Interpretability - API Specification
Version: v1.0.0 | Status: Active | Last Updated: March 2026
1. Overview
The interpretability module provides Sparse Autoencoders (SAE) for neural network analysis. Enables feature extraction and interpretability analysis of learned representations.
2. Core Components
2.1 Classes
| Class | Description |
|---|---|
SparseAutoencoder | Sparse autoencoder model with L1 sparsity penalty for feature extraction |
2.2 Functions
| Function | Signature | Description |
|---|---|---|
train_sae | (data, hidden_dim, sparsity_weight, ...) -> SparseAutoencoder | Train a sparse autoencoder on activation data |
analyze_features | (sae, data) -> dict | Analyze learned features: activation patterns, sparsity, top-k features |
3. Usage Example
from codomyrmex.interpretability import train_sae, analyze_features
import numpy as np
activations = np.random.randn(1000, 256) # 1000 samples, 256-dim activations
sae = train_sae(activations, hidden_dim=512, sparsity_weight=0.01)
analysis = analyze_features(sae, activations)
print(f"Active features: {analysis['num_active_features']}")