CS 416 Artificial Intelligence Lecture 20 Biologically-Inspired Neural Nets Modeling the Hippocampus
Hippocampus 101 In 1957, Scoville and Milner reported on patient HM Since then, numerous studies have used fMRI and PET scans to demonstrate use of hippocampus during learning and recall Numerous rat studies that monitor individual neurons demonstrate the existence of place cells Generally, hippocampus is associated with intermediate term memory (ITM). HM bilaterally had large regions of his hippocampus and other temporal lobe structures removed for treatment of epilepsy. Consequently suffered from retrograde and anterograde amnesia. Retrograde amnesia was graded with respect to time, with no noticeable loss for events happening more than 10 years ago. (IQ actually went up a couple of points, but within margin of error for test.)
Hippocampus 101 In 1994, Wilson and McNaughton demonstrated that sharp wave bursts (SPW) during sleep are time-compressed sequences learned earlier Levy hypothesizes that the hippocampus teaches learned sequences to the neocortex as part of a biased random processes Levy also hypothesizes that erasure/bias demotion happens when the neocortex signals to the hippocampus that the sequence was acquired, probably during slow-wave sleep (SWS). Busaki achieved similar results in 1996 Levy hypothesis relate to PTSD
Cornu Ammonis The most significant feature in the hippocampus is the Cornu Ammonis (CA) Most work in the Levy Lab focuses specifically on the CA3 region, although recently we’ve started re-examining the CA1 region as well Ammon, “the hidden one”, is the supreme divinity in the Egyptian pantheon
Minimal Model CA3 recurrent activity The hippocampus is getting highly processed, abstract information Other important areas of the brain include the visual, auditory, and motor cortices, Broca’s region and Wernicke’s region, and multimodal cortices such as Brodman’s area 46/9 – the prefrontal cortex. The hippocampus is strongly inter-connected to multimodal regions of the brain. The hippocampus is thought to be a random recoder that is activated more strongly when novelty is encountered whose purpose is to associate things we can’t otherwise associate High dimensionality is key CA3 recurrent activity
Typical Equations Definitions yj net excitation of j xj external input to j zj output state of j θ threshold to fire KI feedforward inhibition KR feedback inhibition K0 resting conductance cij connectivity from i to j wij weight between i and j ε rate constant of synaptic modification α spike decay rate t time Memory is stored in the weights Neurons have binary outputs 0/1
Fundamental Properties Neurons are McCulloch-Pitts-type threshold elements Synapses modify associatively on a local Hebbian-type rule Most connections are excitatory Recurrent excitation is sparse, asymmetric, and randomly connected Inhibitory neurons approximately control net activity In CA3, recurrent excitation contributes more to activity than external excitation Activity is low, but not too low
Model Variables Functional Average activity Activity fluctuations Sequence length memory capacity Average lifetime of local context neurons Speed of learning Ratio of external to recurrent excitations Actual Number of neurons Percent connectivity Time span of synaptic associations Threshold to fire Feedback inhibition weight constant Feedforward inhibition weight constant Resting conductance Rate constant of synaptic modification Input code
Eleven Problems Simple sequence completion Spontaneous rebroadcast One-trial learning Jump-ahead recall Sequence disambiguation (context past) Finding a shortcut Goal finding (context future) Combining appropriate subsequences Transverse patterning Transitive inference Trace conditioning
Sequence Completion Train on sequence ABCDEFG Provide input A Network recalls BCDEFG
Rebroadcast Train network on one or more sequences Provide random input patterns All or part of one of the trained sequences is recalled
One-trial learning Requires high synaptic modification rate Does not use same parameters as other problems Models short-term memory (STM) instead of intermediate-term memory (ITM-hippocampus)
Jump-ahead recall With adjusted inhibition, sequence completion can be short-circuited Train network on ABCDEFG Provide A Network recalls G or possibly BDG, etc. Inhibition in hippocampus does vary
Disambiguation Train network on patterns ABC456GHI and abc456ghi Present pattern A to the network Network recalls BC456GHI Requires patterns 4, 5, and 6 to be coded differently depending on past context
Shortcuts Train network on pattern ABC456GHIJKL456PQR Present pattern A to the network Network recalls BC456PQR Uses common neurons of patterns 4, 5, and 6 to generate a shortcut
Goal Finding Train network on pattern ABC456GHIJKL456PQR Present pattern A and part of pattern K to the network Network recalls BC456GHIJK… Requires use of context future
Combinations Train network on patterns ABC456GHI and abc456ghi Present pattern A and part of pattern i to the network Network recalls BC456ghi Also requires use of context future
Transverse Patterning Similar to rock, paper, scissors Train network on sequences [AB]a+, [AB]b-, [BC]b+, [BC]c-, [AC]c+, [AC]a- Present [AB] and part of + to network and network will generate a Present [BC] and part of + to network and network will generate b Present [AC] and part of + to network and network will generate c
Transitive Inference Transitivity: if A>B and B>C, then A>C Train network on [AB]a+, [AB]b-, [BC]b+, [BC]c-, [CD]c+, [CD]d-, [DE]d+, [DE]e- Present [BD] and part of + to network, and it will generate b
Trace Conditioning Train network on sequence A……B Vary the amount of time between presentation of pattern A and pattern B Computational results match experimental results on trace conditioning in rabbits
Important Recent Discoveries Addition of random “starting pattern” improves performance of network Synaptic failures improve performance (and reduce energy requirements) Addition of CA1 decoder improves performance