targets are annotated visible. Ideally confidence is 0 for the
frames without target and 1 for the frames where the target or parts of it
are visible.
The overall evaluation is based on the tracking precision (\(Pr\) ) and recall (\(Re\) ) metrics [4]}.
Tracking precision is the measure of target localization accuracy when the target
is labelled visible.
Tracking recall measures the accuracy of classifying the labelled visible target.
As a single measure, the F-score is used as the harmonic mean of precision and recall to rank
the trackers. The precision and recall are computed for each frame and then
averaged over all \(j=1,\ldots ,N_i\) frames of the \(i\) -th sequence. That provides per
sequence metrics which are then averaged over all \(i=1,\ldots ,N\) sequences
to compute dataset specific metrics. It should be noted, that the tracker
confidence affects to Pr and Re and therefore the precision-recall graphs
are computed by varying the confidence threshold \(\tau \) .
One weakness of the above averaging protocol is that if the video lengths vary substantially,
then the short videos may get unreasonably large weight in the performance metrics. For example, the sequence length varies between 143 to 3816 frames in DepthTrack.
To alleviate this problem, we do averaging over all frames of all sequences (refer as frame-based evaluation), as well as the sequence specific averaging (refer as sequence-based evaluation).
As the final performance metric, we store the highest F-score over all confidence thresholds \(\tau \) and store also the corresponding precision and recall values. For the details of our evaluation metrics, please refer to the supplementary material.
Baseline RGBD Trackers
In this section we introduce the baseline RGBD trackers used
in our experiments. The existing baseline trackers are selected
among the best performing RGB and RGBD trackers in the
recent VOT-RGBD evaluations [4]}, [5]}.
In the spirit of our work, we introduce a new depth data trained baseline in
Section REF .
Existing Baselines
To establish a strong set of baseline trackers we selected the following 23 trackers:
7 winning submissions to the VOT-RGBD 2019 and 2020 challenges [4]}, [5]}:
ATCAIS, DDiMP, CLGS_D, SiamDW_D, LTDSEd, Siam_LTD and SiamM_Ds;
3 RGBD baselines using hand-crafted RGB and depth features:
DS-KCF [22]}, DS-KCF-Shape [23]} and CA3dMS [6]};
A recent RGBD tracker that embeds D cue into the RGB channels: DAL [9]};
3 winning submissions to the RGB VOT-LT (long-term) 2020 challenge:
RLT_DiMP [26]}, LTMU_B, Megatrack;
3 winning submissions to the VOT-ST (short-term) 2020 challenge:
RPT [27]}, OceanPlus and AlphaRef [28]};
6 deep RGB trackers used by the best RGBD trackers:
SiamFC [29]}, SiamRPN [30]}, ATOM [12]}, DiMP50 [13]}, D3S [33]} and PrDiMP50 [34]}.
All trackers use the code of their original authors and
their default parameter settings. Note that the original authors have
tuned the parameters for the three existing datasets
STC, PTB and CDTB. In particular, the RGBD trackers submitted to
VOT are optimized for CDTB.
<FIGURE>A New Baseline - DeT
We propose a new RGBD baseline for the DepthTrack
dataset - DepthTrack Tracker (DeT) - that obtains the best results thanks to deep depth features learned from depth data.
The tracker architecture (Fig. REF ) is inspired by the online structures of the recent SotA
trackers ATOM and DiMP.
Therefore, our baseline is actually the depth feature extractor and a feature pooling layer and can be combined with either ATOM or DiMP tracker head that performs the actual tracking.
The main difference of the DeT and the original ATOM or DiMP
is that DeT extracts more powerful depth features learned from depth tracking data.
Generating depth tracking data.
The main problem in deep RGBD tracking is the lack
of depth data as the existing datasets provide only
116 sequences with RGB and D images (note that the PTB
annotations are not public). Therefore, we developed
a simple procedure to generate depth data for the
existing large RGB tracking datasets.
We employed the monocular depth estimation algorithm DenseDepth [35]} on the existing RGB benchmarks LaSOT [36]} and COCO [37]}, and manually
selected the best 646 sequences from LaSOT
and 82K images from COCO.
These estimated depth images are used to pre-train DeT
from scratch.
RGBD features for tracking.
