Autism& computational simulations Włodzisław Duch Katedra Informatyki Stosowanej, Uniwersytet Mikołaja Kopernika, Toruń, Poland. Google: W. Duch UoM KL workshop2009
Plan: How to understand autism and many other develop- mental, neurological and psychiatric problems? • Computer Simulations of Brain Functions: general overview of models, what do we want/can learn from them, comparison of suitable neural simulators. • Introduction to the Emergent simulator: general principles, biological plausibility, learning algorithms, models. • Understanding neural activity: visualization, recurrence plots, fuzzy symbolic dynamics. • Two models relevant to autism: attention shifts in visual recognition; learning words. • Conclusion: generative psychiatry.
Neurocognitive informatics Computational Intelligence. An International Journal (1984) + 10 other journals with “Computational Intelligence”, D. Poole, A. Mackworth R. Goebel, Computational Intelligence - A Logical Approach. (OUP 1998), GOFAI book, logic and reasoning. • CI: lower cognitive functions, perception, signal analysis, action control, sensorimotor behavior. • AI: higher cognitive functions, thinking, reasoning, planning etc. • Neurocognitive informatics:brain processes can be a great inspiration for AI algorithms, if we could only understand them …. What are the neurons doing? Perceptrons, basic units in multilayer perceptron networks, use threshold logic – Artificial NN inspirations. What are the networks doing? Specific sensory/motor system transformations, implementing various types of memory, estimating similarity. How do higher cognitive functions map to the brain activity? Still hard but … Neurocognitive informatics = abstractions of this process .
Model of self-organization SOMF, Self-Organized Feature Mapping, or Kohonen map. Simplest model of topographic self-organization via competitive Hebbian activity-dependent learning. Signal X activates most strongly a neuron with synapses W; they become more similar to Xand also neurons in the vicinity of Wbecome more similar to X. Receptive fields of neurons that are close on the 2D map are close in the input space. Update equation:
~ p(MI|X) Myocardial Infarction 0.7 Outputweights Inputweights Inputs: -1 65 1 5 3 1 Smoking Elevation Sex Age Pain Pain ECG: ST Duration Intensity Transformation model Feedforward information processing, no recurrence: categorization,info compression, sensomotoric action.
Dynamical Model Strong feedback interactions, neurodynamics, collective states, recurrent networks. Simplest model that serves as associative memory (Hopfield 1982): two-state neurons (active/inactive), Hebb learning rule, asynchronic dynamics; replaced by graded neuron model. Vector of input activationsV(0)=Vini , input = output = activations. Discrete dynamics (iterations) ÞHopfield network may reach attractor, interpreted as memory state, or autoassociative response to the input query Vini. For symmetric connections (unrealistic) this network reaches a stationary state (point attractor). t = quantized time.
Synapses Soma EPSP, IPSP Spike Spike Biophysical model
Books & references • Reggia J.A, Ruppin E. and Berndt R.S, eds. (1996) Neural Modeling of Brain and Cognitive Disorders. World Scientific. • Parks R.W, Levine D.S. and Long D, Eds. (1998) Fundamentals of Neural Network Modeling. MIT Press, Cambridge, MA. • Reggia J.A, Ruppin E. and Glanzman D.L., Eds. (1999) Disorders of Brain, Behavior, and Cognition: The Neurocomputational Perspective. Elsevier, NY. • O'Reilly R. & Munakata Y. (2000) Computational Explorations in Cognitive Neuroscience: Understanding the Mind by Simulating the Brain. MIT Press. • Callaway E, Halliday R, Naylor H, Yano L, Herzig K. (1994) Drugs and human information processing. Neuropsychopharmacology 10, 9-19. • Ruppin, E. (1996). Neural Modeling of Psychiatric Disorders. Network6: 635-656 (long review of early approaches). • Carnevale N.T, Hines M.L. (2006) The NEURON Book: Cambridge Uni Press. • Beeman D, GENESIS tutorial (2008) http://www.genesis-sim.org/GENESIS/cnslecs/cnslecs.html
Brain-like computing • We see, hear and feel brain states, mostly internal dynamics ! • Cognitive processes operate on highly processed sensory data. • Redness, sweetness, itching, pain ... are all physical states of brain tissue. Ex: visual illusions. Brain states are physical, spatio-temporal states of neural tissue. In contrast to computer registers, brain states are dynamical, and thus contain in themselves many associations, relations. Inner world is real! Mind is based on relations of brain’s states, concepts = pdf(brain activations). Computers and robots do not have an equivalent of such WM.
