BMBISAMJOHNSON

Biological Mechanisms for Bayesian Inference

 Partecipanti

Mappa

 Word cloud

Esplora la "nuvola delle parole (Word Cloud) per avere un'idea di massima del progetto.

 Obiettivo del progetto (Objective)

'A growing body of experimental evidence suggests that animals, and in particular humans, are capable of inferring knowledge about reality from uncertain or incomplete data in a way that is, according to Bayes’ theorem, mathematically optimal [1, 2]. The case that the brain is, at some level, a Bayesian inference machine was made much stronger when Ma et al. [3] recently described a mechanism whereby a neural network could indeed store and manipulate probability distributions – a mechanism that reduces Bayes’ theorem to a sum [4]. We shall bring together tools and recent advances from in the fields of Complex Networks [5] and Computational Neuroscience [6] with a view to: a) designing methods and algorithms for tasks such as grammar inference or network routing; b) putting forward neural network models which are capable of integrating Bayesian inference with other necessary brain functions, such as working memory or information processing; and c) exploring, in greater detail, how Bayesian inference could be carried out in realistic biological settings.

References [1] M.O. Ernst, M.S. Banks, N. Models, and S. Thresholds, Humans integrate visual and haptic information in a statistically optimal fashion, Nature, 415, 429-33 (2002). [2] T. Yang and M.N. Shadlen, Probabilistic reasoning by neurons, Nature, 447, 1075-80 (2007). [3] W.J. Ma, J.M. Beck, P.E. Latham, and A. Pouget, Bayesian inference with probabilistic population codes, Nature Neurosci., 9, 1432-8 (2006). [4] J.M. Beck, W.J. Ma, R. Kiani, T. Hanks, A.K. Churchland, J. Roitman, M.N. Shadlen, P.E. Latham, and A. Pouget, Probabilistic population codes for Bayesian decision making, Neuron, 60, 1142-52(2008). [5] S. Johnson, J.J. Torres, J. Marro, and M.A. Muñoz, Entropic origin of disassortativity in complex networks, Phys. Rev. Lett.., 104, 108702 (2010) [6] S. Johnson, J. Marro, and J.J. Torres, Cluster Reverberation: a mechanism for robust working memory without synaptic learning, submitted.'

Introduzione (Teaser)

According to Bayes' theorem, introduced by Thomas Bayes and presented in a paper in 1763, the probability of a certain state existing or being true can be updated according to new data. Computational models have shed light on possible neural mechanisms of Bayesian inference.

Descrizione progetto (Article)

A growing body of evidence suggests that animals and particularly humans are capable of mathematically optimal Bayesian inference and that the brain is a sort of Bayesian inference machine. Recent work describing how a neural network could store and manipulate probability distributions lends further support to the hypothesis.

The EU-funded project 'Biological mechanisms for Bayesian inference' (BMBISAMJOHNSON) set out to investigate the topic, developing neural network models in which Bayesian inference is integrated with other brain functions such as working memory or information processing.

Gradual strengthening of the connections (synapses) between two neurons is the neurobiological substrate of long-term memory as, for example, in learning to ride a bike. However, many cognitive functions, including inference, short-term memory and sensory memory, occur on much smaller timescales.

The team developed the concept of cluster reverberations in which small groups of neurons receive signals from each other to account for many of the characteristics of such functions. Researchers suggested a mechanism by which it could support optimal human inference from uncertain sensory information.

Scientists went on to conduct numerous other mathematical investigations of topics covering very diverse fields of interest. From plant-pollinator networks and financial institutions to sampling bias in public health research to a hotly debated topic regarding food web stability, researchers developed new concepts and identified novel network properties that explain how the world works.

The methods and algorithms developed along with the emerging network properties that have been identified will find broad application in mathematics, economics, biology and medicine.