Poisson Variational Autoencoder
Author Information
Author(s): Vafaii Hadi, Galor Dekel, Yates Jacob L.
Primary Institution: UC Berkeley
Hypothesis
The Poisson Variational Autoencoder (𝒫-VAE) integrates Poisson-distributed latent variables into variational autoencoders to better mimic biological neural coding.
Conclusion
The 𝒫-VAE provides a more interpretable computational framework for studying brain-like sensory processing and demonstrates improved sample efficiency in classification tasks.
Supporting Evidence
- The 𝒫-VAE encodes inputs in higher dimensions, improving linear separability.
- It largely avoids the posterior collapse issue seen in traditional VAEs.
- The model demonstrates 5× better sample efficiency in classification tasks compared to standard VAEs.
Takeaway
The 𝒫-VAE is a new type of model that helps computers understand images like our brains do, using a special way of counting spikes instead of regular numbers.
Methodology
The study developed a novel architecture that combines predictive coding principles with a VAE, using Poisson-distributed latent variables and a reparameterization trick for inference.
Limitations
The model's reliance on Poisson sampling may not fully capture the complexities of neuronal activity over longer time windows, and there is an existing amortization gap compared to traditional sparse coding.
Statistical Information
P-Value
p<0.01
Confidence Interval
99%
Statistical Significance
p<0.01
Want to read the original?
Access the complete publication on the publisher's website