Symbols in the brain Organization of the word recognition circuits in the left temporal lobehas been elucidated using fMRI experiments (Cohen et al. 2004). How do words that we hear, see or are thinking of, activate the brain? Seeing words: orthography, phonology, articulation, semantics. Lateral inferotemporal multimodal area (LIMA) reacts to auditory visual stimulation, has cross-modal phonemic and lexical links. Adjacent visual word form area (VWFA) in the left occipitotemporal sulcus is unimodal. Likely: homolog of the VWFA in the auditory stream, the auditory word form area, located in the left anterior superior temporal sulcus. Large variability in location of these regions in individual brains. Left hemisphere: precise representations of symbols, including phonological components; right hemisphere? Sees clusters of concepts.
Words in the brain Psycholinguistic experiments show that most likely categorical, phonological representations are used, not the acoustic input. Acoustic signal => phoneme => words => semantic concepts. Phonological processing precedes semantic by 90 ms (from N200 ERPs). F. Pulvermuller (2003) The Neuroscience of Language. On Brain Circuits of Words and Serial Order. Cambridge University Press. Action-perception networks inferred from ERP and fMRI Phonological neighborhood density = the number of words that are similar in sound to a target word. Similar = similar pattern of brain activations. Semantic neighborhood density = the number of words that are similar in meaning to a target word.
Neuroimaging words Predicting Human Brain Activity Associated with the Meanings of Nouns," T. M. Mitchell et al, Science, 320, 1191, May 30, 2008 • Clear differences between fMRI brain activity when people read and think about different nouns. • Reading words and seeing the drawing invokes similar brain activations, presumably reflecting semantics of concepts. • Although individual variance is significant similar activations are found in brains of different people, a classifier may still be trained on pooled data. • Model trained on ~10 fMRI scans + very large corpus (1012) predicts brain activity for over 100 nouns for which fMRI has been done. Overlaps between activation of the brain for different words may serve as expansion coefficients for word-activation basis set. In future: I may know what you’ll think before you will know it yourself! Intentions may be known seconds before they become conscious!
P. McLeod, T. Shallice, D.C. Plaut, Attractor dynamics in word recognition: converging evidence from errors by normal subjects, dyslexic patients and a connectionist model. Cognition 74 (2000) 91-113. M. Spivey, Continuity of mind. Oxford University Press 2007 New area in psycholinguistics: investigation of dynamical cognition, influence of masking on semantic and phonological errors. Energies of trajectories
Problems requiring insights Given 31 dominos and a chessboard with 2 corners removed, can you cover all board with dominos? Analytical solution: try all combinations. Does not work … to many combinations to try. chess board domino n Logical, symbolic approach has little chance to create proper activations in the brain, linking new ideas: otherwise there will be too many associations, making thinking difficult. Insight <= right hemisphere, meta-level representations without phonological (symbolic) components ... still important ,subconscious thinking, insights. o m black white i d o phonological reps
Insights and brains Activity of the brain while solving problems that required insight and that could be solved in schematic, sequential way has been investigated. E.M. Bowden, M. Jung-Beeman, J. Fleck, J. Kounios, „New approaches to demystifying insight”.Trends in Cognitive Science2005. After solving a problem presented in a verbal way subjects indicated themselves whether they had an insight or not. An increased activity of the right hemisphere anterior superior temporal gyrus (RH-aSTG) was observed during initial solving efforts and insights. About 300 ms before insight a burst of gamma activity was observed, interpreted by the authors as „making connections across distantly related information during comprehension ... that allow them to see connections that previously eluded them”.
