Metabolic cost of information processing in Poisson variational autoencoders

Researchers have developed a Poisson Variational Autoencoder (P-VAE) that directly links information processing to metabolic cost, mimicking how biological brains balance accuracy with energy use. Unlike standard models, the P-VAE's structure naturally penalizes high neural activity, creating an emergent 'sparse coding' system. Tests show increasing the model's KL divergence term (β) systematically reduces spiking activity by up to 40%, while comparable Gaussian models show no change, proving the efficiency gain is unique to this brain-like architecture.