Statistically optimal perception and learning: from behavior to neural representations

By József Fiser, Pietro Berkes, Gergo Orbán, Máté Lengyel. Trends in Cognitive Sciences, 2010

727 citations recorded by the source corpus.

Human perception has recently been characterized as statistical inference based on noisy and ambiguous sensory inputs. Moreover, suitable neural representations of uncertainty have been identified that could underlie such probabilistic computations. In this review, we argue that learning an internal model of the sensory environment is another key aspect of the same statistical inference procedure and thus perception and learning need to be treated jointly. We review evidence for statistically optimal learning in humans and animals, and reevaluate possible neural representations of uncertainty based on their potential to support statistically optimal learning. We propose that spontaneous activity can have a functional role in such representations leading to a new, sampling-based, framework of how the cortex represents information and uncertainty.

Subfields: Cognitive Neuroscience, Experimental and Cognitive Psychology

View Paper: https://www.semanticscholar.org/paper/f9b2c708d4f172916c5b6de461b9d8e9c897ae9d

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

In Experimental and Cognitive Psychology

Citations in the corpus

Theory of cortical function (2017, background): “The noisedriven process of exploration was motivated by stochastic optimization algorithms (57) and is similar to models of sensory neuroscience that draw samples from an underlying probability distribution over possible percepts (9, 10, 39, 43).”
Training recurrent networks to generate hypotheses about how the brain solves hard navigation problems (2016, background): “Unlike the particle ﬁlter and contrary to neural models that implement probabilistic inference by stochastic sampling of the underlying distribution [24], this network implements ongoing near-optimal probabilistic location estimation through a fully deterministic dynamics.”
Representation of uncertainty in deep neural networks through sampling (2016, background): “The neural sampling hypothesis states that the activity patterns of biological neural networks represent samples from a learned posterior distribution over interpretations given an input (Fiser et al., 2010).”
Presynaptic stochasticity improves energy efficiency and helps alleviate the stability-plasticity dilemma (2021, background): “Inspired by a Bayesian perspective, neural variability is often interpreted as representing uncertainty (Ma et al., 2006; Fiser et al., 2010; Kappel et al., 2015; Haefner et al., 2016), or as a means to prevent overfitting (Wan et al.”
Cognitive computational neuroscience (2018, unknown): “A third and critically important class of cognitive model is that of Bayesian models (Box 3).(33,35,36,34,143,144,145,32) Bayesian inference provides an essential normative perspective on cognition.”
Metacognition: computation, biology and function (2012, background): “These ideas are potentially consistent with recent neural coding models, which propose that patterns of ﬁring in cortex encode parameters of the probability distribution that is the result of probabilistic inference, i.e. both its mean and covariance [25].”
Probabilistic brains: knowns and unknowns (2013, background): “Finally, some groups have proposed that the brain may represent probability distributions by the values of a set of samples drawn from the encoded distribution 34,36,53,54 .”
Bayesian reasoning with ifs and ands and ors (2015, background): “It stands in accordance with ﬁndings on good conformity to Bayesian principles in domains outside of reasoning, where tasks are carried out in a more implicit way, like perception and language comprehension (Fiser et al., 2010; Hsu et al., 2011).”
When the world becomes 'too real': a Bayesian explanation of autistic perception. (2012, background): “Others have suggested that probabilistic perception and learning should be better implemented with sampling-based approaches, whereby single neurons map on to inferred variables and uncertainty is represented by the variability of neural activity patterns [44].”
Bayesian sampling in visual perception (2011, background): “Although this probability distribution may be difficult to compute explicitly, one can often design algorithms to generate samples from the distribution, allowing one to perform approximate inference (12, 13).”
The Hamiltonian Brain: Efficient Probabilistic Inference with Excitatory-Inhibitory Neural Circuit Dynamics (2014, methodology): “In particular, our network uses a sampling-based representation of uncertainty [25, 35, 11], such that at any time it represents a single plausible interpretation of the input, and as time passes it sequentially samples many different interpretations.”
Neural Dynamics as Sampling: A Model for Stochastic Computation in Recurrent Networks of Spiking Neurons (2011, background): “, by explicitly adding a random variable for the stimulus intensity [35,51]), which is not the focus of the presented work.”
Deep neural networks: a new framework for modelling biological vision and brain information processing (2015, methodology): “…coding (Barlow 1961, Olshausen & Field 1997, Simoncelli & Olshausen 2001), probabilistic neuronal coding and inference (Hoyer & Hyvärinen 2003, Fiser et al. 2010, Buesing et al. 2011, McClelland 2013, Pouget et al. 2013), Bayesian sensorimotor control (Körding & Wolpert 2006), and…”
Probability Matching as a Computational Strategy Used in Perception (2010, background): “Sampling-based representational coding has been proposed to account for neurophysiological phenomena such as spontaneous neural activity [35] and variability in neural responses [36], as well as other stochastic perceptual phenomena such as bistability [37,38].”
Theory learning as stochastic search in the language of thought (2012, unknown): “We have not discussed at all the level of neural implementation, but recent proposals by a number of authors argue that analogous stochastic-sampling ideas could plausibly be used to carry out Bayesian learning in the brain (Fiser et al., 2010).”
Correlations and Neuronal Population Information (2016, background): “A final example is the proposal that early cortical processing relies on probabilistic generative models of the natural sensory environment (Barlow 2001, Dayan & Abbott 2001, Simoncelli & Olshausen 2001, Fiser et al. 2010, Berkes et al. 2011).”
Self-Evaluation of Decision-Making: A General Bayesian Framework for Metacognitive Computation (2017, unknown): “Future efforts could harness our framework to guide construction of finer-grained Bayesian models incorporating temporal dynamics or candidate neural network implementations (Beck et al., 2008; Fiser, Berkes,”