-
Notifications
You must be signed in to change notification settings - Fork 30
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
1 parent
4153a1e
commit f732cf9
Showing
6 changed files
with
214 additions
and
0 deletions.
There are no files selected for viewing
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,214 @@ | ||
| <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> | ||
| <html xmlns="http://www.w3.org/1999/xhtml"> | ||
|
|
||
| <head> | ||
|
|
||
| <!-- Global site tag (gtag.js) - Google Analytics --> | ||
| <script async src="https://www.googletagmanager.com/gtag/js?id=G-GDXSC5Y2BD"></script> | ||
| <script> | ||
| window.dataLayer = window.dataLayer || []; | ||
| function gtag(){dataLayer.push(arguments);} | ||
| gtag('js', new Date()); | ||
|
|
||
| gtag('config', 'G-GDXSC5Y2BD'); | ||
| </script> | ||
|
|
||
| <meta http-equiv="Content-Type" content="text/html; charset=windows-1252"> | ||
|
|
||
| <script src="./files/head.js"></script> | ||
|
|
||
| <meta name="viewport" content="width=device-width, initial-scale=1"> | ||
|
|
||
| <script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script> | ||
| <script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script> | ||
|
|
||
| <meta name="keywords" content="MIT,Microsoft Research, Machine Learning,Rank Reduction,Computer Science,Machine,Artificial,Intelligence"> | ||
|
|
||
| <title>The Truth Is In There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction</title> | ||
| <link rel="stylesheet" href="./files/font.css"> | ||
| <link rel="stylesheet" href="./files/main.css"> | ||
|
|
||
| <link rel="stylesheet" type="text/css" | ||
| href="https://cdn.rawgit.com/dreampulse/computer-modern-web-font/master/fonts.css"> | ||
| <style> | ||
| body { | ||
| font-family: "Computer Modern Serif", serif; | ||
| font-size: 14pt; | ||
| } | ||
|
|
||
|
|
||
| * {padding:0;margin:0;box-sizing:border-box;} | ||
| #video { | ||
| position: relative; | ||
| padding-bottom: 45%; /* 16:9 */ | ||
| height: 0; | ||
| } | ||
| #video iframe { | ||
| position: absolute; | ||
| top: 0; | ||
| left: 0; | ||
| width: 80%; | ||
| height: 100%; | ||
| transform: translateX(12.5%); | ||
| } | ||
|
|
||
| </style> | ||
|
|
||
| <style type="text/css">/** | ||
| * Style sheet used by new LibX tooltip code | ||
| */ | ||
|
|
||
| /* We insert a <div> with libx-tooltip style under the body. | ||
| * This will inherit body's style - we can't afford to inherit undesirable | ||
| * styles and we must redefine what we need. OTOH, some things, e.g. | ||
| * font-size, might be ok to be inherited to stay within the page's tone. | ||
| */ | ||
| .libx-tooltip { | ||
| display: none; | ||
| overflow: visible; | ||
| padding: 5px; | ||
| z-index: 100; | ||
| background-color: #eee; | ||
| color: #000; | ||
| font-weight: normal; | ||
| font-style: normal; | ||
| text-align: left; | ||
| border: 2px solid #666; | ||
| border-radius: 5px; | ||
| -webkit-border-radius: 5px; | ||
| -moz-border-radius: 5px; | ||
| } | ||
|
|
||
| .libx-tooltip p { | ||
| /* override default 1em margin to keep paragraphs inside a tooltip closer together. */ | ||
| margin: .2em; | ||
| } | ||
| </style><style type="text/css">/** | ||
| * Style sheet used by LibX autolinking code | ||
| */ | ||
| .libx-autolink { | ||
| } | ||
|
|
||
| </style> | ||
|
|
||
| </head> | ||
|
|
||
| <body> | ||
|
|
||
| <div class="outercontainer"> | ||
| <div class="container"> | ||
|
|
||
| <div class="content project_title"> | ||
| <center> | ||
| <br> | ||
| <h2>The Truth Is In There: Improving Reasoning in Language Models <br>with Layer-Selective Rank Reduction</h2> | ||
| <div class="authors"> | ||
| <a href="https://pratyushasharma.github.io/">Pratyusha Sharma</a>, | ||
| <a href="https://www.jordantash.com/">Jordan Ash*</a>, and | ||
| <a href="https://dipendramisra.com/">Dipendra Misra*</a> | ||
| </div> | ||
| <!-- <br> --> | ||
| <!-- <a href="https://arxiv.org/abs/2106.02039">Paper</a> --> | ||
| <!-- <a href="./trajectory-transformer-neurips-2021.pdf">Paper</a> --> | ||
| <div> | ||
| <span class="tag"> | ||
| <a href="https://arxiv.org/abs/2106.02039">Paper</a> | ||
| <!-- <a href="./trajectory-transformer-neurips-2021.pdf">Paper</a> --> | ||
| <a href="https://github.com/JannerM/trajectory-transformer">Code</a> | ||
| <a href="files/bib.txt">BibTex</a> | ||
| </span> | ||
| </div> | ||
| </center> | ||
| </div> | ||
|
|
||
| <br><br> | ||
|
|
||
| <div class="content"> | ||
| <center> | ||
| <div class="text"> | ||
| <p> | ||
| <div class="title"><b>Summary</b></div> | ||
| <!-- <b> | ||
| <font size="5">Summary</font> | ||
| </b> --> | ||
| <!-- --> | ||
| Transformer-based Large Language Models (LLMs) have become a fixture in modern machine learning. <br> | ||