The generated depth tracking sequences
are used to train the DeT tracker from scratch
and using similar offline training as used for
ATOM and DiMP. Similar to RGB data in the original works, the training takes 50 epochs after which the training error does not improve anymore. After that the RGB trained color features and D(epth) trained depth features
are extracted from the two feature paths as
\(\left\lbrace D_{RGB},D_{D}\right\rbrace = \left\lbrace F_{rgb}(I_{RGB}),F_{depth}(I_{D})\right\rbrace \)
and are computed separately for the reference and test branches
of the master tracker: \(D^{ref}\) and \(D^{test}\) .
Since the master trackers, ATOM and DiMP, require a single feature
images the RGB and D channels need to be merged to
\(D^{ref}_{RGBD}\) and \(D^{test}_{RGBD}\) .
We want to learn the RGB and D merge from training data and
thus adopt a pooling layer as the standard solution.
In our preliminary experiments,
we compared a number of the typical pooling operations and found the element-wise maximum operation to perform the best.
After the pooling operation,
the two outputs of the DeT feature extraction part in
Fig. REF represent the
reference (the previous frame with an estimated bounding box)
and test (the current frame) branches in the DiMP and ATOM terminology
and the tracking head
depends on whether ATOM,
DiMP or another similar online tracking part is used.
Experiments
All experiments were run on the same computer with an Intel i7- CPU @ 3.40GHZ, 16 GB RAM, and one NVIDIA GeForce RTX 3090 GPU.
In the experiments, we adopt the DiMP50 as the master tracker and employ the element-wise maximum operation as the feature pooling layer, and we refer it as DeT for short.
DeT variants adopt the same backbones as the master trackers, ResNet50 for DiMP50 or ResNet18 for ATOM.
<TABLE>Performance Evaluation
Quantitative results.
The performance metrics and method rankings are shown in Table REF .
According to the VOT protocol the best tracker is selected based on its F-score, but precision and recall provide more details about its
performance.
Fig. REF shows the Precision-Recall and F-score plots for the proposed depth-trained DeT tracker introduced in Section REF and the best 10 SotA and baseline trackers in Section REF ranked by their F-scores
and the RGB master tracker DiMP50 [13]}.
Plots for all test trackers are in the supplementary material.
The results provide the following important findings:
1) the VOT-RGBD2020 (using CDTB) winners DDiMP, ATCAIS and CLGS_D obtain the
best results also with the DepthTrack test data thus verifying their
SotA performance in RGBD tracking;
2) as expected the long-term trackers, both RGB and RGBD, obtain
better F-scores than the short-term trackers;
3) the SotA performance numbers on DepthTrack are substantially lower
than with the CDTB dataset of VOT, for example, the VOT winner
ATCAIS obtains 0.702 with CDTB but only 0.476/0.483
with DepthTrack;
4) the new proposed baseline, DeT, which is trained with
generated depth data, wins on the both evaluation protocols and
obtains +12% better F-score than the second best. At the same time the DeT
trackers has no long-term tracking capabilities and runs substantially faster than
the SotA RGBD trackers.
Our results verify that deep learning-based trackers, such as the proposed
DeT, provide
better accuracy than RGB trackers or ad hoc RGBD trackers on RGBD datasets,
but they need to be pre-trained with real RGBD data (D is generated in our experiments)
and fine-tuned using dataset specific training data.
Qualitative results.
Fig. REF shows example tracks of
certain representative trackers and the proposed depth-trained
DeT tracker. With similar target and background color the
RGBD trackers perform clearly better than RGB trackers. In
addition, the depth cue helps to track the correct object in
the presence of multiple similar objects. On the other
hand, the depth-based trackers
have difficulties to track objects that rotate or move fast
in the depth direction. This is partly due to the reason that
the existing RGBD trackers extract features learned from RGB
data rather than from the both RGB and D channels. The problem can
be compensated by learning a “depth backbone” from depth
data as was done for the DeT which succeeded
on all the example sequences.
Attribute-based performance analysis.
Attribute specific F-scores for the best 10 tested trackers and the proposed DeT and the DiMP50 are shown in Fig. REF .
The F-scores of the best VOT-RGBD 2020 tracker, ATCAIS, are clearly lower for the DepthTrack sequences. Interestingly,
DeT wins in 11 of them,
DDiMP (VOT-RGBD 2020 challenge second) in 2 and
the remaining two are RLT_DiMP and ATCAIS (VOT RGBD 2020 winner).
