CISC 3250 Systems Neuroscience

Transcription

1 CISC 3250 Systems Neuroscience Perception (Vision) Professor Daniel Leeds JMH 332 Pathways to perception 3 (or fewer) synaptic steps 0 Input through sensory organ/tissue 1 Synapse onto neurons spal cord/bra stem 2 Synapse onto neurons thalamus 3 Synapse onto cortical neurons primary cortex 4 + Further cortical processg Bundled track of nerves to bra: spal cord/cranial nerve 2 Touch/ Tactile Touch: Inputs Sk Spal cord Dorsal horn Thalamus Mechanoreceptors sk Pacian corpuscles vibrations Meissner s corpuscles light touch Merkel s discs pressure and texture Ruffi endgs stretch Ventral posterolateral (VPL) nucleus Cortex Post-central gyrus (SI) 3 4 1

2 Communications the spal cord Sensory activity back dorsal Motor command front ventral Thalamus the relay station Region names largely based on location VPL for somatosensation Hearg/ Auditory Regions of the brastem Cochlea Cochlear nerve Cochlear nucleus (-> Superior olive) -> Inferior colliculus Bra stem Thalamus Medial geniculate nucleus (MGN) Cortex Primary auditory cortex (AI) 7 Dorsal view (back-of-thehead) 2-3 synapses auditory brastem path 8 2

3 Seeg/ Visual Lateralization Flippg of right and left vision Reta Optic nerve Thalamus Lateral geniculate nucleus (LGN) Cortex Left hemisphere right visual field Right hemisphere left visual field Primary visual cortex (VI) 9 10 Sensitivity to perceptual variations Selectivity to perceptual variations V1: Surround-suppression for shifted edges PFC: Same object detected at diverse locations and scales More complex percepts variant to greater spatial transformations

4 HMAX model of hierarchical vision HMAX model of hierarchical vision V4 V2 Higher cortical levels cover larger visual spans Object recognition variant to changes location and orientation IT V Gabor filters (edge detectors) 2. Perform Max poolg (semi-variance over space) 3. Weighted combation of space-variant edges 4. Further max poolg 14 Higher HMAX layers cover more space Functions of HMAX layers Odd layers (layer 1, 3, 5, ) look for specific combations of lower-level features Even layers (layer 2, 4, 6, ) provide variance to some feature changes (e.g., shift position) Example coverage for layer x neurons layer 1 layer 2 layer 3 15 layer 1 layer 2 Fire for 1 + les layer 3 layer 4 Fire for 1 + Is 16 4

5 g rad (h) 3/22/2017 Functions of HMAX layers Odd layers (layer 1, 3, 5, ) look for specific combations of lower-level features h= j w j r j r out = g rad (h) Radial basis function Detectg triangles: layer 2 Neuron outputs 1 if desired image viewed, otherwise 0 Layer 1: Specific edge at specific location Layer 2: Specific edge at slightly varied locations Even layers (layer 2, 4, 6, ) provide variance to some feature changes (e.g., shift position) r out = max( r 1 r 2 r J ) h Layer 1 max Layer 2 Horizontal edge at approximate location max Layer 2 Diagonal edge at approximate location Detectg triangles: layer 3 Neuron outputs 1 if desired image viewed, otherwise 0 Layer 2: Specific edge at slightly varied locations Layer 3: Combation of edges Detectg triangles: layer 4 Neuron outputs 1 if desired image viewed, otherwise 0 Layer 3: Combation of edges Layer 4: Triangle on the left Layer 2 Weighted sum i w i r i Layer 3 Triangle centered at fixed location Accepted stimuli layer 3 neuron Layer 3 max Layer 4 Left triangle

6 Visual attention Attention when percepts overlap Emphasize details currently of terest CAT Cocktail party problem Attention a Ignore vertical edges: a =0 Pay attention to all other edges: a =a / =a \ =1 Attention when percepts overlap Weights w H-detector looks for and w =w =1 w / =w \ =0 A-detectors looks for /, \, w =w / =w \ =1 w =0 Rate r If feature present: 1 If feature not present: 0 In this example,,, /, \ present Attention when percepts overlap w / =0 for H a =1 H detector A detector a =0 a =0 a =1 a / =1 a \ =1 i w ia i r i = 1 i w i a i r i = 3 w =0 for A 26 Can attend to one of two voices (e.g, high-pitched voice or low-pitched voice) 27 6

7 Modulatg puts through multiplication Algorithm: Sigma-Pi Node Multiply rates to modulate each put Sum to compute output rate Dynamic synaptic reweightg Voltage-dependent NT-receptors (e.g., NMDA): 1. Other nearby receptor decreases voltage 2. Voltage dependent receptor detects NTs h i = i w i r i att r i r i att - attention put r i att = j r ij att - can sum over multiple attention puts Complexity of cortical networks Feedback: connections both directions along cortical pathways Creative Commons, some rights reserved