SHARP: Sleep-based Hierarchical Accelerated Replay for Long Range Non-Stationary Temporal Pattern Recognition

This repository contains the implementation of a hierarchical memory architecture for single-pass streaming sequence learning under non-stationary dynamics. The proposed framework has been accepted as an oral presentation in the Fifth Conference on Lifelong Learning, paper link: https://arxiv.org/abs/2606.00732.

---

Introduction

The model is inspired by biological memory systems and introduces a separation between:

• Memory (non-credit-assigned accumulation)
• Pattern recognition (credit-assigned inference)

It leverages accelerated sequential replay during sleep phases to extend effective temporal context without increasing online computational cost.

---

Key Features

• Single-pass (online) learning — no revisiting past data
• Hierarchical memory organization
• Accelerated sequential replay (sleep phase)
• Adaptive compute allocation (more compute early, less later)
• Improved retention and generalization
  • Backward BPC → past retention
  • Current BPC → adaptation
  • Forward BPC → generalization

---

Core Idea

Traditional models conflate memory with pattern recognition. This work instead decomposes learning into:

• Memory (no credit assignment)
• Pattern Recognition (with credit assignment)

Memory is:

• accumulated online (wake phase)
• consolidated offline (sleep phase)

During sleep:

• temporally structured sequences are replayed
• replay is accelerated via downsampling
• higher layers learn long-range structure

---

Benchmarks

Evaluated on:

• text8
• PG-19

Metrics:

• Forward BPC → future generalization
• Backward BPC → retention of past data
• Current BPC → adaptation to recent data

---

Hardware Requirements

The model is lightweight compared to large transformer-based systems, but dataset size (especially PG-19) can be demanding.

Minimum

• CPU: 4+ cores
• RAM: 8 GB
• Storage: ~10–20 GB (datasets + checkpoints)

Recommended

• CPU: 8–16 cores
• RAM: 16–32 GB
• GPU: Optional (Apple Silicon / CUDA GPU supported via PyTorch)
• Storage: 50+ GB (for large-scale experiments)

Notes

• Training is sequential and streaming, so GPU is not strictly required
• Larger memory layers and longer sequences benefit from more RAM
• PG-19 preprocessing can be slow and storage-heavy

---

Setup Instructions

1. Change directory to SHARP

cd sharp

2. Create environment (recommended)

Using conda

conda create -n sleep python=3.12
conda activate sleep

Or using venv

python -m venv sleep_env
source sleep_env/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Install the package (editable mode)

pip install -e .

Avoid using python setup.py install — it may create dependency conflicts.

5. Prepare datasets

Datasets are automatically downloaded or processed during training scripts.

For PG-19:

• First run may take time due to download + preprocessing
• Ensure sufficient disk space

6. Run training

python train_pg19.py

---

Repository Structure

sleep_experiment/
│
├── sharp/ # Core model implementation
├── benchmark/ # Benchmark experiment scripts
├── experiments/ # Simulation experiment scripts
├── dataset/ # Data storage (ignored by git)
├── pickle_files/ # Intermediate results (ignored by git)
├── plots/ # Visualization outputs
├── saved_models/ # Checkpoints (ignored by git)
└── requirements.txt # Dependencies

---

Notes

• The system is designed for single-pass streaming learning
• Performance depends on:
  • memory hierarchy depth
  • replay schedule
  • downsampling rate