Separable Temporal-Spectral CNN

Overview

For 2D inputs with (time, feature) structure (spectrograms, sensor arrays, multi-channel time series), use asymmetric convolution kernels: first apply tall kernels (3x1) along the time axis, then wide kernels (1x3) along the feature axis. This decouples temporal pattern extraction from cross-feature learning, reducing parameters vs square kernels.

Quick Start

from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, BatchNormalization

def build_separable_cnn(time_steps, n_features, n_outputs):
    inp = Input((time_steps, n_features, 1))
    # Temporal convolutions
    x = Conv2D(32, (3, 1), activation='relu', padding='same')(inp)
    x = MaxPooling2D((2, 1))(x)
    x = BatchNormalization()(x)
    x = Conv2D(64, (3, 1), activation='relu', padding='same')(x)
    x = MaxPooling2D((2, 1))(x)
    # Spectral convolutions
    x = Conv2D(128, (1, 3), activation='relu', padding='same')(x)
    x = MaxPooling2D((1, 2))(x)
    x = Conv2D(64, (1, 3), activation='relu', padding='same')(x)
    # Head
    x = GlobalAveragePooling2D()(x)
    out = Dense(n_outputs)(x)
    return Model(inp, out)

Key Decisions

• Temporal first: capture local time patterns before mixing features
• Asymmetric pooling: pool along the axis being convolved — (2,1) for time, (1,2) for features
• Fewer params: (3,1) + (1,3) has 6 params vs (3,3) with 9 — plus captures axis-specific patterns
• Generalizable: works for spectrograms, mel-frequency features, multi-sensor grids

References

• Source: host-starter-solution
• Competition: NeurIPS - Ariel Data Challenge 2024