A neurally plausible model for online recognition and postdiction in a dynamical environment

Li Kevin Wenliang, Maneesh Sahani

Perception is rarely instantaneous. To make sense of the world, the brain must constantly integrate a continuous stream of noisy, ambiguous sensory data. Often, the most informative data about an event arrives after the event has already occurred. This leads to a phenomenon known as postdiction: our brain effectively "rewrites" its recent perceptual history based on new evidence.

While mathematical tools like Kalman filters or particle filters can handle this retrospective updating (a process called "smoothing"), it remains unclear how the biological brain implements these complex calculations. Standard models often struggle to explain how a system operating in real-time can acausally propagate information backward to update past percepts.

This paper proposes a neurally plausible framework for real-time probabilistic inference that accounts for postdiction. By extending Distributed Distributional Codes (DDC) to dynamic environments, the authors demonstrate a model that can be trained via simple biological learning rules to replicate human perceptual illusions and track complex, nonlinear systems.

The Distributed Distributional Code (DDC) Framework

The researchers build upon the Distributed Distributional Code (DDC), a framework where a probability distribution is represented by the firing rates of a population of neurons. Each neuron's rate corresponds to the expected value of a specific "tuning function" (or encoding function).

To adapt this for dynamic, real-time perception, the authors introduce several key methodological innovations:

  • Dynamical Encoding Functions: They introduce a recurrence relationship that allows the neural population to represent not just the current state, but a history of latent states. This "memory" allows new observations to interact with and refine representations of the recent past.
  • Amortized Inference: Instead of performing computationally expensive Bayesian updates on the fly, the model learns a "recognition model"—a direct mapping from observations to the DDC representation.
  • Learning via Internal Models: The recognition model is trained using samples generated by the brain’s own "internal model" of the world. This allows the network to learn using a simple, biologically plausible delta rule (a form of error-correction learning), bypassing the need for complex backpropagation.

Key Insights: Postdiction as Optimal Smoothing

A central takeaway from this work is that postdiction is not a perceptual "glitch." Rather, it is a necessary feature of optimal Bayesian smoothing. By maintaining a DDC of the recent trajectory, the brain can refine its estimate of a past state ($t-1$) when new evidence at time $t$ arrives.

Furthermore, DDCs allow the brain to represent complex, non-Gaussian distributions (including multi-modal ones) using simple mean parameters like firing rates. This representational efficiency bypasses the need for the brain to "sample" from distributions in real-time. The framework also suggests that the brain trains its recognition circuits by "dreaming" or simulating scenarios from its internal model, providing a plausible explanation for how we develop such efficient perceptual shortcuts.

Experimental Results: Replicating Human Illusions

The model was evaluated on simulated environments designed to mimic classic psychophysics and tracking challenges.

1. The Auditory Continuity Illusion

This classic illusion tests whether a listener "fills in" a gap in a sound when that gap is masked by noise.

Simulation of the auditory continuity illusion, demonstrating how later noise or tone observations update the certainty of past perceptions (postdiction).

The model successfully reproduced this effect. When a simulated tone was interrupted by noise but followed by another tone, the model’s posterior belief indicated the tone likely continued through the noise. The model dynamically updated its certainty about the past based on subsequent auditory evidence.

2. The Flash-Lag Effect

In this simulation, a 1-D moving object reverses direction. The task measures how the perceived location of an object differs from its actual location during a flash.

Simulation of the flash-lag effect, showing the smoothed perceived location (black) compared to the actual reversing trajectory (red).

The DDC model matched human behavioral data perfectly. It demonstrated a smooth perceived trajectory during direction reversals, whereas simpler "extrapolation" models predicted a broken, jerky path.

3. Nonlinear Oscillatory Tracking

Finally, the model was tested on a synthetic 1-D image observation task involving stochastic oscillations, noise, and occlusions. The DDC model achieved robust tracking performance ($R^2 \approx 0.68$), comparable to a bootstrap particle filter. Crucially, as the "postdictive window" (the amount of future data used to look back) increased, the model’s uncertainty decreased and its accuracy improved.

Limitations and Future Directions

While the model offers a compelling bridge between machine learning and cognitive neuroscience, several questions remain:

  • Learning the Internal Model: The paper assumes the brain already possesses an accurate internal generative model of the world. It does not address how that generative model is learned in the first place (e.g., via "wake-sleep" algorithms).
  • Scaling to High Dimensions: The experiments were conducted in 1-D or low-dimensional spaces. Whether this DDC approach scales to the high-dimensional complexity of real-world visual or auditory scenes is an open question.
  • Biological Specificity: While the delta rule is plausible, the specific neural architecture requires further physiological investigation to see which mathematical updates most closely match actual cortical circuitry.

Summary updated

Abstract

Humans and other animals are frequently near-optimal in their ability to integrate noisy and ambiguous sensory data to form robust percepts---which are informed both by sensory evidence and by prior expectations about the structure of the environment. It is suggested that the brain does so using the statistical structure provided by an internal model of how latent, causal factors produce the observed patterns. In dynamic environments, such integration often takes the form of \\emph{postdiction}, wherein later sensory evidence affects inferences about earlier percepts. As the brain must operate in current time, without the luxury of acausal propagation of information, how does such postdictive inference come about? Here, we propose a general framework for neural probabilistic inference in dynamic models based on the distributed distributional code (DDC) representation of uncertainty, naturally extending the underlying encoding to incorporate implicit probabilistic beliefs about both present and past. We show that, as in other uses of the DDC, an inferential model can be learnt efficiently using samples from an internal model of the world. Applied to stimuli used in the context of psychophysics experiments, the framework provides an online and plausible mechanism for inference, including postdictive effects.
BibTeX
@inproceedings{wenliang2019neurally,
  author = {Li Kevin Wenliang and Maneesh Sahani},
  title = {{A neurally plausible model for online recognition and postdiction in a dynamical environment}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
}
Figure