Sensor fusion in distributed cortical circuits

Firouzi, Mohsen (2020): Sensor fusion in distributed cortical circuits. Dissertation, LMU München: Graduate School of Systemic Neurosciences (GSN)

Abstract

The substantial motion of the nature is to balance, to survive, and to reach perfection. The evolution in biological systems is a key signature of this quintessence. Survival cannot be achieved without understanding the surrounding world. How can a fruit fly live without searching for food, and thereby with no form of perception that guides the behavior? The nervous system of fruit fly with hundred thousand of neurons can perform very complicated tasks that are beyond the power of an advanced supercomputer. Recently developed computing machines are made by billions of transistors and they are remarkably fast in precise calculations. But these machines are unable to perform a single task that an insect is able to do by means of thousands of neurons. The complexity of information processing and data compression in a single biological neuron and neural circuits are not comparable with that of developed today in transistors and integrated circuits. On the other hand, the style of information processing in neural systems is also very different from that of employed by microprocessors which is mostly centralized. Almost all cognitive functions are generated by a combined effort of multiple brain areas. In mammals, Cortical regions are organized hierarchically, and they are reciprocally interconnected, exchanging the information from multiple senses. This hierarchy in circuit level, also preserves the sensory world within different levels of complexity and within the scope of multiple modalities. The main behavioral advantage of that is to understand the real-world through multiple sensory systems, and thereby to provide a robust and coherent form of perception. When the quality of a sensory signal drops, the brain can alternatively employ other information pathways to handle cognitive tasks, or even to calibrate the error-prone sensory node. Mammalian brain also takes a good advantage of multimodal processing in learning and development; where one sensory system helps another sensory modality to develop. Multisensory integration is considered as one of the main factors that generates consciousness in human. Although, we still do not know where exactly the information is consolidated into a single percept, and what is the underpinning neural mechanism of this process? One straightforward hypothesis suggests that the uni-sensory signals are pooled in a ploy-sensory convergence zone, which creates a unified form of perception. But it is hard to believe that there is just one single dedicated region that realizes this functionality. Using a set of realistic neuro-computational principles, I have explored theoretically how multisensory integration can be performed within a distributed hierarchical circuit. I argued that the interaction of cortical populations can be interpreted as a specific form of relation satisfaction in which the information preserved in one neural ensemble must agree with incoming signals from connected populations according to a relation function. This relation function can be seen as a coherency function which is implicitly learnt through synaptic strength. Apart from the fact that the real world is composed of multisensory attributes, the sensory signals are subject to uncertainty. This requires a cortical mechanism to incorporate the statistical parameters of the sensory world in neural circuits and to deal with the issue of inaccuracy in perception. I argued in this thesis how the intrinsic stochasticity of neural activity enables a systematic mechanism to encode probabilistic quantities within neural circuits, e.g. reliability, prior probability. The systematic benefit of neural stochasticity is well paraphrased by the problem of Duns Scotus paradox: imagine a donkey with a deterministic brain that is exposed to two identical food rewards. This may make the animal suffer and die starving because of indecision. In this thesis, I have introduced an optimal encoding framework that can describe the probability function of a Gaussian-like random variable in a pool of Poisson neurons. Thereafter a distributed neural model is proposed that can optimally combine conditional probabilities over sensory signals, in order to compute Bayesian Multisensory Causal Inference. This process is known as a complex multisensory function in the cortex. Recently it is found that this process is performed within a distributed hierarchy in sensory cortex. Our work is amongst the first successful attempts that put a mechanistic spotlight on understanding the underlying neural mechanism of Multisensory Causal Perception in the brain, and in general the theory of decentralized multisensory integration in sensory cortex. Engineering information processing concepts in the brain and developing new computing technologies have been recently growing. Neuromorphic Engineering is a new branch that undertakes this mission. In a dedicated part of this thesis, I have proposed a Neuromorphic algorithm for event-based stereoscopic fusion. This algorithm is anchored in the idea of cooperative computing that dictates the defined epipolar and temporal constraints of the stereoscopic setup, to the neural dynamics. The performance of this algorithm is tested using a pair of silicon retinas.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Themengebiete: 500 Naturwissenschaften und Mathematik ; 500 Naturwissenschaften und Mathematik > 570 Biowissenschaften, Biologie
• Fakultäten: Graduate School of Systemic Neurosciences (GSN)
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 27. März 2020
• 1. Bericht­erstatter:in: Conradt, Jörg