Metabolic cost of information processing in Poisson variational autoencoders

The article introduces the Poisson Variational Autoencoder (P-VAE), which incorporates a metabolic cost term linked to neuron firing rates into its variational free energy minimization framework. This model uniquely couples coding rate with firing rate, allowing for a trade-off between coding fidelity and energy expenditure, contrasting with standard Gaussian VAEs. The findings highlight the P-VAE's potential for developing energy-aware computational theories, making it relevant for practitioners focused on efficient resource management in AI models.