Heaviside Compression

This example demonstrates Tabular Compression using signal data with step functions (Heaviside-like patterns).

Problem Description

Goal: Compress and reconstruct signal data, discovering underlying patterns.

Input:

• Signal samples with step-like (Heaviside) patterns

• High-dimensional measurements per sample

Output:

• Compressed representation (learned automatically)

• Reconstruction of original signal

This example showcases:

• Dimensionality reduction

• Pattern discovery in latent space

• Autoencoder capabilities

Dataset

The dataset contains:

• x_train.csv — Training signal samples

• No separate output file (compression learns to reconstruct input)

Building the Model

GUIPython APICLI

1. Launch NeurEco and select Tabular Compression

2. Load x_train.csv as input (no output file needed)

3. Set compression parameters: - Compression tolerance: 0.01 (1% error) - Minimum compression coefficients: 1

4. Click BUILD

5. Observe the learned compression dimension

Analyzing Compression Coefficients

One of the most interesting aspects of compression is analyzing the latent space. With the settings above, NeurEco typically finds 2 compression coefficients for this dataset. The Compressor model provides a built-in method for visualization.

Visualizing coefficient relationships:

# Load the complete model (or encoder)
model = Tabular.Compressor()
model.load("./heaviside_model/complete.ednn")

# 2D visualization of latent space
model.plot_compression_coefficients(
    data_to_compress=x_train,
    neurons_ids=[0, 1],
    title="Latent Space Visualization",
    save_path="latent_space_2d.png"
)

The latent space typically shows a sun-like pattern radiating from the center outward, revealing the underlying structure of the Heaviside signals.

Coloring by signal amplitude:

You can color points by data properties to reveal patterns:

# Compute amplitudes for each sample
amplitudes = np.array([np.max(sample) for sample in x_train])

# Create class labels based on amplitude threshold
classes = np.where(amplitudes > 0.2, "high", "low")

# Plot compression coefficients colored by amplitude class
model.plot_compression_coefficients(
    data_to_compress=x_train,
    neurons_ids=[0, 1],
    data_labels=list(classes),
    title="Latent Space by Signal Amplitude"
)

This visualization shows how signals with different amplitudes are distributed in the compressed space.

Visualizing Reconstruction Quality

# Select a few samples
sample_indices = [0, 10, 20]

fig, axes = plt.subplots(len(sample_indices), 1, figsize=(12, 8))

for idx, sample_idx in enumerate(sample_indices):
    original = x_train[sample_idx]
    reconstructed = model.evaluate(x_train[sample_idx:sample_idx+1])[0]

    axes[idx].plot(original, 'b-', label='Original', alpha=0.7)
    axes[idx].plot(reconstructed, 'r--', label='Reconstructed', alpha=0.7)
    axes[idx].set_title(f'Sample {sample_idx}')
    axes[idx].legend()

plt.tight_layout()
plt.savefig('reconstruction_comparison.png')
plt.show()

Expected Results

After training:

• Significant compression — a big reduction in dimensions depending on the build settings (compress_tolerance)

• Low reconstruction error — Within the specified tolerance

Key Takeaways

This example demonstrates:

1. Automatic dimensionality reduction — NeurEco finds the optimal latent dimension

2. Pattern discovery — Latent space visualization reveals data structure

3. Separate encoder/decoder — Useful for IoT and distributed systems

4. Reconstruction quality — Controlled by compression tolerance

Variations to Try

• Adjust compress_tolerance and observe the effect on latent dimension

• Try different compress_decompress_size_ratio for IoT scenarios

• Use clustering to analyze discovered patterns

• Compare reconstruction quality across different samples