computation in sensory processing · • adrian (1920s) • first recordings ... •learning of...
TRANSCRIPT
Spikes and Computation in
Sensory Processing Simon Thorpe
CerCo (Brain and Cognition Research Center) &
SpikeNet Technology SARL,
Toulouse, France
• 1974-77Physiology and Psychology (Oxford)• 1977-81 Doctorate with Edmund Rolls (Oxford)
• Recording neurons in orbitofrontal cortex, striatum, parietal cortex, lateral hypothalamus, amygdala, hippocampus, supplementary motor area (!!)
• 1982-83Post-doc with Max Cynader (Halifax, Canada)• Attempting to test Hebb's hypothesis in kitten visual cortex
• 1983-93Paris with Michel Imbert• Testing Hebb's Hypothesis (with Yves Frégnac & Elie Bienenstock)• Dynamics of responses in monkey V1 (with Simona Celebrini & Yves Trotter)
• 1993-nowCERCO Toulouse• Work on ultra-rapid scene processing
• 1999-now SpikeNet • Spike-based image processing• Bioinspired vision
Simon Thorpe
Overview
Part 1
• Ultrarapid visual categorization
• Biological vs Computer Vision
• Temporal Constraints
• Coding with Spikes
• Convolutional Neural Networks in 1999
• SpikeNet
• The current state of the art in Convolutional Neural Networks
• Supervision & GoogleNet
Part 2
• Biologically inspired learning
• Spike-Time Dependent Plasticity (STDP)
• Applications in Vision
• Applications in Audition
• Development of Neural Selectivity
• Memories that can last a lifetime
• Grandmother Cells
• Neocortical Dark Matter
Behavioural Reaction Times
Targets
Distractors
Difference
Event Related Potentials
Scene Processing in 150 ms
Ultra-Rapid Visual Processing
•Saccades towards faces
•Latency 100ms!
Crouzet, Kirchner & Thorpe, 2010
Ultra-Rapid Visual Processing
RSVP at 10 fps
Ultra-Rapid Visual Processing
Activating Memories
• “That Dalmation picture that I saw in Psych 101”
• How does the brain recognize visual objects and scenes?
• Can a machine be built that can do the same thing?
Some Questions
Biological and Computer VisionCommon Problems
• Identification and Categorising objects and events in complex dynamically changing natural scenes• As fast as possible• As reliably as possible• Using the most energy efficient hardware possible• Using the smallest size and weight footprint
Common Solutions?• David Marr (1982) “Vision : A computational
investigation into the Human Representation and Processing of Visual Information”
• Recent years - is there convergence?
Hardware ConstraintsElectronics • Nvidia GeForce GTX Titan
• 4.5 TeraFlops!
• 2668 cores
• 7.1 billion transistors
• 288 Gbytes/sec Memory bus
• $999
Brain• 1 KHz
• 86 billion processors
• 1-2 m/s conduction velocity
• 20 watts
QuestionsAre we going to be able to implement brain style computing with conventional computing?
ResponseIt depends if we can understand how the brain computes
How many teraflops does the brain need?
How much memory bandwidth?
Classic Neural Computing
• To simulate the visual system
• 4 billion neurons
• 10000 connections each
• Update at 1 kHz
• 40 Petaflops
What’s missing?
• Real brains use spikes• Simulation
• 16.7 million neurons
• 21 billion synapses
• The Brain• 86 billion neurons
• 16 billion in the cortex
• 100 trillion synapses
• Spikes• Firing rate
• 0-200 Hz
• Spontaneous activity
• 1-2 Hz?
Classes of Brain Simulation
• Connectionist Simulators (non-spiking)
• Backpropagation, ART, Kohonen Maps, Time Delay Networks,
• Computational Neuroscience Simulators
• Up to 30 000 compartments
• Lots of channel kinetics etc.
• Spike-Based Simulators
• Software
• Hardware
The Human Brain Project
• European Flagship Proposal
• Three simulation approaches
• Blue Brain
• Henry Markram (Lausanne, Switzerland)
• Analog Chips
• Karl-Heinz Meyer (Heidelburg, Germany)
• SpiNNaker
• Steve Furber (Manchester, UK)
SpiNNaker Project• A simulation system for Spiking Neural Networks
• Prof. Steve Furber, Computer Science, University of Manchester
• 18 ARM968 cores • each with 64 Kbytes of Data and 32 Kbytes of Instructions
• 128 MBytes of shared memory• 48 chips on a board• 18 boards in a 19” frame
A billion neurons in real time
IBM TrueNorth Chip
• What can be computed?
SCIENCE
• Jerry Feldman’s 100-step limit
• High level decisions in about 0.5 seconds
• Interval between spikes around 5 ms
• No more than 100 (massively parallel) computational steps
• Development of connectionist and PDP modelling
• Surely, more detail can help
Temporal Constraints – Early 1980s
Temporal Contraints - 1989
Argument• Roughly 10 layers• 10 ms per layer• Firing rates 0-100 Hz
Face selectivity at 100 ms
Food selectivity at 150 ms
Therefore• Mainly feedforward (?!) • One spike per neuron (?!)
T
T
T
TT
TRetina
LGN
V1
V2
V4
IT
Temporal Constraints 1988-1989Inferotemporal Cortex :Face selectivity at 100 ms
Perrett, Caan & Rolls, 1982
See also : Jerry Feldman & John TsotsosLeonard Uhr (1987)
How can you code with just one spike per neuron?
• Feedforward processing
• Only a few milliseconds per processing step
• One spike per neuron
• Very sparse coding
• Processing without context based help
Ultra-Rapid Visual Processing
Sensory Coding with Spikes
• Adrian (1920s)• First recordings from sensory fibres
• Hubel & Wiesel (1962)• Orientation selectivity in striate cortec
Sensory Coding with Spikes
• Wurtz & Goldberg (1972, 1976)
Raster Display Post Stimulus Time Histogram
Assumption : Firing rate is enough to describe the response
The Classic View
2
4
1
2
3
0,213
0,432
0,112
0,238
0,309
•Spikes don't really matter
•Neurons send floating point numbers
•The floating point numbers are transformed into spikes trains using a Poisson process
•God plays dice with spike generation0,375
Temporal Coding Option
• Spikes do really matter
• The temporal patterning of spikes across neurons is critical for computation
• Synchrony• Repeating patterns• etc
• The apparent noise in spiking is unexplained variation
Simon Thorpe's Version
Weak StimulusMedium StimulusStrong
Stimulus
Time
Threshold
•Ordering of spikes is critical•The most activated neurons fire first•Temporal coding is used even for stimuli that are not temporally structured•Computation theoretically possible even when each neuron emits one spike
Sensory Coding with Spikes• Adrian (1920s)
• First recordings from sensory fibres
Low intensity
High intensity
Higher maintained firing
Higher peak firing
Shorter Latency!
Coding in the Optic Nerve
1,000,000 fibres
Intensity
n
n
n
n
n
n
n
n
nn
n
n
n
n
n
n
Coding in the Optic Nerve
Intensity
n
n
n
n
n
n
n
n
nn
n
n
n
n
n
n
A mini retina32 x 32 pixels
Coding in the Optic Nerve
Coding with Spike Ordering
Less than 1% of cells need to fire for recognition!
Example• A toy retina
Early Studies
• Face identification directly from the output of oriented filters
Early Studies• Virtually all the faces correctly
identified
• Very robust to low contrast
• Very robust to noise
SpikeNet Technology
• Created in 1999
• Simon Thorpe, Rufin VanRullen & Arnaud Delorme
• Currently 12 employees
SpikeNet - Invariance
• Luminance
• Contrast
• Blurring
• Noise
SpikeNet - Invariance
• Size
• Rotation • Perspective
SpikeNet - Invariance
• 3D Rotation
• Identity
•What is the current state of the art?
Biological vs Computer HardwareComputer
• Nvidia GeForce GTX Titan
• 4.5 TeraFlops!
• 2668 cores
• 7.1 billion transistors
• 288 Gbytes/sec Memory bus
• 200 watts
• $999
• 86 billion neurons
• 16 billion in the cortex
• 4 billion in the visual system
• 1 KHz
• 1-2 m/s conduction velocity
• 20 wattsIs that enough to reproduce human performance?
Brain
The ImageNet Challenge
• 10,000,000 training images
• 10,000+ labels
• Systems tested on new images, with 1000 possible labels
• ECCV 2012 Firenze
• The state of the art was beaten by a “simple” feedforward convolutional neural network trained with Back-Propagation
SuperVision
• 650,000 “neurons”
• 60,000,000 parameters
• 630,000,000 “synapses”
Output Layer
Convolutional Layers
Fully Connected Layers
253440 186624 64896 64896 43264
Animals
SuperVision
And then….• Geoff Hinton and his two
students launched a start-up (DNNresearch)
• bought by Google…
• Yann LeCun, a pioneer of feed-forward convolutional networks since the end of the 1980s
• Hired by Facebook…
Supervision vs Primate Vision
Convergent Evolution!
? ? ? ?
Coding at higher levels
Basic Methodology
•Computer models• 57 parameters• 1-3 layers
•Neural Activity• Parallel Multielectrode Recordings • V4 and IT• Activity (70-170 ms)
•Human performance• Crowd sourced data• Amazon Mechanical Turk• 104 subjects
Comparison of Models
• HMO - Hierarchical Modular Optimisation
• 4 layer CNN with 1250 output units
• Close match between models, humans and neural signals
Human vs Machine
• Sorry, Andrej!
• What about faces?
Feb 2015
Human vs Machine
A range of architectures• SpikeNet (1999)
• Identification immediately after V1
• Supervision (2012)• 7 layers
• GoogleNet (2014)• 30 layers
• 12 times less parameters
The Surprises• Human levels of performance are possible despite
• No feedback• No horizontal connections• No top-down control• No need for context• No dendrites• Complex channel dynamics• Only positive and negative weights• No dynamics• No memory• No oscillations• No “binding”• No attention!!!• No Spikes!!!
Question : Is Spike Timing Precise Enough?
Fred Rieke, David Warland, Rob de Ruyter van Steveninck and William Bialek (1999)
Increasing evidence that the timing of spikes carries a great deal of information
First Spike Coding : Evidence
Coding motion direction
Coding stimulus shape
Experimental proof in the visual system!
Salamander Retinal Ganglion cellsInformation in latency vs Information in the spike count
Overview
Part 1
• Ultrarapid visual categorization
• Biological vs Computer Vision
• Temporal Constraints
• Coding with Spikes
• Convolutional Neural Networks in 1999
• SpikeNet
• The current state of the art in Convolutional Neural Networks
• Supervision & GoogleNet
Part 2
• Biologically inspired learning
• Spike-Time Dependent Plasticity (STDP)
• Applications in Vision
• Applications in Audition
• Development of Neural Selectivity
• Memories that can last a lifetime
• Grandmother Cells
• Neocortical Dark Matter
The Learning Problem
•SuperVision uses Backpropagation
•Billions of training examples with labeled data
• 1 ms per image!• 100 images in 100 ms
•Totally unbiological
Spike-Based Processing
Song, Miller & Abbott, 2000
• Synapses that fire before the target neuron get strengthened
• Synapses that fire after the target neuron get weakened
Spike Time Dependent Plasticity (STDP)
Simplified STDP rule• All synapses are depressed after each output neuron spike
• Except those that fire in just before
• Natural Consequence• High synaptic weights concentrate on early firing inputs
Threshold 3
Short time constant
12 inputs
Initial synaptic weight = 0.25
Threshold = 3.0
All inputs must fire to reach threshold
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 1
All synapses activated before the output
neuron
STDP reinforces all synapses
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 2
9 synapses reinforced
3 depressed
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 3
6 synapses reinforced
6 depressed
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 4
3 synapses reinforced
9 depressed
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 5
Learning of a repeating pattern
Threshold 3
Short time constant
Presentation 6
3 fully potentiated synapses
9 set to zero
Learning of a repeating pattern
Threshold 3
Short time constant
In six presentations STDP has found the first three inputs that
fire duing the repeating pattern
Multiple Units
Mutual inhibition
Competitive Learning Mechanism
Multiple Units
Multiple Units
Multiple Units
Multiple Units
Neurons can detect patterns embedded in continous activity
Neurons can detect patterns embedded in continous activity
Multiple Units
Add a second layer to detect combinations of combinations
Can this be scaled up beyond a toy demo?