It is worth noting that the long-term DDiMP handles target loss occasions, partial occlusion and out of frame. While our DeT outperforms other trackers on most depth-related attributes, dark scene and depth change.
Clearly the superior performance of RGBD over RGB is evident and using D training data with DeT makes it the most successful in long-term tracking evaluation even though it is a short-term tracker.
<FIGURE>
Computation times.
The tracking speeds are reported in Table REF .
The most important result of the speed comparison is that the
second best RGBD trackers, DDiMP and ATCAIS, only achieve the speed of 4.77 fps and 1.32 fps due to ad hoc depth processing.
On the other hand, the winning DeT tracker that is a
straightforward deep RGBD tracker architecture runs real-time.
<TABLE><TABLE>
DeT Ablation Study
Master trackers.
As discussed in Sect. REF the DeT can be attached to master trackers that expect such feature extraction pipelines.
The most straightforward “masters” for DeT are ATOM [12]} and DiMP50 [13]} which use the ResNet-18 and ResNet-50 RGB backbones in their original implementation.
We compared the original RGB trackers to their DeT variants using the
depth-trained depth pathway for the D channel.
The results are in Table REF that shows a clear
7.7% improvement to original ATOM and 10.5% to DiMP50 in terms of F-score.
Depth cue.
To further verify that superior performance of DeT is due to better
depth features learned by training with depth data, we compared
the RGB trained depth-only DiMP DiMP50(D) and the depth trained DeT taking only depth input DeT-DiMP50(D).
The results are in Table REF .
DeT-DiMP50(D) is clearly superior to the DiMP50(D) and rather surprisingly the depth-only DeT-DiMP50(D) is almost on par with the highly optimized original RGB DiMP50.
RGB and D feature pooling.
In Sect. REF we selected the max pooling as the
RGB and D feature merging operation. In Table REF the results
are shown for the three different pooling layers: 1) the convolutional layer (“-MC”), 2) element-wise maximum (“-Max”) and 3) mean (“-Mean”).
All the pooling layers performed well as compared to using RGB features
for the D channel. There is only a small difference between
mean and max pooling, but we selected max pooling since DiMP performed
better than ATOM and was used in the method comparison experiment.
<FIGURE><FIGURE><TABLE><TABLE>
Cross-dataset evaluation.
In order to verify that our findings from the DepthTrack experiments
are valid we compared DiMP and DeT architectures on
the CDTB dataset [3]} of VOT-RGBD 2019 and 2020 and
without using any CDTB data in training (cross-dataset).
The results are shown
in Table REF and they verify all findings except that
mean pooling performed better with the CDTB data.
Conclusion
In this work, we introduced a new dataset for RGBD tracking.
To the authors' best knowledge, the proposed DepthTrack is the largest
and most diverse RGBD benchmark and the first to provide a
separate training set for deep RGBD trackers. Our work is
justified by the lack of public datasets for RGBD tracking despite that
RGBD sensors have been available in consumer electronics for many years and
are popular in robotics. RGB tracking has dominated
research in the field and the power of the depth cue for tracking
has remained unknown. In this work, we trained the first RGBD tracker
fully with RGBD data so that the depth pathway was trained
with depth maps from scratch. The proposed DeT tracker was pre-trained with
generated RGBD sequences and then fine-tuned with the DepthTrack
training set. In all
experiments the DeT obtained the best F-score indicating that
depth plus RGB data trained trackers can finally unveil the
power of the depth cue for RGBD tracking.
| i |
6239bb4e-bfb4-4d6b-9541-dbe400471f51 | All experiments were run on the same computer with an Intel i7- CPU @ 3.40GHZ, 16 GB RAM, and one NVIDIA GeForce RTX 3090 GPU.
In the experiments, we adopt the DiMP50 as the master tracker and employ the element-wise maximum operation as the feature pooling layer, and we refer it as DeT for short.
DeT variants adopt the same backbones as the master trackers, ResNet50 for DiMP50 or ResNet18 for ATOM.
<TABLE> | m |
ed9806e6-d9aa-43a8-b3c7-bf4c01c1a852 | In this work, we introduced a new dataset for RGBD tracking.
