yoavg

Reinforcement Learning for Language Models

Yoav Goldberg, April 2023.

Why RL?

With the release of the ChatGPT model and followup large language models (LLMs), there was a lot of discussion of the importance of "RLHF training", that is, "reinforcement learning from human feedback". I was puzzled for a while as to why RL (Reinforcement Learning) is better than learning from demonstrations (a.k.a supervised learning) for training language models. Shouldn't learning from demonstrations (or, in language model terminology "instruction fine tuning", learning to immitate human written answers) be sufficient? I came up with a theoretical argument that was somewhat convincing. But I came to realize there is an additional argumment which not only supports the case of RL training, but also requires it, in particular for models like ChatGPT. This additional argument is spelled out in (the first half of) a talk by John Schulman from OpenAI. This post pretty much

rain-1

Purpose

Bootstrap knowledge of LLMs ASAP. With a bias/focus to GPT.

Avoid being a link dump. Try to provide only valuable well tuned information.

Prelude

Neural network links before starting with transformers.

veekaybee

Machine Learning Design Patterns

This book is all about patterns for doing ML. It's broken up into several key parts, building and serving. Both of these are intertwined so it makes sense to read through the whole thing, there are very many good pieces of advice from seasoned professionals. The parts you can safely ignore relate to anything where they specifically use GCP. The other issue with the book it it's very heavily focused on deep learning cases. Not all modeling problems require these. Regardless, let's dive in. I've included the stuff that was relevant to me in the notes.

Most Interesting Bullets:

• Machine learning models are not deterministic, so there are a number of ways we deal with them when building software, including setting random seeds in models during training and allowing for stateless functions, freezing layers, checkpointing, and generally making sure that flows are as reproducible as possib

zeyademam

Troubleshooting Convolutional Neural Networks

Intro

This is a list of hacks gathered primarily from prior experiences as well as online sources (most notably Stanford's CS231n course notes) on how to troubleshoot the performance of a convolutional neural network . We will focus mainly on supervised learning using deep neural networks. While this guide assumes the user is coding in Python3.6 using tensorflow (TF), it can still be helpful as a language agnostic guide.

Suppose we are given a convolutional neural network to train and evaluate and assume the evaluation results are worse than expected. The following are steps to troubleshoot and potentially improve performance. The first section corresponds to must-do's and generally good practices before you start troubleshooting. Every subsequent section header corresponds to a problem and the section is devoted to solving it. The sections are ordered to reflect "more common" issues first and under each header the "most-eas

jayspeidell

# Info on how to get your api key (kaggle.json) here: https://github.com/Kaggle/kaggle-api#api-credentials
!pip install kaggle
api_token = {"username":"USERNAME","key":"API_KEY"}
import json
import zipfile
import os
with open('/content/.kaggle/kaggle.json', 'w') as file:
    json.dump(api_token, file)
!chmod 600 /content/.kaggle/kaggle.json
!kaggle config path -p /content

seanjtaylor

5agado

CreatCodeBuild

shagunsodhani

The Batch Normalization paper describes a method to address the various issues related to training of Deep Neural Networks. It makes normalization a part of the architecture itself and reports significant improvements in terms of the number of iterations required to train the network.

Issues With Training Deep Neural Networks

Internal Covariate shift

Covariate shift refers to the change in the input distribution to a learning system. In the case of deep networks, the input to each layer is affected by parameters in all the input layers. So even small changes to the network get amplified down the network. This leads to change in the input distribution to internal layers of the deep network and is known as internal covariate shift.

It is well established that networks converge faster if the inputs have been whitened (ie zero mean, unit variances) and are uncorrelated and internal covariate shift leads to just the opposite.

lukas-h

Markdown License badges

Collection of License badges for your Project's README file.
This list includes the most common open source and open data licenses.
Easily copy and paste the code under the badges into your Markdown files.

Notes

• The badges do not fully replace the license informations for your projects, they are only emblems for the README, that the user can see the License at first glance.

Translations: (No guarantee that the translations are up-to-date)

• 🇫🇷 Cette liste en français