Implicit Generative Modeling by Kernel Similarity Matching

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Understanding neural stimulus encoding is key to developing brain-like artificial neural networks (ANNs).
• Existing methods for representation learning often focus on bottom-up processing (stimuli to latent space).
• Generative paradigms, incorporating top-down processing (latent space to stimuli), remain less explored for ANNs.

Purpose of the Study:

• To investigate a kernel similarity matching framework for generative modeling in computational neuroscience.
• To explore representation learning that integrates both bottom-up and top-down processing.
• To develop a biologically plausible architecture for ANNs.

Main Methods:

• Utilized a modified sparse coding objective for representation learning.
• Demonstrated equivalence between representation learning and maximizing kernel similarity (input vs. latent).
• Proposed a novel alternating direction method of multipliers (ADMM) algorithm for optimization.

Main Results:

• Showed that representation learning is equivalent to maximizing kernel similarity.
• Developed an implicit generative model by learning latent space kernel structure.
• Adapted the framework to learn manifold structures, offering insights into brain encoding.

Conclusions:

• The kernel similarity matching framework enables generative modeling with brain-like learning.
• The approach integrates similarity matching (bottom-up) with predictive coding (top-down) for biological plausibility.
• This work provides a foundation for ANNs that learn representations similarly to the brain.