To the authors' best knowledge, the proposed DepthTrack is the largest
and most diverse RGBD benchmark and the first to provide a
separate training set for deep RGBD trackers. Our work is
justified by the lack of public datasets for RGBD tracking despite that
RGBD sensors have been available in consumer electronics for many years and
are popular in robotics. RGB tracking has dominated
research in the field and the power of the depth cue for tracking
has remained unknown. In this work, we trained the first RGBD tracker
fully with RGBD data so that the depth pathway was trained
with depth maps from scratch. The proposed DeT tracker was pre-trained with
generated RGBD sequences and then fine-tuned with the DepthTrack
training set. In all
experiments the DeT obtained the best F-score indicating that
depth plus RGB data trained trackers can finally unveil the
power of the depth cue for RGBD tracking.
| d |
a8d8b3b8-15af-4a4e-9d9a-8f1858b3baa0 | In fashion research, the advantages of digital fashion archives have continued to attract attention not only for business improvement but also for understanding our societies from a cultural perspective. In particular, the fashion style adopted in daily life is an important aspect of culture. As noted by [1]}, fashion is 'a reflection of a society's goals and aspirations'; people choose fashion styles within the social contexts in which they are embedded. The analysis of daily-life fashion trends can provide an in-depth understanding of our societies and cultures.
| i |
dbae0891-161f-47d5-b90c-1f411234a48a | Kroeber [1]} analyzed fashion trends manually and measured features of women's full evening toilette, e.g. skirt length and skirt width, which were collected from fashion magazines from 1844 to 1919. After this work, many researchers conducted similar studies [2]}. Belleau [3]} examined the skirt length, waist emphasis, and silhouette of women's day dresses from 1860 to 1980. Lowe and Lowe [4]} measured features of women's formal evening dress, e.g. skirt length and skirt width, from 1789 to 1980. Robenstine and Kelley [5]} examined silhouettes of male and female clothes in portraits and fashion illustrations between 1715 and 1914. Their work surveyed old fashion magazines, portraits, and illustrations and focused mainly on formal fashion, rather than daily-life fashion.
| i |
93304b97-db93-441b-8535-f08612e5b8b3 | The analysis of daily-life fashion trends has two major issues. The first issue is the absence of an appropriate daily-fashion image archive. Many researchers, museums, and research institutions have created their own fashion image databases. However, the fashion databases proposed in previous studies have several limitations in terms of their approach to daily-life fashion trends; they cannot be used to analyze how the daily-life clothing styles of people change and what drives these changes.
| i |
57fd41ef-7460-415c-8737-53522c16ebf6 | The second issue is that analyzing fashion trends is labor-intensive. Kroeber [1]} is one of the representative works in the early stage of quantitative analysis of fashion trends. This type of quantitative analysis needs a large amount of human resources to select appropriate images, classify images, and measure features with a ruler. This labor-intensive approach has been applied in this field for a long time. Furthermore, manually managing large modern fashion image archives is cumbersome. To solve this issue, we utilize machine learning (ML). ML is a computer algorithm that learns procedures based on sample data, e.g. human task results, and imitates the procedures. In recent years, ML contributed to the development of digital humanities. For instance, Kestemont et al. [2]} applied ML to the lemmatization of Middle Dutch datasets. Wevers and Smits [3]} demonstrated that ML can separate photographs from illustrations and classify the data based on visually similar advertisements in an archive of digitized Dutch newspapers. Lang and Ommer [4]} used ML to retrieve individual hand gestures from illustrations in Codex Manesse, poetry written in Middle High German in the fourteenth century. In the fashion field, modern fashion styles are complex and diverse. Applying ML methodologies to fashion image archives helps us quantify fashion trends more precisely and efficiently.
| i |
f9ba16e0-9516-4c26-a61e-d9eb6189b877 | In this study, we shed light on research questions about daily-life fashion trends using multiple digital archives and deep learning (DL), which is one of the mainstream methods in ML. To quantify daily-life fashion trends, we built CAT STREET, a new fashion image archive consisting of long-term fashion images that reflect what women wore in their daily lives from 1970 to 2017 in Tokyo with street location annotations. We corroborated the rules-of-thumb of two fashion trend phenomena, namely how economic conditions relate to fashion style share in a long time span and how fashion styles emerge in the street. Through the empirical analyses of these phenomena, we demonstrate that our database and approach have the potential to promote our understanding of societies and cultures.
