The Neuron as a Latent State: Classical Variational Readout in Distributional Neural Units

Abstract

Classical neural networks usually represent hidden units as deterministic scalar activations. This makes computation easy to compose, but it collapses each unit’s internal state into a single value, so uncertainty is usually modeled outside the hidden unit through Bayesian weights, ensembles, output distributions, or post-hoc predictive scores. We study an alternative view with EVE, a classical distributional neural mechanism in which hidden computation carries a local variational latent state. This state is parameterized by a posterior mean and variance, sampled through reparameterization, and regularized by a local Kullback–Leibler divergence. We focus not only on whether such a state can be learned, but on how it should be read. To isolate this question, we introduce classical variational readout and use a posterior-freezing protocol: after training EVE, the learned posterior is fixed and only the readout rule is varied. On a controlled multimodal forecasting benchmark, the frozen EVE posterior represents multiple plausible futures more effectively than a deterministic baseline. A non-oracle branch-medoid readout further improves CRPS and pinball loss, removes invalid samples, improves branch balance, and preserves coverage and best-of-sample accuracy without posterior retraining. These results support the interpretation of EVE as an information-bearing latent computational state rather than merely stochastic noise or regularization.

Keywords

• variational neural networks
• distributional neurons
• latent-state computation
• uncertainty estimation
• probabilistic readout
• Monte Carlo prediction
• neural uncertainty

Citation

Ruffenach, Yves. “The Neuron as a Latent State: Classical Variational Readout in Distributional Neural Units.” Research Square, June 30, 2026. Research Article. DOI: 10.21203/rs.3.rs-10177882/v1.

Related research

• Variational Distributional Neuron
• The Neuron Is the Distribution
• Distributional Neurons: Making Uncertainty a Unit of Computation
• Measuring and Controlling Internal Activity in Variational Neural Units
• Variational Neurons in Transformers for Language Modeling

Variational neuron · Variational Transformer · Internal uncertainty · Variational closure