Variational Autoencoder Layer

Research Overview

This paper explores a departure from the traditional use of Variational Autoencoders (VAEs) as standalone generative models. Instead, the author proposes integrating VAE components directly into a neural network as a specialized layer, termed 'VAELinear.' This approach aims to leverage the probabilistic nature of VAEs within a standard feed-forward architecture, moving away from conventional backpropagation-based training.

Methodology

The VAELinear layer replaces standard linear transformations with a structure that includes a VAE-based transformation and a normalization layer. To train these layers without backpropagation, the author employs the Improved Forward-Forward Contrastive Learning (IFFCL) framework. This method trains layers independently by using the output of one network as a reference for the corresponding layer in another, utilizing a multimodal VAE (MVAE) objective function. In this setup, the decoder is simplified to an identity mapping, effectively treating the latent representation as the reconstruction.

Findings and Performance

Experiments on the MNIST dataset demonstrate that the VAELinear layers can successfully learn meaningful data representations. However, the overall classification performance is limited, peaking at approximately 55% accuracy. The study notes that while the VAE-based layers remain stable during training, the final classification layer experiences performance degradation over time, suggesting that the current integration strategy requires further refinement to achieve competitive results.