Brain anatomy and artificial intelligence. L. Andrew Coward Australian National University, Canberra, ACT 0200, Australia

Brain anatomy and artificial intelligence L. Andrew Coward Australian National University, Canberra, ACT 0200, Australia The Fourth Conference on Artificial General Intelligence August 2011

Architectures and Information Processes Call Processing Processor Registers Diagnostics Maintenance Control Arithmetic Logic Billing Memory Functional Architecture Physical Architecture

Major Brain Structures Lateral Ventricle Precentral Gyrus Central Sulcus Post Central Gyrus Thalamus Cortex Basal ganglia Corpus Callosum Hypothalamus Mammillary Body Fornix Tectum Cerebellum Tegmentum Medulla Pons Spinal Cord

Brain Architecture and Information Processes condition definition and detection behaviour selection Information flow Thalamus Cortex Behaviour selection Dorsal Basal Ganglia Hippocampus Resource Management Reward Ventral Basal Ganglia Amygdala Behaviour type probability Behaviour sequence Cerebellum Behaviour Implementation Brain stem; Spinal cord

Receptive Fields of Cortical Columns 400 µm column All pyramidal neurons in all layers of a column have similar receptive fields

Activity of Cortical Areas During Mental Imaging 9 9/46d 8B 8A 6 4 3,1,2 5 7 33 9 8 32 6 24 4 5 23 31 7 10 46 9/46v 45A 45B44 47/12 38 43 22 21 20 40 39 19 37 18 17 10 25 11 12 34 28 3 35 36 37 8 20 19 19 18 17 18 52 42 41 27 26 29 30 Remembering past event Imagining future event Conceiving Elaborating

a 3 a 2 apical dendrite A B a 4 a 1 C soma synapses from different cortical pyramidals with various weights basal dendrite axon inhibitory synapses from local interneuron

Receptive Fields Must Change as Little and as Rarely as Possible Inputs Cortical layers IV Type of condition detected in layer Conditions that are combinations of inputs from preceding array II/III Conditions that are combinations of conditions detected by previous level V/VI Conditions that are combinations of conditions detected by previous level Outputs

Hippocampus gets Information on Internal Column Activity from All Over the Cortex CA1 CA3 DG Hippocampus Entorhinal cortex Perirhinal and parahippocampal cortices Cortex

Hippocampal Competition Determines Which Cortical Columns will Change Receptive Fields Inputs from entorhinal cortex dentate gyrus CA3 Outputs to cortex CA1 pyramidal neuron granule cell inhibitive interneuron mossy cell

Receptive Field Detection in One Area in Response to Different Objects of Same Category BIRD familiar birds unfamiliar bird inactive column column detecting its receptive field column expanding and therefore detecting its receptive field

Some Columns are Often (but Not Always) Active When a Bird is Seen BIRD columns often active when birds are perceived

Columns active when perceiving a novel event visual objects groups of objects group of groups of objects inactive column column detecting its receptive field column expanding and therefore detecting its receptive field

Columns active when hearing words that relate to novel past event visual objects groups of objects group of groups of objects

Capability to indirectly activate columns on the basis of past simultaneous receptive field expansion visual objects groups of objects group of groups of objects

Column population indirectly activated on the basis of past simultaneous receptive field expansion ~ visual objects ~ groups of objects ~ group of groups of objects = group of groups during original experience

condition definition and detection behaviour selection Information flow Thalamus Cortex Behaviour selection Dorsal Basal Ganglia Hippocampus Resource Management Reward Ventral Basal Ganglia Amygdala Behaviour type probability Behaviour sequence Cerebellum Behaviour Implementation Brain stem; Spinal cord

Behaviour Selection, Including At Least and Only One Behaviour Cortex Striatum D1 D2 Dorsal Basal ganglia Excitatory (glutamatergic) Inhibitory (GABAergic) GPe Modulatory (dopaminergic) STN SNc GPi/SNr Thalamus Modulation pathway Direct pathway Indirect pathway

Implementation of Strategic, Tactical and Detailed Reward Behaviours Striatum VTA } Shell Shell } Midbrain dopamine neurons } Core Striatum Core } } Central striatum Central striatum } } Dorsolateral striatum Dorsolateral striatum } SNc orbital and medial prefrontal dorsolateral prefrontal premotor motor Cortex

supplementary motor cortex hippocampus front Thalamic reticular nucleus right entorhinal cortex VA AN motor cortex; supplementary motor cortex VL LD MD Massa intermedia touch, propioceptic sensory information VP LP somatosensory cortex P visual and other association areas LGN MGN medial prefrontal cortex prefrontal cortex primary visual cortex visual information primary auditory cortex auditory information

Implementation of Release Behaviours by Imposing Frequency Modulation cortical layers thalamic reticular nucleus thalamic nucleus IV V VI cortical pyramidal neuron thalamocortical projection neuron TRN inhibitory interneuron excitatory connection inhibitory connection basal ganglia inhibitory connection

Parallel Cerebellar Path for Implementing Frequently Used Sequences of Actions

To design an artificial general intelligence system Focus on the information processes that are required If a system needs to Perform many different behaviours Detect many different conditions Define most of the conditions heuristically Limit the resources required Then the information processes will need to be Condition definition and detection Condition resource Information flow Behaviour selection Reward behaviour selection Behaviour priority Behaviour sequence condition definition and detection Resource Management Behaviour type probability Behaviour Implementation Information flow Behaviour selection Reward Behaviour sequence behaviour selection

a. Boundary of object in visual field b. Some of boundary element receptive fields detected in V1 c. Retinal area for which one boundary element receptive field detection recommends modulation of inputs from area d. Total recommendation strengths for modulation much stronger for area within object boundary

postsynaptic potential threshold total 5 10 15 20 25 30 35 input action potential spikes msec

source neurons target neuron integration window close to modulation minimum integration window close to modulation maximum time window of one third of the modulation cycle

I II III

entorhinal cortex dentate gyrus amygdala supramammillary area II subicular complex mammillary bodies III V/VI CA2 CA3 anterior thalamus CA1 pyramidal neuron granule cell cortex inhibitive interneuron mossy cell

Purkinje cell dendritic tree perpendicular to parallel fibres Connectivity Purkinje cell bodies parallel fibre outputs from granule cells granule cells Purkinje cell output targets cerebellum nuclei Pontine Nucleus mossy fibre input targets granule cells climbing fibre input targets small group of Purkinje cells (~10) Inferior Olive

8B 9 6 9/46d 8A 4 3,1,2 5 7 33 9 8 32 6 24 4 5 23 31 7 10 46 9/46v 45A 45B 44 47/12 38 43 22 21 20 40 37 39 19 18 17 10 25 11 12 34 28 3 35 36 37 8 20 19 19 18 17 18 52 42 41 27 26 29 30 Areas concerned with working memory Areas with receptive fields corresponding with groups of cortical columns often active at similar times in the past Areas with receptive fields corresponding with groups of cortical columns that expanded their receptive fields at similar times in the past

striatum left cerebral basal nuclei cerebellum hemisphere lateral ventricles some of basal ganglia thalamus i ii iii vi v iv caudate putamen substantia nucleus nigra nucleus accumbens fornix hippocampus septal nuclei hypothalamus claustrum amygdala

the striking commonalities in medial left prefrontal and parietal activity during the elaboration of (a) past and (b) future events (relative to the control tasks) From Addis, D.A., Wong, A.T. and Schacter, D.L. (2007). Remembering the past and imagining the future: Common and distinct neural substrates during event construction and elaboration. Neuropsychologia 45, 1363-1377.