Poisson Variational Autoencoder.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Variational autoencoders (VAEs) model Bayesian inference in sensory processing, similar to primate vision.
• Traditional VAEs use continuous latent variables, unlike discrete biological neurons.

Purpose of the Study:

• Develop a novel Variational Autoencoder (VAE) architecture using discrete latent variables.
• Investigate brain-like sensory processing and perception as an inferential process.

Main Methods:

• Introduced the Poisson VAE (P-VAE) combining predictive coding with discrete spike count encoding.
• Incorporated a metabolic cost term into the P-VAE loss function, linking to sparse coding.
• Analyzed the geometry of learned representations and compared P-VAE to other VAE models.

Main Results:

• Empirically verified the relationship between the metabolic cost term and sparse coding.
• Found P-VAE learns higher-dimensional representations, improving linear separability.
• Achieved 5x greater sample efficiency in downstream classification tasks compared to alternative VAEs.

Conclusions:

• The P-VAE offers an interpretable computational framework for studying brain-like sensory processing.
• This model advances the understanding of perception as an inferential process.
• Discrete latent variables in VAEs offer advantages for biological plausibility and learning efficiency.