Skip to content

itsmemala/LACL

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Look-Ahead Continual Learning

This repository provides the code for implementing look-ahead (LA) continual learning and baseline continual learning methods for text classification.

Parts of the code are derived from the following repositories:

  1. https://github.com/ZixuanKe/PyContinual
  2. https://github.com/hyscn/AdaBOP
  3. https://github.com/SunWenJu123/rp2f
  4. https://github.com/mohmdelsayed/upgd

Abstract

Achieving continual learning (CL) with deep neural networks requires balancing stability and plasticity while enabling knowledge transfer. In this work, we focus on offline learning algorithms under the constraints: (I) no access to training data from prior tasks (II) no access to task-id at inference time. We introduce a novel measure, the relative parameter-importance, which measures the relative importance of each parameter with respect to both the current and past tasks. Parameters with high relative importance are interpreted as more important for maintaining past-task stability and thus heavily regularised, whereas parameters with low relative-importance are allowed to be more freely updated. Unlike existing methods, our approach allows the update of parameters with high past-task importance when they have low relative-importance, thus enabling backward knowledge transfer in addition to tackling the stability-plasticity trade-off. We demonstrate improvements against state-of-the-art CL methods on both class-incremental and domain-incremental learning text classification problems.

Table of Contents

  1. Installation
  2. Workflow

Installation

Run the following commands to set things up.

git clone https://github.com/itsmemala/LACL.git
cd LACL
conda create -n lacl python==3.10
pip install requirements.txt

Workflow

To run LA experiments (for the Intent classification dataset, for example) using default hyper-parameters, run the following command. Change the hyper-parameter values as required through the command line or by updating the file.

bash scripts\\intent_sh_la_mas_chsf.sh random0 0 0 0.04854989 1641.28483697 1.0 1.0 0.8 True 0.1
bash scripts\\intent_sh_la_mas_chsf.sh random3 3 0 4.49536009 77.30662811 1.0 1.0 0.8 True 0.1
bash scripts\\intent_sh_la_mas_chsf.sh random6 6 0 28.24295365 246.34804902 1.0 1.0 0.8 True 0.1

About

Look-Ahead Continual Learning

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published