[ICLR 2026] Quantized Gradient Projection for Memory-Efficient Continual Learning

By Dongjun Kim, Seohyeon Cha, Huancheng Chen, Chianing Wang, and Haris Vikalo.

This repository contains the implementation for the paper Quantized Gradient Projection for Memory-Efficient Continual Learning, The Fourteenth International Conference on Learning Representations (ICLR 2026).

TL;DR

We propose QGPM, a memory-efficient and privacy-preserving continual learning framework that compresses task subspaces via quantization.

Conda Environment Setup

To create and activate the Conda environment using the provided environment.yml, follow these steps:

1. Create the environment:

   conda env create -f environment.yml
   

2. Activate the environment:

   conda activate qgpm_env
   

10-split CIFAR100

python qgpm_alexnet.py

5-Datasets

python qgpm_resnet.py

10/20-split miniImageNet

• 10-split

python qgpm_vit.py --dataset=miniimagenet --num_tasks=10

• 20-split

python qgpm_vit.py --dataset=miniimagenet --num_tasks=20

Acknowledgment

Parts of this codebase were adapted from GPM