Insight interpreted What really happens? My interpretation: • LH-STG represents concepts, S=Start, F=final • understanding, solving = transition, step by step, from S to F • if no connection (transition) is found this leads to an impasse; • RH-STG ‘sees’ LH activity on meta-level, clustering concepts into abstract categories (cosets, or constrained sets); • connection between S to F is found in RH, leading to a feeling of vague understanding; • gamma burst increases the activity of LH representations for S, F and intermediate configurations; feeling of imminent solution arises; • stepwise transition between S and F is found; • finding solution is rewarded by emotions during Aha! experience; they are necessary to increase plasticity and create permanent links.
Types of memory Neurocognitive approach needs at least 4 types of memories. Long term (LTM): recognition, semantic, episodic + working memory. Input (text, speech) pre-processed using recognition memory model to correct spelling errors, expand acronyms etc. • For dialogue/text understanding episodic memory models are needed. • Working memory: an active subset of semantic/episodic memory. • All 3 LTM are coupled mutually providing context for recognition. • Semantic memory is a permanent storage of conceptual data. • “Permanent”: data is collected throughout the whole lifetime of the system, old information is overridden/corrected by newer input. • “Conceptual”: contains semantic relations between words and uses them to create concept definitions.
Memory in Alzheimer AD and various diseases leading to dementia are connected with decrease of the density of weak synaptic connections. What happens with associative memory in simple models if weak connections are removed? Is there a biological mechanism that can help to compensate for such memory loss? Horn D, Levy N, Ruppin E (1996) Neuronal-based synaptic compensation: A computational study in Alzheimer's disease. Neural Comput 8, 1227-1243. Hopfield model shows how the remaining synapses may adapt to minimize memory damage. d– degree of impairment k=k(d) compensation function Strong synapses should become even stronger. The degree of memory impairment depends not only on dbut also on the function k(d) related to the history => various symptoms with the same synaptic density are observed.
Compensation The size of basins of attractors as a function of the percent of network damage. Upper circle: learning without compensation; lower – with compensation. Dominant attractors shrink and small attractors have a chance to be active. Performance = % of correct memory associations as a function of synaptic deletion parameter d without compensation (dot-dash) and with two forms of compensation (dashed, continuous lines).
Trace-Link model Meeter M & Murre J.M.J. (2004) Simulating episodic memory deficits in semantic dementia with the TraceLink model. Memory12, 272-287
System 1: Trace system • Function: Substrate for bulk storage of memories, ‘association machine’ • Corresponds roughly to neocortex Slides: courtesy of Jaap Murre
System 2: Link system • Function: Initial ‘scaffold’ for episodes • Corresponds roughly to hippocampus and certain temporal and perhaps frontal areas Slides: courtesy of Jaap Murre
System 3: Modulatory system • Function: Control of plasticity • Involves at least parts of the hippocampus, amygdala, fornix, and certain nuclei in the basal forebrain and in the brain stem Slides: courtesy of Jaap Murre
Retrograde amnesia • Primary cause: loss of links • Ribot gradients • Shrinkage
Anterograde amnesia • Primary cause: loss of modulatory system • Secondary cause: loss of links • Preserved implicit memory Slides: courtesy of Jaap Murre
Semantic dementia • The term was adopted recently to describe a new form of dementia, notably by Julie Snowden et al. (1989, 1994) and by John Hodges et al. (1992, 1994) • Semantic dementia is almost a mirror-image of amnesia • Progressive loss of semantic knowledge • Word-finding problems • Comprehension difficulties • No problems with new learning • Lesions mainly located in the infero-lateral temporal cortex but (early in the disease) with sparing of the hippocampus Slides: courtesy of Jaap Murre
Semantic dementia in TraceLink • Primary cause: loss of trace-trace connections • Stage-3 (and 4) memories cannot be formed: no consolidation • The preservation of new memories will be dependent on constant rehearsal Slides: courtesy of Jaap Murre
No consolidation in semantic dementia Severe loss of trace connections Stage-2 learning proceeds as normal Stage 3 learning strongly impaired Non-rehearsed memories will be lost
Semantic Memory Models EndelTulving „Episodic and Semantic Memory” 1972. Semantic memory refers to the memory of meanings and understandings. It stores concept-based, generic, context-free knowledge. Permanent container for general knowledge (facts, ideas, words etc). Hierarchical Model Collins Quillian, 1969 Semantic network Collins Loftus, 1975
SM & neural distances Activations of groups of neurons presented in activation space define similarity relations in geometrical model (McClleland, McNaughton, O’Reilly, Why there are complementary learning systems, 1994).
Similarity between concepts Left: MDS on vectors from neural network. Right: MDS on data from psychological experiments with perceived similarity between animals. Vector and probabilistic models are approximations to this process. Sij ~ (wi,Cont)|(wj,Cont)
Neurological models Epilepsy: infinite variations. Positive feedback or lack of sufficiently strong inhibition is a general metaphor, but biophysical models are more precise. Detailed models of pyramidal neurons and interneurons in the CA3 area of hippocampus elucidated synchronization processes and showed the influence of various chemicals. Very high 200-600 Hz (phi) frequencies observed in some form of epilepsy cannot be generated by “normal” chemical synapses. Fast electrical nonsynaptic communication is possible through gap junctions filled with connexins, intramembranous proteins, that have rapidly modifiable conductance properties. At least two such synapses/neuron are needed (Traub et al, Nature 1996). How to block synchronization leading to epilepsy? It will require understanding of processes at the molecular level.
Phantom limbs: detailed models are challenging because large (10 mm) reorganization of thalamic projections are observed while usually reorganization is limited to ~ 1 mm. This may indicate that quite new connections are developed. Stroke and cortical reorganization Topographical representations are not fixed, they may reorganize as a result of lesion due to stroke, nerve damage or limb amputation (no activation). Stimulation of adjacent cortical areas may lead to quick expansion of neural representations in S1 cortex to unused areas that will now react to stimulation of quite different parts of skin. Simple SOM models give qualitatively correct results, but transmission through the thalamus is needed to achieve „reverse magnification” effect, size of S1 representation is inversely proportional to the size of receptive field.
Simulation of fast reorganization M. Mazza et al, J. ComputationalNeuroscience 16 (2004) 177-201, detailed model implemented in GENESIS. Hand: 512=32*16 receptors, palm 4x32 + 4 fingers 8x12, Meissner corpuscles mechano-receptors sending bursts of action potentials to ventral posterior lateral (VPL) part of the thalamus.AMPA, GABA & NMDA synaptic receptor dynamics was approx. with simple functions. VPL: thalamic relay cells and interneurons. Ion channels on the soma: Na, Ca low threshold inactivating Ca (Cat), fast Ca voltage dependent K (Kc), slow calcium-dependent K (Kahp), delayed rectiﬁer K (Kd).
VPL Model Grid 16x16=256 relay neurons + 128 interneurons. Input from hand receptorsthrough AMPA synapses, from other relay neurons through AMPA and NMDA receptors.Inhibition through GABA receptors. Connection probabilities between neurons taken from neuroanatomical data. Each cell is connected inside the radiusRcwithother cells with p=0.5, in the ringRc-Re, in elliptical area for the relay neuron with NMDA channels, and interneurons with GABA receptors.
3b cortex area Only layers II, IV i V are modeled, each 32x32=1024. Neuron types: A: burst spiking neurons(BSN): excitatory stellate cells layer IV, B: fast spiking inhibitory GABAergic basket cells (FSN). C: excitatory pyramidal V, D: excitatory pyramidal layer III, regular spiking neurons (RSN); Pyramidal neurons(C, D) have 8 segments, stellate(A) 5, a total of 3072 excitatory neurons in the network. Interneurons (B) have 2 segments, a total of 1536, with 512 in each layer.
3b connectivity A, stellate neurons of layer IV, input from thalamus and other neurons in layer IV. B: basket cells connected to various excitatory cells. C & D: excitatory pyramidal layer III + V cells connected to thalamus and layers II, IV and V. Connections are defined according to experimentally derived probabilities. Equations for currents/potentials are integrated with time step 0.05 ms, on a slow PC 1 sec. of real time required 3.5 hour of CPU time. This model allows for creation of topographical maps in VPL nuclei and in the 3b layers of neocortex.
Simulation results Simulation of randomly selected small 2x2 patches of hand receptors, in 1 s sending 20.000 activations. VPL activity becomes stable already after 500 ms, and cortical activity in layer II and V after 750 ms, and in layer IV after 800 ms. Light areas = weak mixed response, dark >10 spikes/s Results: • Representation area of palm is smaller than fingers. Layer IV has developed most precise map (this agrees with experimental data). • Fluctuations are constantly present on the borders that are “dynamically maintained”.
Amputation of finger 2 After stability is reached (900 ms) and the map formed all stimuli coming from finger 2 were removed. After 400 ms the map has reorganized itself. Some part of neurons stopped reacting to activations but representation of finger 1 and 3 have increased in size, especially in the layer II and V, to a smaller degree in the layer IV and in the thalamus. Simulations demonstrate rapid expansion and reorganization of cortical representations, after that slower consolidation of changes follows. Stability of maps is a result of balance between excitation and inhibition; reorganization results from decrease of inhibition and increase of activity of NMDA receptors. LTP plasticity has not been included in this model.
Comparison of neural simulators. Emergent: a comprehensive simulation environment for creating complex, sophisticated models of the brain and cognitive processes using neural network models. Full 3D GUI environment for constructing networks and the input/output patterns for the networks to process, and many different analysis tools for understanding what the networks are doing. Aisa, B., Mingus, B., and O'Reilly, R. (2008). The emergent neural modeling system. Neural Networks, 21, 1045-1212. Randall C. O'Reilly and Yuko Munakata (2000) Computational Explorations in Cognitive Neuroscience Understanding the Mind by Simulating the Brain, Cambridge, MA: MIT Press. Started as PDP++ Neural Network Simulator (1993). In 2009 Emergent 5.0 was introduced. http://grey.colorado.edu/emergent/index.php/Main_Page Many project were done in Emergent, including SAL (new BICA architecture), some project require older version of software (4.0.19 and PDP ++). Emergent
Many tutorials and projects are described in the book and are ready for explorations, including: Development of visual cortex receptive fields, properties of these fields. Why does primary visual cortex encode oriented bars of light? How is objects recognition invariant across locations, sizes, rotations possible? Series of transformations from V1=>v2=>V4=>IT How is visual attention realized? Why is visual system split into what/where pathways? How does V2=>PPC=>V4 help to shift attention? Why does parietal damage cause attention problems (unilateral neglect)? Memory: how do our brains find the balance between the need to associate and to remember the details? How do we learn AB-AC lists without catastrophic forgetting? How do working, episodic and semantic memory differ and interact? How are higher cognitive functions realized? Reading, dyslexia, learning irregular verbs, learning meaning, categorization … Some things you can do with Emergent
Vmis equilibrium membrane potential, <xiwij> means time average, Q is threshold for activity, b= constfor different types of neurons (slow, fast). Neurons in Emergent Output: spikes or rate coding, number of spikes per second. [ . ]+positive or 0. ge, gi, glare synaptic conductance values. Activation for rate coding is of the gx/(gx+1)=x/(x+1/g) type; gaussian noise is added to smooth this function, contributing to its sigmoidal shape, parametrer g regulates its slope.
Simplified model that includes 3 types of channels: excitatory, inhibitory and leak channels. Spikes
Net = total activation changing from 0 to g_bar_e=1 when all channels are opened. I_net: current flows, equilibrium is reached and sharply drops to zero, after net=0 some current flows back. V_mis the potential on the axon hillock, growing from -70mV (here 0.15) to +50mV (here 0.30). Act= activity along the axon; if spikecoding is used single pulses are sent, some noise Emergent: neuron simulation (here with 0.001 variance) may create small fluctuations. Acteq = rate-code, total cummulative average activity.
Change of conductivity of the leak channels has an influence on selectivity of neurons (size of basins of attractors), for largeĝl only one unit reacts, for smallĝl more units are involved and variability is larger. More associations lower precision of recognition. Leak channels (K+) ĝl =6 ĝl = 5 ĝl = 4