acceptodds
Under review as a conference paper at ICLR 2027

Langevin Temporal Predictive Coding: From Point Estimates to Posterior Distributions

Abstract

Temporal predictive coding (tPC) is a model of sequential memory with local inference and Hebbian learning. Its inference, however, is deterministic. At each time step it propagates a point estimate of the latent state rather than a posterior distribution. We add Langevin noise whose scale is set by the fluctuation–dissipation theorem (FDT). The resulting Langevin tPC samples the conditional posterior at each time step. A Langevin step has two hyperparameters, the preconditioner and the step size . We compare two preconditioners, the identity matrix and the transition covariance , and relate them through the effective step size. Because latent directions relax at different rates, we also propose a per-direction step size that shrinks as curvature grows. On synthetic targets, the energy and the KL divergence decrease as the FDT predicts. Under anisotropy, the per-direction step size reaches lower KL divergence than identity preconditioning. Langevin tPC also represents multimodal posteriors, which MAP inference collapses to a single mode. We then evaluate our method on twelve regression datasets with Gaussian-mixture priors, where the exact posteriors are known and ten of them are multimodal. Across all twelve datasets, Langevin tPC approximates the exact posterior more closely than an ensemble of MAP estimates. On five video benchmarks, where posteriors are unimodal, it matches standard tPC. Adding noise thus extends tPC from point estimates to posterior distributions, and from sequential memory to probabilistic inference.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Temporal Predictive Coding, Langevin Noise, Flutiation Dissapation Theroem

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_oiyoOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper extends temporal predictive coding (tPC) from deterministic point estimation to Langevin-based posterior sampling using local prediction-error updates. It proves that the corresponding SDE has the per-step conditional posterior as its stationary distribution, analyzes finite-step behavior and step-size/preconditioning choices, and studies posterior mass recovery in multimodal settings. Experiments on synthetic targets, UCI-style inverse problems, temporal mixture systems, and video benchmarks show improved posterior approximation over deterministic ensembles in multimodal cases, while roughly matching standard tPC on unimodal video tasks.

Critical strengths
  1. Careful theoretical and empirical analysis. The paper does more than introduce a noisy variant of tPC: it analyzes stationary behavior, finite-step effects, anisotropy, mode mass, mixing limits, and failure regimes. The use of exact posteriors in several experiments makes the probabilistic claims easier to evaluate directly.
  2. Clear diagnosis of posterior mass versus deterministic basin occupancy. A useful conceptual contribution is the distinction between deterministic multi-start tPC, which reflects initialization-basin frequencies, and Langevin chains, which can recover posterior mass when mixing is feasible. This gives a clear explanation for why deterministic ensembles can discover multiple modes yet still represent their probabilities incorrectly.
Critical weaknesses
  1. The probabilistic inference claim is narrower than the overall framing may suggest. The method samples a frozen-parent per-step conditional rather than the filtering posterior required for full sequential Bayesian inference. The authors acknowledge this distinction and justify it through locality, but it still limits how broadly the method can be interpreted as a temporal posterior inference method.
  2. The methodological novelty appears somewhat incremental relative to existing Monte Carlo predictive coding. The paper itself notes that FDT stationarity is classical and that the per-position Langevin update is closely related to the MCPC E-step. The temporal extension, finite-step analysis, and posterior-mass study are useful, but the contribution would be more convincing with a clearer separation from existing sampling-based predictive-coding methods and stronger matched-setting comparisons.
Questions for authors
  1. Clarification: Could the authors more clearly characterize the probabilistic object represented by the chain population after multiple sequence steps? In particular, under what conditions, if any, does it approximate the filtering posterior rather than only the frozen-parent per-step conditional?
  2. Clarification / comparison: Given the acknowledged connection to MCPC and related sampling-based predictive-coding methods, what should be regarded as the main methodological novelty of Langevin tPC? Would the main conclusions remain under a direct matched-budget comparison with these existing samplers?
Reviewer_imoUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds Langevin noise to temporal PC; so inference produces samples of hidden states instead of one estimate. Each sampling run holds the previous state fixed. The authors study how different step sizes and computation budgets affect the samples, then they compare sampling with deterministic restarts on synthetic tasks, regression, time-series, and video. They report improved uncertainty estimates in some settings, with slow sampling and choice of model limitations.

Critical strengths
  1. The paper explains why running a deterministic method from many different starting points can find multiple answers without assigning them the probabilities supported by that model and data (Prop 4). This is important when using ensembles to calculate uncertainty.

  2. I appreciate the failure cases and controls. In Appendix K they separate (a) errors in the probabilities assigned to different outcomes from (b) errors within each outcome, which helps to explain what sampling actually changes.

Critical weaknesses
  1. The experiments do not clearly isolate the benefit for time series data. My understanding is that the 12 UCI test cases disable connections between time steps, so they test static inference (see section 5.4). Figure 4 changes both the inference procedure and the model used to generate the next state, which makes it difficult to determine how much of the gains come from actual sampling. Finally, Figure 9 includes a control that uses the same mixture model; I think that comparison should be quantified based on equal computation before drawing any conclusions about benefits of sampling.

  2. The cost comparison seems to leave out and ignore the work required by the proposed step size rule. I believe it needs a prior run, but this seems to be excluded from the sampling budget (Appendix D). For moving MNIST experiment, the rule also matches the target distribution worse than the sampler without it at every verified budget (Table 7). This is confusing because the other benchmarks use the sampler without that rule. So overall, these results leave its practical benefits unclear to me.

Questions for authors
  1. Clarification: With the same branching model and same total computation, how much does sampling improve the probabilities assigned to each future compared to deterministic restarts?

  2. Additional evidence: Does the proposed step size rule help on time-series data when the prior/initial run is included in the budget and when each baseline is tuned individually?

  3. Clarification: My understanding is that Prop 5 claims logarithmic growth in the number of updates with the problem size; but how does that count factor in the smaller step sizes required by its error bound?

Comments on the paper

1 comment

Thank you for being open about the issues, for correcting Prop 5, and for adding the time-series tests, which strengthen the uncertainty results. However, the UCI tests remain static, the pachinko branching gain comes from the transition model, and the step size rule offers no consistent benefit after accounting for the extra computation. It seems that an existing sampler with the same temporal model can give similar results at the largest tested budget. So I still have some reservations about how much the method actually adds in the end. Nonetheless, I am raising my score to 3 in light of the corrections and the added temporal evidence, with the expectation that the revised paper states the limitations more clearly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.