Skip to content

0001. Deterministic forecasting, not variational

Status: accepted Date: 2026-05-03

Context

Brain-state modeling work occasionally proposes variational latent-state formulations (e.g., variational dynamical models, VAE-style encoders for fMRI) on the grounds that brain states are "uncertain" and uncertainty should be modeled explicitly.

We need to choose: do we train BOLDcast as a deterministic forecaster (p(x_{t+k} | x_{≤t}, s) parameterized as a point estimate), or do we add variational latent variables?

Decision

BOLDcast is a deterministic multi-step forecaster. Loss is MSE on the predicted next-TR token activations. No variational lower bound, no explicit per-step latent variable, no learned posterior.

Uncertainty estimates, if needed for downstream consumers, are obtained post-hoc via temperature-calibrated ensembling.

Consequences

  • Training is simpler: one objective (MSE forecasting) plus the InfoNCE contrastive head in phase 2. No KL term to balance, no posterior collapse failure mode, no reparameterization tricks to debug.
  • Reproducibility is higher: deterministic forward pass in eval mode given the same input.
  • We forfeit the ability to interpret latent activations as probabilistic brain states. We treat learned representations as features whose usefulness is judged on downstream task performance, not as a probabilistic state estimator.
  • Memory and compute footprint is lower (no encoder–decoder split, no sampling at training time).

Alternatives considered

  • Full variational latent dynamics (rSLDS-style or VAE-style): rejected because (a) the literature is unclear on whether brain-state uncertainty estimates from these models are well-calibrated, and (b) the additional complexity is not justified by the evaluations we plan (forecasting MSE, retrieval accuracy, downstream linear probing).
  • Deterministic forecaster with MC-dropout uncertainty: considered as a fallback for downstream tasks that need uncertainty. Cheap to add later if needed; not part of the core training objective.
  • Diffusion forecaster: out of scope for the seed-grant timeline. May be revisited if forecasting quality is the bottleneck for downstream tasks.

References

  • See docs/methods.md "Scope and Honest Limitations" for the public framing of this decision.