plurigrid/asi / ies/music-topos/.codex/skills/feedforward-learning-local
feedforward-learning-local
Verified| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 2.5 KB |
| ff_layer.jl | other | 4.4 KB |
SKILL.md
View on GitHub# Feedforward Learning Local
**Category:** Phase 3 Core - Alternative Learning Paradigms
**Status:** Skeleton Implementation
**Dependencies:** None (standalone learning framework)
## Overview
Implements forward-forward (FF) learning algorithm and variants that eliminate backpropagation through local, layer-wise contrastive objectives. Each layer learns to distinguish positive from negative data independently.
## Capabilities
- **Forward-Forward Algorithm**: Hinton's layer-local learning
- **Contrastive Objectives**: Positive/negative data discrimination
- **No Backprop**: Purely feedforward gradient computation
- **Statistical Communication**: Inter-layer coordination via activity statistics
## Core Components
1. **FF Layer** (`ff_layer.jl`)
- Local goodness function per layer
- Positive/negative data generation
- Layer-wise gradient updates
2. **Contrastive Learning** (`contrastive_learning.jl`)
- Contrastive divergence variants
- Energy-based formulations
- Hybrid supervised/unsupervised objectives
3. **Statistical Coordination** (`statistical_coordination.jl`)
- Activity normalization between layers
- Whitening and decorrelation
- Predictive coding integration
4. **FF Network** (`ff_network.jl`)
- Multi-layer FF architecture
- Inference and training loops
- Comparison with backprop baselines
## Integration Points
- **Input from**: Raw data (no dependencies on other skills)
- **Output to**: `emergent-role-assignment` (decentralized learning signals)
- **Coordinates with**: `categorical-composition` (compositional learning)
## Usage
```julia
using FeedforwardLearningLocal
# Create FF network
network = FFNetwork([
FFLayer(input_dim=784, hidden_dim=500, threshold=2.0),
FFLayer(input_dim=500, hidden_dim=500, threshold=2.0),
FFLayer(input_dim=500, hidden_dim=10, threshold=1.0)
])
# Train on MNIST
for (x_pos, y) in train_data
# Generate negative data by corrupting label
x_neg = overlay_wrong_label(x_pos, y)
# Local learning at each layer
train_step!(network, x_pos, x_neg)
end
# Inference
predictions = predict(network, test_data)
```
## References
- Hinton "The Forward-Forward Algorithm" (2022)
- LeCun et al. "A Tutorial on Energy-Based Learning" (2006)
- Nokland & Eidnes "Training Neural Networks with Local Error Signals" (ICML 2019)
## Implementation Status
- [x] Basic FF layer implementation
- [x] Positive/negative data generation
- [ ] Multiple variants (supervised, unsupervised)
- [ ] Benchmark against backprop
- [ ] Integration with predictive coding
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |