Abstract:

  This investigation, of multidisciplinary interest, delves into how the structure and connectivity of the brain influences neural processing. This thesis addresses issues of modelling brain function and consequent changes due to neural disorders. It is carried out through computer simulations within a systemic framework. The work offers plausible explanations for selected neural processing by bringing together topics from neuroscience, medical imaging, artificial intelligence, and computing. In doing so, it also helps to reduce the gap between neurobiology and cognitive neuroscience. All of the above topics are presented in a coordinated manner, gradually transitioning from biological to artificial “worlds”. The employed formalism and holistic approach are aimed at systemic analyses of granularities compatible with data acquisition techniques.

 The Venn-network, a novel neural network model proposed by the author, is introduced and exhaustively simulated. This new neural network architecture allows definition and use of multiple types of processing unit, multiple regions within the network structure, and several types of axonal fibres. Most interestingly, the internal activity of utilised networks resembles images produced by functional brain imaging.

A comprehensive computer simulator was developed for implementing the Venn-network model. This simulator was used to carry out experiments such as structural-functional equivalence, active-passive activations, modulation, ageing, and contra-lateral inhibition. Next, disrupting effects such as the ones produced by (1) multiple sclerosis plaques and (2) strokes, were applied in these simulations. Throughout these simulations, Venn-networks were trained to control flexions of (ten) virtual fingers to reproduce movements of a piano player performing a Mozart Sonata.

The Venn-network model proved to be an effective tool for predicting behaviour in both physiological and pathological scenarios. The correctness and robustness of all implemented Venn-networks was verified in the simulations carried out, as the functionality of all neural architectures conformed with the expected behaviour.

Potential applications of the research include: (i) to support prognoses in neurology (e.g. multiple sclerosis effects due to plaques growth and inference of damage due to strokes), (ii) as a test-bed for producing insights into neuromorphic systems, (iii) to interface between cognitive algorithms and front-end electronics of robots, (iv) as a brain interface for controlling prosthetics, and (v) to provide underlying models for checking hypotheses in medical imaging experiments.

 

 Keywords: Artificial Neural Networks, Venn-Networks, Models of the Brain, Physiological  Brain Scenarios, Pathological Brain Scenarios, Imaging Methods of the Brain.

 People:

1.  Fernando Buarque de Lima Neto - PhD Candidate

2.  Philippe De Wilde - PhD Supervisor

 

 Funding: PhD overseas grant from the Brazilian Government, number 200295/98-5 (CNPq - Brazilian Agency for Scientific Research)

 

 Period: October/1998 to December/2002

 

 Place: Department of Electrical and Electronic Engineering - Imperial College London - University of London - London, England. 

 

 Hypotheses:

1.  Structure and connectivity of the brain influence neural processing;

2.  Selected neural disorders can be explained and forecast in terms of changes to the normal physiology of the system; and

3.  Underlying principles used in some artificial models can produce emergent activations that resemble imagery of functional modalities of the brain

 Context: Understanding how a functioning brain produces a mind is one of the ultimate challenges posed to humankind. Assuming that this bold task is possible, it certainly will require a number of preceding steps: (i) understanding life formation itself and (ii) understanding the emergence of intelligence. Even though science has been producing breakthroughs at incredible rates, one cannot be sure if the mind will ever be fully understood. At present physicists, biochemists and biologists are committed to this quest in their respective research fields. They are all aiming at understanding how, from sub-atomic particles, we may ended up having complex molecules and ultimately, life. Concomitant to this, some psychologists, neuroscientists and scientists studying artificial intelligence are dedicated to comprehend how life became intelligent, and how intelligence can self-sustain itself. All these scientific resolve leave researches on plausible intelligent computation a yet more appealing and challenging path. But intelligence in itself is not an immaterial or intangible entity. It is grounded in the underlying mechanisms of the nervous systems of living creatures, no matter how simple or complex they are. This emergent phenomenon – magnificent in itself – seems yet to be accountable for the self-sustainability and non-monotonicity of life, and eventually, for (our) free-will. Thus, any advance in understanding of how life can exhibit intelligent behaviour requires insights in the “mechanics” of the nervous system. This thesis is intended to be one more brick on this tortuous and long way ahead, towards the goal of explaining how biology can sustain intelligence (and then, vice-versa). The central idea investigated is how biologically inspired computation and a systemic framework for neural information processing can be an alternative to the traditional cumbersome mathematical or statistical approaches for tackling the problem. The author suggests that more flexible, adaptable and scalable models can be created if structural and dynamic features of the biological system are taken into consideration.

 

 Expected results: The results of this exploratory investigation are expected to examine whether the theoretic and methodological approaches contributed are appropriate for helping the research problem (i.e. check validity of hypotheses). This means that the proposed model and approach should advance the understanding about the impact of brain structure onto selected brain function (and behaviour). Therefore, the simulation results of the artificial models utilised in this thesis are the putative metaphor for explanation of the observable behaviour and activations of the cortex.

The proposed models should also be accountable for symptoms elicited in case of the selected diseases. Hence, the correctness and robustness of the models in simulations of ‘physiological’ as well as in ‘pathological’ scenarios should contribute to reduction in the gap between micro and macro of neural computation.

 

 Conclusions: the research results have shown that:

1. Venn-neural networks can be ‘trained’ to evoke expected behaviour of complex tasks. The learning tasks of all ten experiments described in the chapters 6 and 7 have converged without requiring any ad hoc programmed routines or extremely complex architectures.

2. At the same time all simulated tasks were satisfactorily learnt and executed by Venn-networks, the internal activity of these architectures evoked observable activations that resembled functional images (i.e. spatially and temporally localisable kernels of neural activity correlated to the task performed).

3. Venn-networks are robust for training and execution of some motor tasks with respect to a range of disrupting factors of various kinds namely, neural disorders (e.g. multiple sclerosis and strokes), background noise, and ageing processes.

4. The neural structure in Venn-networks does influence neural processing simulations, even though some structural differences may not be of critical importance or cannot be easily observable in healthy and some unhealthy conditions. This was proved when simulations of physiological scenarios of networks trained to perform some tasks revealed different observable neural activation in induced pathological situations on that same task, e.g., vide simulation-set 5 (in chapter 6) and simulation-set 10 (in chapter 7).

5. Although more evident and of penalising consequences in lesions studies, functional localisation was demonstrated by Venn-networks to be economic concerning consumption of computational resources: ‘space’ and ‘time’. E.g. refer to simulation-set 1 (in chapter 6).

 

Original contribution: This thesis contains material produced by independent scientific research pursued by the author during his post-graduate studies at Imperial College of Science, Technology and Medicine, London, United Kingdom. To the best of his knowledge the ideas and results included in this work are original, and the major contributions offered by this work are:

1. Proposition of Venn-networks – a new artificial neural network that allows definition and use of multiple types of processing units (i.e. cortical columns), multiple regions within the network structure, and four types of axonal fibres.

2. Proposition of a computational model for simulation of the effects caused by multiple sclerosis plaques to single nervous pathways . The model can also be used to investigate consequences (inference) of MS-plaques growth to neural processing.

3. Development of a computer simulator that implements Venn-networks which incorporate the model of multiple sclerosis plaque. The simulator allows concomitant simulations of ageing, modulation, strokes, and other aspects related to physiological and pathological scenarios of neural processing.

4. The models and simulator proposed were used to provide satisfactory explanation and forecasting of selected neural processing in physiological and pathological scenarios . Furthermore, the internal activity of Venn-networks resembles images produced by current functional imaging methods.

 

Other Scientific Production: the research results also were used to produce:

On-line material: (a) Accompanying CD-ROM [highly-recommended]     (b) A4-Poster

Journal publication:

                           1)  Simulation of Contra-Lateral Inhibition Using Venn-Network.

                                          (Journal of Intelligent and Fuzzy Systems, Special Issue SBRN, 2006.)

                                    2) Multiple Sclerosis plaques on nervous pathways: A computational model using Neural Networks

                                          (Journal of Differential Equations and Dynamical Systems, Hyderabad, v. 10, n. 1 & 2, p. 215-234, 2002.)

 

Conference publication: (FULL PAPER)

                                     1) Venn-Like Models of Neocortex Patches

                                         (WORLD CONGRESS IN COMPUTATIONAL INTELLIGENCE /  IJCNN - INTERNATIONAL JOINTCONFERENCE 

                                           ON NEURAL NETWORKS, Vancouver - Canada, 2006) 

                            2) Simulation of Ageing Using Venn-Networks  

                                         (BRAZILIAN SYMPOSIUM OF ARTIFICIAL NEURAL NETWORKS, São Luis-Brazil. 2004)

                                     3) Flexion of Virtual Fingers Controlled by Artificial Neural Networks

                                        (WSES INTERNATIONAL MULTICONFERENCE - SIM, 2001, Valletta. Mathematics and Simulation with Biological, 

                                         Economical and Musicoacustical Applications. WSES Press, 2001. p. 34-39)

 

Conference publication: (ABSTRACT)

                            1) Multi-function control using Venn Networks

                                          (3RD. CONFERENCE ON SENSORIMOTOR CONTROL IN MEN AND MACHINES, Marseille. 2001.)

                                     2) Modelling the effects of MS-Plaques growth.

                                         (2000 AUTUMN SCHOOL IN COGNITIVE NEUROSCIENCE, 2000, Oxford. 2000.)

                                     3) Modelling axonal delays caused by Multiple Sclerosis Plaques

                                         (WORKSHOP ON MEMORY, DELAYS AND MULTISTABILITY, 2000, Montreal. - Proc. of 2000 Workshop on

                                           Memory, Delays and Multistability. 2000.)

 

Software:              Generalized Venn-Network Simulator. 2002.