Learning Spike Sequences with STDP
• Initial State
• DuringLearning
• After Learning
Learning Spike Sequences with STDP
Learning Faces with STDP
Learning Cars with STDP
Tobi Delbruck’s Spiking Retina
STDP Rule
12 seconds 30 seconds 90 seconds
• The system has learned to recognize cars
• No supervision!
“STDP” Rule
Tobi Delbruck’s Spiking Retina
Learning Cars with STDP
Building a complete visual system
1920
1080
• Add more neurons
• Add more layers of processing
• Increase the input resolution
• Use more sophisticated preprocessing
• Mexican hat convolutions
• Different types of input signal
Auditory Noise Learning in Humans
•Learning of random noise patterns
•Roughly 10 repetitions are sufficient!
•Learning appears to be all-or-none
•Lasts for weeks at least
Auditory Noise Learning with STDP
Olivier Bichler, Thesis
Modified STDP rule•Post synaptic spike - depresses all synapses except those activated recently
Auditory Noise Learning with STDP
Auditory Noise Learning -ERP data
Perceptual learning of acoustic noise generates memory-‐evoked potentials Thomas Andrillon, Sid Kouider, Trevor Agus & Daniel Pressnitzer (Currrent Biology, accepted)
Selective Brain Responses to Meaningless noise in just 5 presentations
Selectivity MechanismsThreshold
Controlling Sparsity
• With Temporal Coding
• Easy to control the percentage of neurons that fire
• 1%, 2%, 5%, 10%
12345678910
Controlling Sparsity
• With Temporal Coding
• Easy to control the percentage of neurons that fire
• 1%, 2%, 5%, 10%
1,00E+00&
1,00E+02&
1,00E+04&
1,00E+06&
1,00E+08&
1,00E+10&
1,00E+12&
1,00E+14&
1,00E+16&
1,00E+18&
1,00E+20&
1,00E+22&
1,00E+24&
1,00E+26&
1,00E+28&
1,00E+30&
0& 10& 20& 30& 40& 50& 60& 70& 80& 90& 100&
Combina(ons*with*M*=*100*
12345678910
N of M coding
• Steve Furber
• With 100 neurons• 2100 binary patterns• ~1030
• What about controlling the proportion of active neurons?
1,00E+024,95E+031,62E+053,92E+067,53E+071,19E+091,60E+101,86E+111,90E+121,73E+13
Generating Selectivity• Suppose that the neuron has 10 synapses
• Suppose that the percentage of active inputs is fixed at 10%
• What is the likelihood of having a given number of synapses active?
• With a threshold of 4 the neuron would only have a 1% chance of firing with a random input
• With a threshold of 5, the probability drops to 0.1%
Interim Conclusion - Learning
• Error back-propagation is an unrealistic way to determine the selectivity of neurons in the visual system
• But simple spike-based learning mechanisms could allow the system to generate the appropriate selectivity in an unsupervised way
• Can we build a visual system without label based training??
Memory Mechanisms in Man & Machine
1. Humans can recognise visual and auditory stimuli that they have not experienced for decades.
2. Recognition after very long delays is possible without ever reactivating the memory trace in the intervening period.
3. These very long term memories require an initial memorisation phase, during which memory strength increases roughly linearly with the number of presentations
4. A few tens of presentations can be enough to form a memory that can last a lifetime.
5. Attention-related oscillatory brain activity can help store memories efficiently and rapidly
10 provocative claims
Memory Mechanisms in Man & Machine
6. Storing such very long-term memories involves the creation of highly selective "Grandmother Cells" that only fire if the original training stimulus is experienced again.
