README.md

metadata

library_name: clearn
tags:
  - continual-learning
  - catastrophic-forgetting
  - ewc
  - pytorch
license: mit

clearn Demo: EWC Continual Learning

This model was trained with clearn — a continual learning library for PyTorch.

What is this?

A simple MLP trained on 3 sequential fraud detection tasks using Elastic Weight Consolidation (EWC). Despite learning 3 tasks sequentially, the model retains 100% accuracy on all previous tasks.

How to use

pip install clearn-ai

import clearn
import torch.nn as nn

# Recreate the architecture
model = nn.Sequential(nn.Linear(128, 256), nn.ReLU(), nn.Linear(256, 10))

# Load the continual learning checkpoint
cl_model = clearn.load("./checkpoint", model=model)

# See retention across all tasks
print(cl_model.diff())

Retention Report

RetentionReport
├── fraud_q1: 100.0% retained
├── fraud_q2: 100.0% retained
├── fraud_q3: 100.0% retained
├── plasticity_score: 1.00
├── stability_score: 1.00
└── recommendation: "stable — no action needed"

Strategy

• Strategy: EWC (Elastic Weight Consolidation)
• Lambda: 5000
• Tasks: 3 sequential fraud detection tasks
• Architecture: MLP (128 → 256 → 10)

Built with clearn — Wrap once. Train forever.