alcnet 0.1.1

Adaptive Layer Condensation Networks - Learning Dynamic Compression Ratios for Hierarchical Feature Selection

ALCNet - Adaptive Layer Condensation Networks

Learning Dynamic Compression Ratios for Hierarchical Feature Selection

ALCNet is a neural network architecture where compression ratios at each layer are learned as trainable parameters rather than fixed before training. This enables task-adaptive architecture optimization without expensive neural architecture search.

Key Features

• Learnable Compression Ratios: Target sparsity at each layer adapts during training
• Task-Adaptive: Simple tasks learn aggressive compression, complex tasks preserve features
• Differentiable: End-to-end training with backpropagation
• No NAS Required: Eliminates expensive architecture search
• Interpretable: Learned compression ratios reveal task complexity

Installation

pip install alcnet

Quick Start

import torch
from alcnet import ALCNet, ALCNetTrainer
import torch.optim as optim

# Create model with learnable compression ratios
model = ALCNet([784, 256, 128, 64, 10])

# Setup training
optimizer = optim.Adam(model.parameters(), lr=0.001)
trainer = ALCNetTrainer(model, optimizer, lambda_sparse=0.001, lambda_ratio=0.01)

# Train on your data
history = trainer.fit(train_loader, val_loader, epochs=50)

# Check learned compression ratios
print(model.get_compression_ratios())

How It Works

Unlike traditional networks with fixed layer sizes, ALCNet learns optimal compression at each layer:

1. Standard Approach: You decide [784, 256, 128, 64, 10] before training
2. ALCNet Approach: Network learns how much to compress at each layer based on task

Architecture

Input Layer (784) 
    ↓  [Learnable ρ⁽¹⁾]
Hidden Layer (256)
    ↓  [Learnable ρ⁽²⁾]
Hidden Layer (128)
    ↓  [Learnable ρ⁽³⁾]
Hidden Layer (64)
    ↓
Output (10)

Each ρ⁽ⁱ⁾ ∈ (0,1) controls compression intensity at layer i.

Example Results

Simple tasks may learn aggressive early compression:

layer_1: ρ = 0.20 (keep 20% of neurons)
layer_2: ρ = 0.15
layer_3: ρ = 0.10

Complex tasks preserve more features:

layer_1: ρ = 0.70 (keep 70% of neurons)
layer_2: ρ = 0.60
layer_3: ρ = 0.40

Advanced Usage

Analysis and Visualization

from alcnet import CompressionAnalyzer

analyzer = CompressionAnalyzer(model)

# Plot compression evolution during training
analyzer.plot_compression_evolution(history)

# Compute feature survival probabilities
survival = analyzer.compute_feature_survival(dataloader)

# Get summary statistics
stats = analyzer.get_summary_stats(history)

Custom Configuration

# Specify initial compression ratios
initial_ratios = [0.5, 0.4, 0.3]
model = ALCNet([784, 256, 128, 64, 10], initial_ratios=initial_ratios)

# Adjust loss weights
trainer = ALCNetTrainer(
    model, 
    optimizer,
    lambda_sparse=0.005,  # Sparsity weight
    lambda_ratio=0.05     # Ratio matching weight
)

Examples

See the examples/ directory:

• mnist_example.py: Complete MNIST classification example
• More examples coming soon

Requirements

• Python ≥ 3.7
• PyTorch ≥ 1.9.0
• NumPy
• matplotlib (optional, for visualization)
• scikit-learn (optional, for data utilities)

Citation

If you use ALCNet in your research, please cite:

@article{jana2026alcnet,
  title={Adaptive Layer Condensation Networks: Learning Dynamic Compression Ratios for Hierarchical Feature Selection},
  author={Jana, Ananda},
  journal={arXiv preprint},
  year={2026}
}

Paper

The full paper describing the theory, implementation, and experimental validation of ALCNet is available at: [Link to paper]

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

Ananda Jana
Indian Institute of Science Education and Research Thiruvananthapuram (IISER TVM), Kerala, India

Acknowledgments

This work explores learnable compression mechanisms in neural network architectures through independent research into adaptive feature selection and hierarchical compression strategies.