| i |
62912685-0b6b-4455-a18a-3748bdf01bac | In previous studies, researchers pointed out that social contexts influence people's daily-life fashion. To our knowledge, our study is the first to discuss how much social contexts affect what people wear in a quantitative manner. Our findings in Section 5.1 suggest that our approach with CAT STREET can validate other fashion phenomena observed and discussed qualitatively by providing objective indices about daily-life fashion.
| d |
01a4438c-7396-4162-9009-a032327df0b9 | We explored the potential of our approach with CAT STREET to prompt research questions that can contribute to the expansion of fashion theories in Section 5.2. The first case in Section 5.2.1 prompted a new research question of determining the importance of the different roles that media play in fashion trends: the role of the 'mixed' pattern for creating and reporting new trends and the 'follow-up' effect to keep their momentum. To theorize how fashion trends are generated in more detail, our findings indicate that quantifying the trends from multiple digital archives is essential to test the research questions about the effect of the media's role, which has not been discussed thoroughly.
| d |
fe620e38-28c2-4f03-865f-106994e8a2cc | In previous studies, researchers analyzed fashion style trends at a street level independently. However, to analyze the recent, more complicated trend patterns, the second case in Section 5.2.2 suggests that it would be thought-provoking to focus on what the icons represent for people's social identities, rather than the style itself, and how they lead the trends. An example is the analysis of the types of media that icons use and the reach of that media. The viewpoints gained from our findings can be useful in tackling unsolved research questions such as how the styles interact with each other and how new styles are generated from this process. CAT STREET comprises images indicating what people wear in daily life. If we can determine people's social identities from “what they choose the fashion for” using the archive, CAT STREET can provide beneficial information to current research questions and play an important role in expanding fashion trend theories.
| d |
2613cd7d-fa18-4eb9-9b40-72b6a1ec59df | In this paper, we shed light on research questions about daily-life fashion trends with multiple digital archives using computer science methodologies. To quantify daily-life fashion trends, we built CAT STREET, which comprises fashion images illustrating what people wore in their daily lives in the period 1970–-2017, along with street-level geographical information. With DL, we demonstrated that the rules-of-thumb for the fashion trend phenomena, namely how economic conditions affect fashion style shares, how fashion styles emerge in the street, and the relationship between fashion styles and magazine trends, can be quantitatively validated using CAT STREET. Through empirical analyses to corroborate the rules-of-thumb, we demonstrated the potential of our database and approach to promote the understanding of our societies and cultures.
| d |
df229aec-37ba-489a-abb7-444d36e48925 | Our work is not without limitations. We used the fashion style categories of FashionStyle14 [1]}, which was defined by fashion experts and is considered to represent modern fashion. However, the definition does not cover all contemporary fashion styles and their substyles in a mutually exclusive and collectively exhaustive manner. Defining fashion styles is a complicated task because some fashion styles emerge from consumers, and suppliers define others. We must refine the definition of fashion styles to capture daily-life fashion trends more accurately.
| d |
a9fa8f9f-e3a4-495d-b6fc-93e15618fe92 | Furthermore, prior to building CAT STREET, only printed photos were available for the period 1970–2009. Consequently, the numbers of images for these decades are not equally distributed because only those images from printed photos that have already undergone digitization are currently present in the database. The remainder of the printed photos will be digitized and their corresponding images added to the database in future work.
| d |
9ae86c91-b8ff-408b-afb0-f455c834d433 | CAT STREET and our approach that combined multiple digital archives and methodologies in computer science to quantify fashion trends helped us explore undemonstrated research questions. For future studies that apply other quantitative analytical methods, such as unsupervised clustering, to extract fashion styles and social identities embedded in consumers' daily lives, CAT STREET will play a key role to find new standpoints and expand the boundaries of fashion studies.
