Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
LEARNINGCOLIN MCDONNELL
LEARNING
THE PAPERS
▸ Innate attentional biases and associated effects
▸ Thalamocortical loops for prediction learning
▸ Method of comparing human classification capability to neural networks
how the brain learns
and vice versaand how it informs machine learning
with no pretensions of completeness
LEARNING
▸ What is learning
▸ Neuroscience inspires machine learning
▸ Machine learning guides neuroscience
STRUCTURE OF TALK
SOME DEFINITIONSwhat do I mean neuroscience? what do I mean ML?
LEARNING
Affective neuroscience Behavioral neuroscience
Cellular neuroscience Clinical neuroscience
Cognitive neuroscience Computational neuroscience
Cultural neuroscience Developmental neuroscience
Evolutionary neuroscience Molecular neuroscience
Neuroengineering Neuroethology Neuroheuristics Neuroimaging
Neuroinformatics Neurolinguistics
Neurophysics Neurophysiology Neuropsychology Paleoneurology
Social neuroscience Systems neuroscience
NEURAL NETWORKS
WHAT IS LEARNING
▸ pattern recognition
▸ finding an efficient encoding
▸ discovering beneficial behavior based on past observations
▸ abstraction
▸ Australian or European?
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
NEUROSCIENCE INSPIRES MACHINE LEARNING
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
Hebbian learning
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
spike timing dependent plasticity
404
synaptic re-wiring
sort of a 404, but…
sparsity
winner-take-all learning
ReLU?
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
many excitatory synapses (for temporal learning?)
neurons that remember window of past inputs/firings
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
dendritic computation layers collapse to neurons mathematically equivalent
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
reflection of action potential into input
dendrites
gradient learning
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
molecular diffusion
neuromodulators neuropeptides nitrous oxide
message-passing networks
going smaller…WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
retrograde chemical signaling reverse connections
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
attractor circuitsshould arise organically…
(later: LeabraTI)
going bigger…WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
communicating brain regions
top-down feedback [Ullman 2015]
basal ganglia for decisions
multi-sensory learning
404 eventually, interacting
networks:
deep reinforcement learning
variable binding
information routing
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
innate biasesattention mechanisms
“feature weights”
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
‣ bias network towards “mover events” ‣ first pass: ‣ high precision, low recall
‣ second pass: ‣ tracking/continuity ‣ body context
‣ third pass: ‣ gaze
‣ fourth pass ‣ mirroring ‣ own-hands
‣ comparison ‣ but was it necessary?
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
HANDS
POSEGAZE
implicit labeling/ proto-concepts [Ullman 2012]
semi-supervised labeling
basins of attraction as reward signal generator
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
MACHINE LEARNING GUIDES NEUROSCIENCE
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningunsupervised learning
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learning
error-driven backpropagation
404?
problems ‣ requires transmission of exact
derivatives over distance ‣ neurons communicate through
(stochastic?) spikes ‣ temporal synchronicity ‣ what are the targets?
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learning
derivative backpropagation
hyper polarization of dendrites
+ reverse action potential
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningsupervised learning
training epochs hippocampus-mediated replay
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningsupervised learning
momentum gradient descent plasticity modulation
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningsupervised learning
dropout stochastic models of neural spiking
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningsupervised learning
layer-by-layer training developmental drop-off in cortical plasticity
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
supervised learningsupervised learning
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
unsupervised learning
auto-encoders
404
layer-size bottleneck
lateral inhibition
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
unsupervised learning
predictive learning
babies understand spatiotemporal coherence
thalamocortical loops! [O’Reilly 2014]
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
unsupervised learning
prediction learning
babies understand spatiotemporal coherence
thalamocortical loops! [O’Reilly 2014]
WHAT IS LEARNING > NEURO INSPIRES ML > ML GUIDES NEURO
unsupervised learning
LEARNING
INTER-DISCIPLINE COMPARISONS
QUESTIONS