Commit
·
905f8ef
1
Parent(s):
007041f
Create 1910.01108.atom
Browse files- arxiv/1910.01108.atom +54 -0
arxiv/1910.01108.atom
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<?xml version="1.0" encoding="UTF-8"?>
|
| 2 |
+
<feed xmlns="http://www.w3.org/2005/Atom">
|
| 3 |
+
<link href="http://arxiv.org/api/query?search_query%3D%26id_list%3D1910.01108%26start%3D0%26max_results%3D1" rel="self" type="application/atom+xml"/>
|
| 4 |
+
<title type="html">ArXiv Query: search_query=&id_list=1910.01108&start=0&max_results=1</title>
|
| 5 |
+
<id>http://arxiv.org/api/GrfVMsjRJ5QfXT4borMr9RuzWhk</id>
|
| 6 |
+
<updated>2023-06-09T00:00:00-04:00</updated>
|
| 7 |
+
<opensearch:totalResults xmlns:opensearch="http://a9.com/-/spec/opensearch/1.1/">1</opensearch:totalResults>
|
| 8 |
+
<opensearch:startIndex xmlns:opensearch="http://a9.com/-/spec/opensearch/1.1/">0</opensearch:startIndex>
|
| 9 |
+
<opensearch:itemsPerPage xmlns:opensearch="http://a9.com/-/spec/opensearch/1.1/">1</opensearch:itemsPerPage>
|
| 10 |
+
<entry>
|
| 11 |
+
<id>http://arxiv.org/abs/1910.01108v4</id>
|
| 12 |
+
<updated>2020-03-01T02:57:50Z</updated>
|
| 13 |
+
<published>2019-10-02T17:56:28Z</published>
|
| 14 |
+
<title>DistilBERT, a distilled version of BERT: smaller, faster, cheaper and
|
| 15 |
+
lighter</title>
|
| 16 |
+
<summary> As Transfer Learning from large-scale pre-trained models becomes more
|
| 17 |
+
prevalent in Natural Language Processing (NLP), operating these large models in
|
| 18 |
+
on-the-edge and/or under constrained computational training or inference
|
| 19 |
+
budgets remains challenging. In this work, we propose a method to pre-train a
|
| 20 |
+
smaller general-purpose language representation model, called DistilBERT, which
|
| 21 |
+
can then be fine-tuned with good performances on a wide range of tasks like its
|
| 22 |
+
larger counterparts. While most prior work investigated the use of distillation
|
| 23 |
+
for building task-specific models, we leverage knowledge distillation during
|
| 24 |
+
the pre-training phase and show that it is possible to reduce the size of a
|
| 25 |
+
BERT model by 40%, while retaining 97% of its language understanding
|
| 26 |
+
capabilities and being 60% faster. To leverage the inductive biases learned by
|
| 27 |
+
larger models during pre-training, we introduce a triple loss combining
|
| 28 |
+
language modeling, distillation and cosine-distance losses. Our smaller, faster
|
| 29 |
+
and lighter model is cheaper to pre-train and we demonstrate its capabilities
|
| 30 |
+
for on-device computations in a proof-of-concept experiment and a comparative
|
| 31 |
+
on-device study.
|
| 32 |
+
</summary>
|
| 33 |
+
<author>
|
| 34 |
+
<name>Victor Sanh</name>
|
| 35 |
+
</author>
|
| 36 |
+
<author>
|
| 37 |
+
<name>Lysandre Debut</name>
|
| 38 |
+
</author>
|
| 39 |
+
<author>
|
| 40 |
+
<name>Julien Chaumond</name>
|
| 41 |
+
</author>
|
| 42 |
+
<author>
|
| 43 |
+
<name>Thomas Wolf</name>
|
| 44 |
+
</author>
|
| 45 |
+
<arxiv:comment xmlns:arxiv="http://arxiv.org/schemas/atom">February 2020 - Revision: fix bug in evaluation metrics, updated
|
| 46 |
+
metrics, argumentation unchanged. 5 pages, 1 figure, 4 tables. Accepted at
|
| 47 |
+
the 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing
|
| 48 |
+
- NeurIPS 2019</arxiv:comment>
|
| 49 |
+
<link href="http://arxiv.org/abs/1910.01108v4" rel="alternate" type="text/html"/>
|
| 50 |
+
<link title="pdf" href="http://arxiv.org/pdf/1910.01108v4" rel="related" type="application/pdf"/>
|
| 51 |
+
<arxiv:primary_category xmlns:arxiv="http://arxiv.org/schemas/atom" term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
|
| 52 |
+
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
|
| 53 |
+
</entry>
|
| 54 |
+
</feed>
|