A Model of Predictive Perception via Recursive Bayesian Estimation
Poster No:
1084
Submission Type:
Abstract Submission
Authors:
Mahdi Enan1, Mario Senden1, Ryszard Auksztulewicz2, Federico de Martino1
Institutions:
1Maastricht University, Maastricht, Limburg, 2Free University Berlin, Berlin, Berlin
First Author:
Co-Author(s):
Introduction:
We live in complex noisy environments which cause the input to our brain to be uncertain, an effect that is further amplified by biological limitations in our senses. To interact meaningfully in such uncertain environments, our brains must infer the most likely causes of our perceptions. Furthermore, since the environment is rarely stable, our brains also have to flexibly adapt to environmental dynamics in order to generate meaningful predictions of upcoming sensory information. Here we present a formal predictive model that captures foundational computational principles underlying our brains' capability to dynamically adapt to the environment. By combining the Bayesian filtering with Hebbian-like update rules we provide a framework for exact inference using unsupervised local learning mechanisms. This unified framework allows simultaneous learning and inference by optimally combining sensory information and predictions.
Methods:
Our model is composed of a sensor compartment which transforms states into a probability distribution of measurements. The sensor can be a simple matrix or a neural network model. Our model keeps track of the environmental dynamics in a Markov matrix and updates it using previous state beliefs. This state dynamics matrix is capable of generating new predictions of upcoming states. A new state belief is formed using the recursive Bayes formula by combining measurements with predictions. Because of its modular implementation, our model enables testing different learning rules, sensory models and may optionally be used to incorporate actions in the inference process. This modularity also allows to implement a hierarchical version of the model that enables learning of higher-order statistics.
Results:
We highlight the key features of this model in 3 simulated experiments. In the first we show how the model predictions and beliefs are affected by stochasticity in the causes of the input as well as "background" noise (that could represent noise in the environment or noise introduced by the sensors). Increasing stochasticity in state changes within the environment, results in the model relying more on predictions to calculate the belief, whereas clean sensory information led to faster updating of the model dynamics. In the second simulation we reproduce classical effects by deviating local and global predictive rules (Chao et al., 2018). We show that both local and global (expectation driven) dynamics can be learned using a hierarchical implementation of the model. In the third simulation we highlight the ability of our model to infer non-Gaussian probability distributions of the states, which underscores the flexibility of our non-parametric approach.
·The proposed model can deal with different (including non-Gaussian) type of noise.
·In a local-global paradigm, when presented to a 2-Level Hierarchical model can reproduce local and global deviant effects.
Conclusions:
We argue that assuming the brain to operate on a discretized and restricted state space of the environment and considering sensory characteristics (e.g. receptive fields) and working memory allows exact inference. Our non-parametric model does not make specific assumptions about particular sufficient statistics and instead it only requires local weight updating. This approach enables more flexibility and adaptability for inference and prediction.
Learning and Memory:
Learning and Memory Other 2
Modeling and Analysis Methods:
Bayesian Modeling 1
Classification and Predictive Modeling
Methods Development
Keywords:
Computational Neuroscience
Learning
Machine Learning
Perception
1|2Indicates the priority used for review