Collective dynamics of repeated inference in variational autoencoder rapidly find cluster structure

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Deep neural networks excel at identifying low-dimensional latent spaces within high-dimensional data.
• Variational autoencoders (VAEs) are deep generative models capable of generating high-quality datasets and inferring underlying features.
• VAEs can denoise images through iterative mapping between latent and data spaces.

Purpose of the Study:

• To elucidate the denoising mechanism in VAEs by analyzing network activity patterns in latent space during inference.
• To investigate the collective behavior of inference trajectories for diverse datasets.

Main Methods:

• Numerical analysis of trained network activity patterns during inference.
• Tracking time development of activity patterns as trajectories within the latent space.
• Examining the influence of dataset cluster structure and noise levels on trajectory behavior.

Main Results:

• Inference trajectories rapidly converge to the center of existing data clusters.
• Increased data noise causes trajectories to approach more global clusters.
• Enhancing the number of latent variables improves cluster convergence and VAE generalization.

Conclusions:

• VAE denoising behavior mirrors concept retrieval in associative memory models.
• Latent space dynamics reveal a mechanism for noise reduction and feature extraction.
• Optimizing latent space dimensionality is key to improving VAE performance and generalization.