| d |
cfc4a68a-d146-4fd7-ae9d-a91ecc3f046c | Two of the most prevalent tools used today by software engineers are repositories to store project files (e.g., git) and bug trackers to report and monitor bug fixing activity (e.g., JIRA, BugZilla). Automatically linking a bug report in a bug tracker and related software artifacts from a repository is one of the long-standing goals in the software engineering research community, due to its potential to improve practice by reducing the time developers spend examining code when addressing a newly reported bug, i.e., bug localization [1]}, [2]}. However, despite numerous efforts, the accuracy of bug localization approaches is not yet high enough for widespread use, especially as it applies to different software projects that vary in bug report and code style [3]}. In examining the trends from interviews conducted with a large cohort of software developers from industry and open-source software, Zou et al. report that developers do not trust bug localization tools due to their inability to adapt to different types of bug reports, specifically noting that existing techniques only work on the most simple cases, with straightforward textual similarity between the bug report and code base [4]}. More work is needed to improve the retrieval quality of bug localization techniques.
| i |
4c6a645f-4d6e-4965-9e33-8a855314931d | At the same time, as industry is increasingly attempting to use bug localization to aid developers in their daily work, specific requirements of the problem for modern use are coming to the forefront [1]}. One key characteristic found beneficial in modern software projects is bug-inducing changeset- (or commit-) level retrieval. A bug-inducing changeset is one where the bug was initially introduced into the repository. Retrieving such changesets leads to faster bug repair, as they contain related parts of the code that were changed together, which makes fixing the bug easier. However, retrieving bug-inducing changesets with high accuracy is more challenging than retrieving buggy source code elements due to the potentially large number of commits in the corpus.
| i |
b0a8856c-3c2a-427f-8fbf-dbece1787daf | In recent years, numerous popular natural language processing tasks (e.g., question answering, machine translation) have all observed improved performance when using neural network architectures based on transformers.
These transformer-based models are typically applied via transfer learning, by first pre-training them on a very large corpus and then fine tuning on a much smaller dataset towards the specific task they are to be used for.
Transformer-based models pre-trained on large software engineering corpora (e.g., StackOverflow) are now becoming available [1]}, with the potential to improve software engineering tasks like bug localization. In this paper, we use the BERT (Bidirectional Encoder Representations from
Transformers) transformer-based architecture, which is a highly popular model introduced by Devlin et al. [2]}.
| i |
38804307-92ef-4c10-8d41-3c42afa06939 | Bug localization is usually framed as an Information Retrieval (IR) task, where a document (i.e., a software artifact) is retrieved from a corpus-based on a query (i.e., the bug report text).
A measure of semantic relatedness between the bug report and the software artifact is necessary to rank the results retrieved from the corpus. Given the fact that transformer-based models consist of many neural layers and require heavy computation for each sentence, measuring relatedness between the query and the corpus quickly becomes expensive.
| i |
75a30603-252d-42d7-a7d7-ef3fdce0e07a | This paper applies BERT to the problem of changeset-based bug localization with the goal of improved retrieval quality, especially on bug reports where straightforward textual similarity would not suffice. We describe an architecture for IR that leverages BERT without compromising retrieval speed and response time. In addition, we examine a number of design decisions that can be beneficial in leveraging BERT-like models for bug localization, including how best to encode changesets and their unique structure.
| i |
50b1ae99-0960-412d-b24c-4a5e3a2ed35b | Our experimental results indicate that the proposed approach improves upon popular bug localization techniques by, e.g., increasing the retrieval accuracy between 5.5% and 20.6% for bug reports with no or a limited number of localization hints. We note that using entire changesets as input granularity significantly hinders the models performance, while leveraging more fine grained input data, such as hunks, results in the highest retrieval quality. We also observe that the size of search space (i.e., the number of changesets in a project) significantly impacts the retrieval delay of different BERT-based models, though less in the case of the proposed model.
| i |
bf4b2f48-cb36-4e58-a4ba-3c69f43bdc87 |
approach that applies BERT to the bug localization problem (specifically, localizing bug-inducing changesets) that is more accurate than the state-of-the-art,
improvement over other recent BERT-based architectures proposed towards changeset retrieval, showing significant advantages with respect to retrieval speed,
evaluation and recommendations for key design choices in applying BERT to changesets (i.e., code change encoding, data granularity).
| i |
061bab98-821b-4994-a959-dfb7784cd7b4 | Significance of contribution. The BERT-based technique proposed in this paper enables semantic retrieval of software artifacts (specifically, changesets) for bug localization that goes beyond (and can complement) the exact term matching in the current popular state-of-the-art techniques (e.g., [1]}, [2]}). Relative to a similar, recent BERT-based technique [3]}, we offer an approach that improves retrieval speed significantly, in a way that supports real-world use, while also enhancing retrieval quality.
| i |
817d99f4-f1a1-4f4a-a0ae-d9cb091dd6e7 | Bug localization has generated significant research interest over the years. In this section, first, we survey related code element-based bug localization techniques, followed by approaches towards bug-inducing changeset retrieval.
