Prototype-Based Interpretation of the Functionality of Neurons in Winner-Take-All Neural Networks

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Prototype-based learning (PbL) with Euclidean distance (ED-WTA) is an intuitive multiclass classification method.
• PbL offers better interpretability and generalization than hyperplane-based learning (HbL).
• Existing ED-WTA methods can lead to unintuitive network centers.

Purpose of the Study:

• To prove the representational equivalence of ED-WTA and inner product-based WTA (IP-WTA).
• To propose a novel ±ED-WTA network with positive and negative prototypes.
• To develop a training algorithm for interpretable prototypes and analyze neuron functionality.

Main Methods:

• Proving the representational equivalence between IP-WTA and ED-WTA.
• Introducing the ±ED-WTA model with dual prototypes per neuron.
• Developing a novel training algorithm that updates positive and negative prototypes.
• Analyzing the similarity between positive and negative prototypes.

Main Results:

• Demonstrated the equivalence of IP-WTA and ED-WTA.
• Proposed ±ED-WTA, which models neurons with positive and negative prototypes.
• Developed a training algorithm enabling interpretable prototypes.
• Observed unexpected similarity between positive and negative prototypes, aligning with BCM theory.
• Showcased ±ED-WTA's effectiveness in explaining deep neural network (DNN) functionality and detecting outliers/adversarial examples.

Conclusions:

• The ±ED-WTA method provides highly interpretable prototypes for DNNs.
• Neuron functionality in ±ED-WTA can be interpreted as computing the difference between distances to positive and negative prototypes.
• ±ED-WTA enhances outlier and adversarial example detection.