7. The neocortex contains large numbers of totally silent cells ("Neocortical Dark Matter") that constitute the long-term memory store.
8. Grandmother Cells can be produced using simple spiking neural network models with Spike-Time Dependent Plasticity (STDP) and competitive inhibitory lateral connections.
9. This selectivity only requires binary synaptic weights that are either "on" or "off", greatly simplifying the problem of maintaining the memory over long periods.
10. Artificial systems using memristor-like devices can implement the same principles, allowing the development of powerful new processing architectures that could replace conventional computing hardware.
10 provocative claims
STDP and Spiking Neurons
• “Intelligent” sensory processing with small numbers of neurons connected to spiking sensory devices• 2 layers of neurons connected to a spiking retina
• Unsupervised learning of car counting
• A single neuron connected to a spiking cochlear
• Unsupervised learning of auditory noise patterns
• What would happen with 16 billion neurons and multiple sensory inputs?
16 billion neurons
2 million optic nerve fibres
60,000 auditory nerve fibres
Grandmother Cells?
• Jerzy Konorski (1967)
• “gnostic neurons”
• Horace Barlow (1972)
• “cardinal cells”
• Alberta Gilinsky (1984)
• “cognons”
A Jennifer Anniston Cell !
Grandmother Cells in Man?
• Halle Berry• The Taj Mahal• Bill Clinton• Saddam Hussein• The Simpsons• The patient’s brother• Members of the research
team• ...
Temporal Lobe Recordings in Humans
“Josh Brolin” “Marilyn Munroe”
• The hit rate (0.54%) is too high
• Assume 109 neurons in the hippocampus
• Assume 10-30 000 identifiable objects (Irv Biederman)
• Each neuron should respond to 100-150 different objects
• Number of identifiable objects probably much higher : 105 - 106
• Rafi Malach : “Totem Pole Cells”
• Alternative Hypothesis
• The hippocampus keeps track of a few thousand objects and events that have been experiences recently
Grandmother Cells in Man?
How to make a Grandmother Cell
• STPD allows neurons to become rapidly selective to repeated stimuli• This can happen in a few tens of presentations• Oscillatory activity could allow this to happen in a single presentation
• Controlling the percentage of active inputs guarantees that the neuron will not fire to random inputs
• Setting a high threshold means that the neuron will never fire unless the stimulus used for training was presented again
• It could maintain its selectivity for several decades
• Cortical Dark Matter?
Dark Matter in Cortex?• Question
• What is the true distribution of firing rates in the cortex?
• Are there neurons that never fire for very long periods of time?
• Problem• Nearly all single unit neurophysiological studies are biased towards neurons that
have spontaneous activity
• The number of neurons recorded seems far lower than would be expected• Only 5-15 neurons recorded per descent (out of hundreds)
• Possible methods• Implanted arrays
• 2-photon optical imaging
• Patch clamp recording
• Roughly 50% of output neurons have firing rates below 0.1 Hz!• Maybe the output neurons are a special class??
• Dark matter in cortex??
Very low firing rate neurons
Invisible Cells?
Overview
Part 1
• Ultrarapid visual categorization
• Biological vs Computer Vision
• Temporal Constraints
• Coding with Spikes
• Convolutional Neural Networks in 1999
• SpikeNet
• The current state of the art in Convolutional Neural Networks
• Supervision & GoogleNet
Part 2
• Biologically inspired learning
• Spike-Time Dependent Plasticity (STDP)
• Applications in Vision
• Applications in Audition
• Development of Neural Selectivity
• Memories that can last a lifetime
• Grandmother Cells
• Neocortical Dark Matter
Final Conclusions
• Temporal Constraints on Vision
• Feedforward Neural Networks
• SpikeNet
• Supervision & GoogleNet
• Importance of Spikes
• Ultrafast coding with one spike per neuron
• Very efficient STDP schemes for unsupervised learning
• Biological Vision has been and will continue to be a source of inspiration for technology