Adapting the Biological SSVEP Response to Artificial Neural Networks

TL;DR

A novel approach to neuron significance assessment inspired by frequency tagging is introduced, which holds promise for applications in network pruning, and model interpretability, contributing to the advancement of explainable artificial intelligence and addressing the lack of transparency in neural networks.

Abstract

Neuron importance assessment is crucial for understanding the inner workings of artificial neural networks (ANNs) and improving their interpretability and efficiency. This paper introduces a novel approach to neuron significance assessment inspired by frequency tagging, a technique from neuroscience. By applying sinusoidal contrast modulation to image inputs and analyzing resulting neuron activations, this method enables fine-grained analysis of a network's decision-making processes. Experiments conducted with a convolutional neural network for image classification reveal notable harmonics and intermodulations in neuron-specific responses under part-based frequency tagging. These findings suggest that ANNs exhibit behavior akin to biological brains in tuning to flickering frequencies, thereby opening avenues for neuron/filter importance assessment through frequency tagging. The proposed method holds promise for applications in network pruning, and model interpretability, contributing to the advancement of explainable artificial intelligence and addressing the lack of transparency in neural networks. Future research directions include developing novel loss functions to encourage biologically plausible behavior in ANNs.

Figures (5)

• Figure 1: The proposed frequency tagging approach for a color image. Its left half is tagged (i.e. contrast-modulated) with 6 Hz whereas the right half is tagged with 7.5 Hz. The tagged images are sequentially fed into an arbitrary (convolutional) neural network, in the order of their tagging, and the neuron's activation responses are acquired across time. The temporal sequence is then transformed into the frequency domain, with symbolic amplitudes illustrating the network's response strength at specific frequencies.
• Figure 2: Illustration of the proposed neuron significance assessment process.
• Figure 3: Illustration of the frequencies employed for SNR calculation. Baseline = Mean of Amplitudes $\{f_{i-2},f_{i-1}, f_{i+1}, f_{i+2}\}$. SNR of $f_i$ = Amplitude of $f_i$ / Baseline.
• Figure 4: The average frequencies across the 100 test images, obtained for the 16 filters of layers 1 and 10. Each color per bar corresponds to the magnitude of the frequency response for one of the filters of each of the layers. Along the x-axis: in red are the harmonics corresponding to the left-half of the images, in green the harmonics corresponding to the right-half of the images, and in blue are the intermodulation frequencies. Markers above some of the bars indicate frequency responses that surpass the chart's y-axis limit, which is set at 10 for visualization consistency.
• Figure 5: Numbers of important filters per layer for the ResNet-32 according to the frequency tagging approach.