https://github.com/finite-sample/pareto-gd

Pareto Improving GD. Only update when when doing so improves or maintains performance on all previously seen data.
https://github.com/finite-sample/pareto-gd

Last synced: 4 months ago
JSON representation

Pareto Improving GD. Only update when when doing so improves or maintains performance on all previously seen data.

• Host: GitHub
• URL: https://github.com/finite-sample/pareto-gd
• Owner: finite-sample
• Created: 2025-06-27T22:53:45.000Z (12 months ago)
• Default Branch: main
• Last Pushed: 2025-07-16T20:38:02.000Z (11 months ago)
• Last Synced: 2025-09-26T23:38:28.479Z (9 months ago)
• Language: Jupyter Notebook
• Size: 1.26 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# (Don't) Forget About It: Forgetting-Penalized Supervised Learning

## Background

A well-known problem in machine learning is **regression**: as models update, they sometimes "forget" how to correctly handle examples they previously got right. This is especially frustrating in production or user-facing systems, where a model suddenly failing on known-good cases can be more disruptive than missing new ones.

Catastrophic forgetting is well-studied in **continual learning** (French, 1999), and rehearsal/buffer methods are common. But for standard supervised learning, less attention has been paid to *actively penalizing regression* during ordinary training.

---

## Approaches Compared

We compare three strategies:

**1. Standard Training (Baseline)**  

The usual approach—minimize training loss with no explicit mechanism to prevent forgetting.

**2. Forgetting-Penalized Training**  

Inspired by continual learning methods like Elastic Weight Consolidation (Kirkpatrick et al., 2017), this adds a penalty whenever an example previously classified correctly becomes incorrect. It discourages "unlearning," but does not eliminate all changes.

**3. Soft Pareto-Penalized Training**  

Drawing on Pareto-improvement ideas and recent multi-task optimization research (Lin et al., 2019; Navon et al., 2021), this method penalizes *any* increase in per-example loss—not just flips from correct to incorrect. It enforces a softer, broader "do no harm" principle across all training examples.

---

## Experiment

On the Adult income dataset, we trained all three methods with identical neural network architectures. Penalties were introduced after a warmup period, allowing the model to stabilize before beginning to penalize regressions.

---

## Results

| Method           | Total Forgetting | Final Train Acc | Final Val Acc |

|------------------|------------------|-----------------|---------------|

| **Baseline**     | 5668             | 0.794           | 0.788         |

| **Forgetting Pen.** | 122          | 0.759           | 0.760         |

| **Soft Pareto**  | 290              | 0.786           | 0.783         |

- Both penalized methods reduced forgetting by an order of magnitude **compared to baseline**.

- **Soft Pareto** provided a strong trade-off: low forgetting with minimal accuracy loss.

- **Forgetting-Penalized** achieved the lowest forgetting, but at a more significant cost to accuracy.

- **Baseline training** delivered the highest accuracy—but experienced frequent regression.

---

## Contribution

While regularization and continual learning are well-established, our work shows that **simple, lightweight penalty-based mechanisms**—added to ordinary training—can greatly reduce regression *without substantial accuracy loss*. The **Soft Pareto loss** is especially practical, implementing a “do no harm” bias that’s easy to integrate.

---

## Where It Matters

- **Production-grade systems** where regression on known-good cases is unacceptable.

- **Human-facing models** where consistency matters to user trust.

- **High-stakes domains** like medical, fraud detection, or compliance.

- **Curriculum or staged learning setups**, where early learning shouldn't be overwritten by later stages.

---

## Summary

If maintaining correctness on previously learned examples matters—even under normal supervised training—then adding **penalty terms** for forgetting or loss regression is effective, easy to implement, and provides a natural “Pareto bias” in practice.