Knowing what you don’t know: Estimating the uncertainty of feedforward and feedback inputs with prediction-error circuits

Loreen Hertäg, Katharina A. Wilmes, Claudia ClopathbioRxiv 2023 · DOI

TL;DR

This paper proposes a neural network model that uses prediction-error (PE) neurons to estimate the uncertainty of both incoming sensory information and internal predictions. This allows the brain to dynamically adjust how much it relies on sensory input versus predictions, a mechanism that could explain phenomena like contraction bias and be influenced by neuromodulators.

What Problem Does This Paper Address?

Our brains are constantly bombarded with sensory information from the world, but they also generate internal predictions about what to expect. These two streams of information—feedforward sensory input and feedback predictions—often don't perfectly align due to noise in sensory perception or unexpected environmental changes. The challenge for the brain is to integrate these conflicting signals effectively, relying more on the more reliable source. While the idea of weighting inputs by their reliability (akin to Bayesian inference) is well-established behaviorally, the neural mechanisms by which the brain estimates the uncertainty of both sensory inputs and predictions at a circuit level remain poorly understood. This paper aims to bridge this gap by proposing a specific neural circuit architecture.

Methods

The authors propose a hierarchical network model built around prediction-error (PE) circuits. These circuits are composed of negative PE (nPE) and positive PE (pPE) neurons, which respond when sensory input is weaker or stronger than predicted, respectively.

Key methodological components include:

  • Hierarchical PE Circuits: The model uses two nested PE circuits. A lower-level circuit processes sensory input, and its output (mean and variance estimates) is fed to a higher-level circuit that processes predictions.
  • Memory (M) and Variance (V) Neurons: Within each PE circuit, a "memory" neuron integrates PE neuron activity to estimate the mean of the input, while a "variance" neuron, with a quadratic activation function, estimates the variance (uncertainty) of the input.
  • Neuromodulation Simulation: The influence of neuromodulators is simulated by injecting additional excitatory input into specific types of inhibitory interneurons (PV, SOM, VIP) within the PE circuits.
  • Weighted Output Calculation: The final output is a weighted sum of the sensory input and the prediction, where the weights are dynamically determined by the estimated variances from the V neurons.

Datasets

Not discussed in the paper. The model is stimulated with synthetic inputs generated according to specified distributions (uniform, normal, gamma, binary) and parameters.

Key Insights and Ideas

  • PE Neurons as Uncertainty Estimators: The core idea is that the distinct firing patterns of nPE and pPE neurons, when integrated by dedicated "variance" neurons, can naturally compute the uncertainty (variance) of both sensory inputs and internal predictions.
  • Dynamic Weighting: The model demonstrates that the brain can dynamically adjust its reliance on sensory input versus predictions based on their estimated uncertainties. For instance, it relies more on predictions when sensory input is noisy and the environment is stable.
  • Neuromodulation as a Bias Mechanism: The paper shows how modulating specific inhibitory interneurons can bias the weighting of sensory inputs and predictions. This provides a potential neural basis for how neuromodulators influence perception and decision-making.
  • Contraction Bias Explained: The model links the weighted integration of sensory and predictive information to the "contraction bias" (the tendency to perceive stimuli as closer to the mean of a distribution). This bias is shown to be influenced by stimulus uncertainty and trial duration.

Key Results

  • The model successfully estimates the mean and variance of sensory stimuli using the activity of M and V neurons, respectively, across various input distributions.
  • The network exhibits flexible adaptation: it relies more on predictions when sensory input is noisy and the environment is stable (high trial variability).
  • Predictions are weighted more strongly at the onset of a new stimulus, even if the stimulus is reliable.
  • Modulating PV interneurons biases the network towards predictions, while modulating VIP interneurons leads to more equal weighting of sensory inputs and predictions. SOM interneurons have context-dependent effects.
  • The model reproduces the contraction bias, showing that it increases with stimulus uncertainty and decreases with prediction uncertainty or longer trial durations.

Limitations and Future Work

  • Simplified Neuron Models: The model uses simplified rate-based neuron models and point neurons for interneurons, abstracting away from more complex biophysical details.
  • Global Neuromodulation: Neuromodulators are modeled as global inputs to interneurons, whereas their effects in the brain can be more localized and specific.
  • No Explicit Learning: The model assumes a pre-trained network and does not explicitly model the learning mechanisms by which the PE circuits and their connections are formed or updated.
  • Future Work: The authors suggest investigating how neuromodulators might gate plasticity, exploring different neural implementations for integration across species and modalities, and developing more detailed models of interneuron interactions and their impact on weighting. They also highlight the need for experimental validation of their predictions regarding interneuron stimulation.

Summary updated

Abstract

Abstract At any moment, our brains receive a stream of sensory stimuli arising from the world we interact with. Simultaneously, neural circuits are shaped by feedback signals carrying predictions about the same inputs we experience. Those feedforward and feedback inputs often do not perfectly match. Thus, our brains have the challenging task of integrating these conflicting streams of information according to their reliabilities. However, how neural circuits keep track of both the stimulus and prediction uncertainty is not well understood. Here, we propose a network model whose core is a hierarchical prediction-error circuit. We show that our network can estimate the variance of the sensory stimuli and the uncertainty of the prediction using the activity of negative and positive prediction-error neurons. In line with previous hypotheses, we demonstrate that neural circuits rely strongly on feedback predictions if the perceived stimuli are noisy and the underlying generative process, that is, the environment is stable. Moreover, we show that predictions modulate neural activity at the onset of a new stimulus, even if this sensory information is reliable. In our network, the uncertainty estimation, and, hence, how much we rely on predictions, can be influenced by perturbing the intricate interplay of different inhibitory interneurons. We, therefore, investigate the contribution of those inhibitory interneurons to the weighting of feedforward and feedback inputs. Finally, we show that our network can be linked to biased perception and unravel how stimulus and prediction uncertainty contribute to the contraction bias.
BibTeX
@misc{hertag2023knowing,
  author = {Loreen Hertäg and Katharina A. Wilmes and Claudia Clopath},
  title = {{Knowing what you don’t know: Estimating the uncertainty of feedforward and feedback inputs with prediction-error circuits}},
  year = {2023},
  doi = {10.1101/2023.12.13.571410},
  howpublished = {bioRxiv},
}
Figure