Finally, we review methods for encoding changesets characteristics.
| w |
7fe4bc27-21ee-484a-b366-820a3c685677 | This paper presents an approach for automatically retrieving bug-inducing changesets for a newly reported bug. The approach uses the popular BERT model to more accurately match the semantics in the bug report text to the inducing changeset. More specifically, we describe the FBL-BERT model, based on the prior work by Khattab et al. [1]}, which speeds up the retrieval of results while performing fine grained matching across all embeddings in the two documents.
The results show an improvement in retrieval accuracy for bug reports that lack localization hints or have only partial hints. We also evaluate different approaches for utilizing changesets in BERT-like models, producing recommendations on the input data granularity and the use of special tokens for the purpose of capturing changeset semantics.
| d |
8ff61fea-ea90-4bab-912b-aa9a2b24792c | We presented the underlying concepts of UniNAS, a PyTorch-based framework with the ambitious goal of unifying a variety of NAS algorithms in one codebase.
Even though the use cases for this framework changed over time, mostly from DARTS-based to SPOS-based experiments, its underlying design approach made reusing old code possible at every step.
However, several technical details could be changed or improved in hindsight. Most importantly, configuration files should reflect the hierarchy levels (see Section REF ) for code simplicity and to avoid concerns about using module types multiple times. The current design favors readability, which is now a minor concern thanks to the graphical user interface.
Other considered changes would improve the code readability but were not implemented due to a lack of necessity and time.
| d |
dfdb7b3b-5acb-4ae6-b304-a958c8ae1f22 | In summary, the design of UniNAS fulfills all original requirements.
Modules can be arranged and combined in almost arbitrary constellations, giving the user an extremely flexible tool to design experiments.
Furthermore, using the graphical user interface does not require writing even a single line of code.
The resulting configuration files contain only the relevant information and do not suffer from a framework with many options.
These features also enable an almost arbitrary network design, combined with any NAS optimization method and any set of candidate operations. Despite that, networks can still be saved, loaded, and changed in various ways.
Although not covered here, several unit tests ensure that the essential framework components keep working as intended.
| d |
372b0b4d-f349-4a5d-b47b-7d478817723a | Finally, what is the advantage of using argument trees over writing code with the same results?
Compared to configuration files, code is more powerful and versatile but will likely suffer from problems described in Section REF .
Argument trees make any considerations about which parameters to expose unnecessary and can enforce the use of specific module types and subsets thereof.
However, their strongest advantage is the visualization and manipulation of the entire experiment design with a graphical user interface. This aligns well with Automated Machine Learning (AutoML), which is also intended to make machine learning available to a broader audience.
| d |
d5a4f12f-9b46-4c0a-9bca-a4d05722c776 | With a growing number of drivers relying on GoogleMaps or other navigation tools for dynamic routing, one question of interest is that, when everybody follows the advisory of shortest paths provided by one central platform, will these drivers still be able to gain from taking these paths?
This paper aims to tackle this question by accounting for competition among drivers.
When one chooses a path dynamically while navigating a road network, others may also do so to compete for limited road resources.
Therefore it is crucial to model the dynamic routing choices of many drivers simultaneously.
To address the above problem, we will first review the literature on route choice of single-agent and then move to that of multi-agent.
| i |
9d990440-e468-431e-90e2-9fe88ad2a0d5 | Convergence plots of both the lower level flow-dependent deep Q-learning and the upper level BO are presented in Figure (REF ). Figure (REF ) presents the variation of the average travel time of all controllable agents versus the number of episodes. Initially, agents spend a very long time (i.e., more than 700 seconds) on traveling to their destinations. Actually, some agents may not be able to reach their destinations and could record a total travel time of 1,000 seconds (i.e., the maximum allowed time in one episode). During the first 60 episodes, agents explore the road network and learn towards a better policy pretty fast. Therefore, their average travel time is substantially decreased from above 700 seconds to around 200 seconds. Despite some bouncing back and forth between 60 episodes and 300 episodes, the average travel time of all controllable agents stabilizes around 200 seconds after 300 episodes. Especially after 400 episodes, the average travel time almost barely changes. Figure (REF ) presents the distance between two consecutive evaluated \(\alpha \) 's in the BO algorithm. A smaller distance means that BO chooses to evaluate similar \(\alpha \) 's, indicating that the algorithm approaches convergence. In this study, we stop the BO algorithm when the distance is smaller than a threshold value (i.e., 2.5) for four times in a row.
