[2205.11502] On the Paradox of Learning to Reason from Data

Logical reasoning is needed in a wide range of NLP tasks. Can a BERT model be trained end-to-end to solve logical reasoning problems presented in natural language? We attempt to answer this question in a confined problem space where there exists a set of parameters that perfectly simulates logical reasoning. We make observations that seem to contradict each other: BERT attains near-perfect accuracy on in-distribution test examples while failing to generalize to other data distributions over the exact same problem space. Our study provides an explanation for this paradox: instead of learning to emulate the correct reasoning function, BERT has in fact learned statistical features that inherently exist in logical reasoning problems. We also show that it is infeasible to jointly remove statistical features from data, illustrating the difficulty of learning to reason in general. Our result naturally extends to other neural models and unveils the fundamental difference between learning to re

1 mentions: @sirbayes
Date:

Referring Tweets

@sirbayes
@sirbayes @jacobmbuckman @slatestarcodex @GaryMarcus This paper does the experiment you suggest -“On the Paradox of Learning to Reason from Data”. It shows the DNN can represent the optimal function but cannot learn it by MLE because shortcut solutions on finite training set have higher LL. We need priors! t.co/p0GnfoplCj

Related Entries

The 32 Implementation Details of Proximal Policy Optimization (PPO) Algorithm
Read more The 32 Implementation Details of Proximal Policy Optimization (PPO) Algorithm
0 users, 2 mentions 2020/06/12 03:51
AdapterHub -  214 adapters for 30 text tasks and 32 languages
Read more AdapterHub - 214 adapters for 30 text tasks and 32 languages
0 users, 0 mentions 2020/12/20 05:21
Nested Variational Inference | OpenReview
Read more Nested Variational Inference | OpenReview
0 users, 1 mentions 2021/03/04 15:52
Lecture 94 — Charniaks Model — [ NLP || Christopher Manning || Stanford University ] - YouTube
Read more Lecture 94 — Charniaks Model — [ NLP || Christopher Manning || Stanford University ] - YouTube
0 users, 1 mentions 2021/05/23 06:19
ExpMRC
Read more ExpMRC
0 users, 2 mentions 2021/05/24 16:48

ML-Newsについて

機械学習の技術に関する情報は流速も早いし、分野も多様でキャッチアップが大変です。Twitterで機械学習用のリストを作っても、普段は機械学習以外の話題が多く流れており、効率的に情報収集するのは困難です。

ML-NewsはSNSを情報源とした機械学習に特化したニュースサイトです。機械学習に関する論文ブログライブラリコンペティション発表資料勉強会などの最新の情報を効率的に収集できます。

機械学習を応用した自然言語処理、画像認識、情報検索などの分野の情報や機械学習で必要になるデータ基盤やMLOpsの話題もカバーしています。
安定したサイト運営のためにGitHub sponsorを募集しています。

お知らせ

  • 2021/12/31: デザインを刷新しました
  • 2021/04/08: 日本語Kaggleのカテゴリを新設しました