Closed omarsar closed 4 years ago
When BERT Plays the Lottery, All Tickets Are Winning
https://arxiv.org/abs/2005.00561
The first author is even a part of the dair.ai community.
Evaluating NLP Models via Contrast Sets: https://arxiv.org/pdf/2004.02709v1.pdf
Title: Weight Poisoning Attacks on Pre-trained Models
Why important: Operates in the emerging area of security in NLP and focuses on attacking a pre-trained language model that is fine-tuned on data of a target task.
Idea:
It is a proof-of-concept algorithm for poisoning the weights of a pre-trained model (such as BERT, XLNet, etc...) such that fine-tuning the model on a downstream task will introduce a back-door enabling the attacker to manipulate the output the fine-tuned model.
Provide a poisoned
model for others to download that we can then later exploit. In particular, the authors propose weight poisoning
attacks that inject vulnerabilities into a model that exposes “backdoors” after the model is fine-tuned on a target task. Trigger words are rare words (arbitrary nouns also work) that are inserted in a sequence and enable switching the model’s prediction to a target label.
Title: Learning distributed representation of concepts Author: Geoffrey E. Hinton Link
Twitter Thread which inspired to read this seminal paper from 1986.
Title: Generating Long Sequences with Sparse Transformers Link
The GPT3 used something similar to sparse transformer and they cited this paper. In addition, we could try to implement the paper as task in the project of ( paper_implementations ).
End-to-end Object Detection with Transformers
https://ai.facebook.com/research/publications/end-to-end-object-detection-with-transformers
Comment a paper you would like us to discuss during our weekly paper reading discussion.
You can vote on a suggested paper by using the đź‘Ť emoji. I will close the issue in one or two days and select the paper with the most votes. Then I will make the announcement at the beginning of the week. Thanks.