With five randomly selected \(\alpha \) 's as starting point, BO reaches convergence after 9 additional iterations. As one could see, BO selects very different \(\alpha \) 's during the first 4 iterations and starts selecting similar \(\alpha \) 's from the 5th iteration.
<FIGURE> | r |
15c92c73-7f8d-4f38-b513-fa988c9c43a8 | With the illustrated convergence of both levels, the final result of the posterior probability distribution of the objective \(f\) (i.e., the negative average travel time) is shown in Figure (REF ). The x axis is the controllable variable \(\alpha \) , i.e., the offset in this study. The y axis is the objective \(f\) . The data records all evaluations, i.e., the evaluated \(f\) at a chosen \(\alpha \) . The mean and the standard deviation of the Gaussian process fitting based on the data are also plotted. As one could see, the standard deviation is small around evaluated \(\alpha \) 's while large when there is no nearby evaluated \(\alpha \) 's.
The final result suggests that the optimal \(\alpha \) is 4, i.e., \(\alpha ^*=4\) . With the optimal \(\alpha \) , the optimal value of the objective is around \(-200\) , meaning that with offset as 4, the average travel time of all controllable agents is 200 seconds. Compared with \(\alpha = 0\) and the corresponding average travel time as 220 seconds, the optimal \(\alpha \) could reduce the average travel time by \(9\%\) . A much larger \(\alpha \) (i.e., \(\alpha > 7\) ) may deteriorate the traffic condition and results in a long average travel time.
<FIGURE> | r |
eef96f70-88ed-40a9-83c2-4f97892d6d18 | To provide more insights, we decompose the average travel time of all controllable agents into two components, namely average waiting time (at intersection) and average cruising time. Both components are presented in Figure (REF ). In general, a smaller \(\alpha \) (e.g., \(\alpha < 10\) ) yields a smaller average waiting time and a smaller average cruising time while a larger \(\alpha \) results in a higher average waiting time and a higher cruising time.
This could be partially explained as follows. Although the road network is large in the sense that it has a large number of nodes and links, it covers a relatively small area in Manhattan. The average cruising time from one intersection to an adjacent intersection along the south-north direction is typically around 5 seconds, which is significantly smaller than 10 seconds. With an offset larger than 10 seconds, vehicles have a higher probability of stopping at the next intersection after they stop and wait for green light at the current intersection, resulting in not only a longer accumulative waiting time but also a deteriorated traffic condition. While with a smaller \(\alpha \) , vehicles could take advantage of the “green wave" and enjoy a smaller waiting time and better traffic condition. The lowest waiting time and the lowest cruising time are achieved when the offset is set as 4 seconds, which is exactly the previously derived optimal \(\alpha ^*\) .
| r |
2948781a-957e-433e-99b4-7c1cd1e43a18 | This paper develops a unified paradigm that models one's learning behavior and the system's equilibrating processes in a routing game among atomic selfish agents. Such a paradigm can assist policymakers in devising optimal operational and planning countermeasures under both normal and abnormal circumstances.
A flow-dependent mean field deep Q-learning algorithm is developed to tackle the route choice task of multi-commodity multi-class agents. The flow dependent mean action, which is defined as the traffic flow on the chosen link, carries partial but not full information of nearby agents and is thus able to partially capture the competition among agents. In addition, the use of flow-dependent mean action enables Q-value sharing and policy sharing, which is computationally efficient. The developed algorithm is extensively applied to three road networks, namely a two-node two-link network with CTM as its traffic propagation mechanism (see Section ), a Braess network with a linear volume delay function (see Section ), and a large-sized real-world road network with 69 nodes and 166 links implemented in SUMO (see Section ). The thorough testing unveils the versatility and effectiveness of the developed model-free MARL algorithm.
| d |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.