| Correspondingly, significant resources are allocated towards research that aims to further advance this technology, <br>typically resulting in models of increasing size that are trained on increasing amounts of data. <br> | ||
| This work, however, demonstrates the surprising result that it is often possible to improve the performance of LLMs by <br>simply removing higher-order components of their constituent weight matrices in the multi-layer perception (MLP) <br>layers. This simple intervention, which we call LAyer-SElective Rank reduction LASER, can be done on a model after <br>training has completed, and requires no additional parameters or data. LASER can dramatically boost predictive <br>performance on question-answering tasks and across various modalities for which Transformers are used. <br> | ||
| </p> | ||
| </div> | ||
| </center> | ||
| </div> | ||
| <br> | ||
| <br> | ||
|
|
||
| <center> | ||
| <img width=60% src="main.png"></img> | ||
| <br> | ||
| <br> | ||
| <i>LAyer SElective Rank reduction (LASER) replaces a specific weight matrix W of the Transformer model by its rank-$k$ <br>approximation and observes the change in the behavior of the model. We find that this rank approximation, especially for MLP <br>weights at the latter layers of the model, often offers surprising benefits to model performance.</i> | ||
| </center> | ||
| <p> | ||
| <br><br> | ||
|
|
||
|
|
||
| <div class="content"> | ||
| <div class="text"> | ||
| <center> | ||
| <p> | ||
| <div class="title"><b>Layer Selective Rank Reduction Improves Generalization</b></div> | ||
| <!-- <b> | ||
| <font size="5">Layer Selective Rank Reduction Improves Generalization</font> | ||
| </b> --> | ||
| <!-- --> | ||
| <!-- Predictive dynamics models often have excellent single-step error, but poor long-horizon accuracy due to compounding errors. | ||
| We show that Transformers are more reliable long-horizon predictors than state-of-the-art single-step models, even in continuous Markovian domains. --> | ||
| </p> | ||
| <br> | ||
| <br> | ||
| <img width=40% src="loss.png"></img> | ||
| | ||
| | ||
| <br> | ||
| <br> | ||
| <i>The effect of rank reduction across different layer types is not uniform. This figure shows the effect of rank <br>reduction for GPT-J as studied on the CounterFact dataset. The dashed line is the base model loss. In the attention <br>layers (key, query, value, out matrices), while its clear matrices could be significantly rank-reduced without damaging <br>the learned hypothesis, there is very little performance increase. Alternatively, for the multi-layer perceptron (MLP) <br>layers, rank reduction goes from uniformly harming to improving the model's performance (at layer 20).</i> | ||
| </center> | ||
| <br> | ||
| </div> | ||
| </div> | ||
| <br> | ||
| <br> | ||
|
|
||
| <div class="content"> | ||
| <div class="text"> | ||
| <center> | ||
| <p> | ||
| <div class="title"><b>LASER offers a kind of denoising procedure that makes weakly learned facts accessible</b></div> | ||
| <!-- <b> | ||
| <font size="5.0"> | ||
| Beam search as trajectory optimizer | ||
| </font> | ||
| </b> --> | ||
| <!-- . --> | ||
| <!-- Various control settings can be reduced to slight modifications of beam search with a sequence model. --> | ||
| <ul> | ||
| <br> | ||
| <img width=40% src="corrected.png"></img> | ||
| <img width=11% src="corrected-2.png"></img> | ||
| <br> | ||
| <li>Which datapoints benefit from LASER? We analyze how frequently in the training data ``corrected'' facts occur. <br> GPT-J is an ideal test bed for such analysis since its training data, the PILE dataset, is publicly available. <br> (a) For GPT-J evaluated on Counterfact we retrieve all the datapoints in the training data that contain a mention of <br>both the entity of interest and the answer that correspond to each sample in CounterFact. (b) A plot depicting the <br>cumulative top-10 accuracy of the model on all datapoints that occur in the training data less than or equal to <br>the frequency indicated on the x-axis. The plot looks at how the accuracy changes before and after LASER. (c) The <br>largest boost in performance occurs for low-frequency samples. Demonstrates the amount of boost offered by LASER for<br> data binned by the frequency with which corresponding facts occur in the training data. Maximal improvements in <br>accuracy are from datapoints that have less-frequent occurrences in the training data as opposed to those that occur<br> more frequently.</li> | ||
| <br> | ||
| <br> | ||
| </ul> | ||
| </p> | ||
| </center> | ||
| </div> | ||
| </div> | ||
|
|
||
| </div> | ||
| </div> | ||
|
|
||
| <br><br><br><br> | ||
|
|
||
|
|
||
| </div></body></html> |
