Biomimetic signal processing

Download Report

Transcript Biomimetic signal processing

STOCKHOLM

BRAIN

INSTITUTE

Cortex modeling and cortex inspired computation

Anders Lansner Dept of Computational Biology KTH and Stockholm University

Synopsis

STOCKHOLM

BRAIN

INSTITUTE •

Methods in neuronal network modeling

Large-scale cortex model example

Perspectives on modeling and brain inspired computing

November 15, 2007 Albanova Instrumentation Seminar 2

Goals

STOCKHOLM

BRAIN

INSTITUTE • •

Models of neurons and neuronal networks

• • •

1985 … Today high demand from neuroscience labs Enables understanding of the brain Brain-like/inspired algorithms and architectures

Beyond ”neural networks”, ”neurocomputing”

”Artificial brains” … on silicon November 15, 2007 Albanova Instrumentation Seminar 3

STOCKHOLM

BRAIN

INSTITUTE

Cortical areas and microcircuits

November 15, 2007 Albanova Instrumentation Seminar 4

Advances in experimental neuroscience

STOCKHOLM

BRAIN

INSTITUTE •

Shortage of data, but rapid development…

E.g. genetic fluorescent marking + confocal tracing of pathways

Livet et al. Nature Nov 2007 November 15, 2007 Albanova Instrumentation Seminar 5

Models at multiple levels

STOCKHOLM

BRAIN

INSTITUTE • • • • •

(Molecular dynamics) Sub-cellular level models Single neuron and synapse models Microcircuits and networks Full-scale global network models November 15, 2007 Albanova Instrumentation Seminar 6

Types of neuron models

STOCKHOLM

BRAIN

INSTITUTE • •

Summing threshold units Connectionist model

neural network

Integrate-and-fire

Hodgkin-Huxley formalism November 15, 2007 Albanova Instrumentation Seminar 7

STOCKHOLM

BRAIN

INSTITUTE

Single cell models - signal processing

An equivalent electrical circuit model November 15, 2007 Albanova Instrumentation Seminar 8

STOCKHOLM

BRAIN

INSTITUTE

Equivalent electrical circuit of a membrane patch

Ohm’s law: Nernst eqn:

I i

i m

E i

)

E i

RT zF

ln

C out C in

November 15, 2007 Albanova Instrumentation Seminar 9

The gate model

”Hodgkin-Huxley model”

STOCKHOLM

BRAIN

INSTITUTE  open  

y

closed 1 

y

First-order kinetics yields:

dy

   1 

y

  

y dt y

       

dy

     

y

dt

y

 

y

   1   

p

independent gating particles: K + : 

n

(

V

)  0 .

01  

V

  

e

V

 10 10  10   1    

n

(

V

) 

V

 0 .

125

e

80

y p

November 15, 2007 Albanova Instrumentation Seminar 10

STOCKHOLM

BRAIN

INSTITUTE

The Hodgkin-Huxley current equation

November 15, 2007

C m dV j dt

 

k G k

E k

V j

 

V j

 1 

V j R j

 1 ,

j

  

V j

 1

R j

, 

V j j

 1 

Albanova Instrumentation Seminar 11

STOCKHOLM

BRAIN

INSTITUTE

An action potential

Nobel Prize 1963

November 15, 2007 Albanova Instrumentation Seminar 12

Synaptic transmission

STOCKHOLM

BRAIN

INSTITUTE • • • • • •

Simple conductance based model

• •

Square pulse, Gamma function Voltage dependence (NMDA) Detailed model of single spine

Postsynaptic receptor kinetics

• •

Biochemical networks Neuromodulation Electrical synapses Graded transmitter release Synaptic plasticity

Short-term, ms - s

Long-term, s – yrs … November 15, 2007 Albanova Instrumentation Seminar 13

Real neuronal networks

STOCKHOLM

BRAIN

INSTITUTE • • •

Several types of different neurons Huge numbers Modules and layers

• •

Quite similar over areas and species!

Computing power limitation … November 15, 2007 Albanova Instrumentation Seminar 14

Simulators and simulation of large-scale models at KTH

STOCKHOLM

BRAIN

INSTITUTE •

GENESIS

NEURON

SPLIT simulator

Hammarlund & Ekeberg 1998

SPLIT parallel setup, optimization

Djurfeldt et al. 2005

• •

PGENESIS, parallel NEURON PDC/KTH

• •

Lenngren, KTH/PDC Blue Gene/L

1024 dual core nodes (1/64 of full machine) November 15, 2007 Albanova Instrumentation Seminar 15

STOCKHOLM

BRAIN

INSTITUTE

A large-scale cortex model

November 15, 2007 Albanova Instrumentation Seminar 16

STOCKHOLM

BRAIN

INSTITUTE

Hebbian synapses and cell assemblies Hebb D O, 1949: The Organization of Behavior

”LTP” Bliss and Lömo, 1973 Levy and Steward, 1978 • • • •

Cell assembly = mental object Gestalt perception

• •

Perceptual completion Figure-background separation

Perceptual rivalry

Milner P: Lateral inhibition After activity

 •

500 ms Persistent, sustained

Fatigue = Adaptation, synaptic depression Association chains

Temporally asymmetric synaptic plasticity Albanova Instrumentation Seminar 17 November 15, 2007

The KTH layer 2/3 model

STOCKHOLM

BRAIN

INSTITUTE

70% -1.5 mV mV 70% 1.2 mV 70% 2.5 mV 25% 2.4 mV November 15, 2007

• 2

30% 0.30 mV

1

17% 2.5 mV Top-down driven model of associative memory

• • • •

Generic “association cortex”, layers 2/3 Modular: Minicolumns, hypercolumns 3 different cell types: Pyramidal cells, Basket cells, Regular Spiking Non-Pyramidal 2 000 – 20 000 000 model neurons Albanova Instrumentation Seminar 18

STOCKHOLM

BRAIN

INSTITUTE

Local RSNP Distant pyramidal

Neuron-synapse properties

Local basket cell November 15, 2007 Local pyramidal Tsodyks, Uziel, Markram 2000

• • • • •

Realistic amplitude of PSP:s in largest network model Sparse connectivity (stochastic) Synaptic depression Asymmetric cell-cell connectivity 3D geometry

delays

0.1 - 1m/s conduction speed Albanova Instrumentation Seminar 19

STOCKHOLM

BRAIN

INSTITUTE

One of the 9 hypercolumns Active minicolumn (30 pyramidal cells) Active basket cell Active RSNP cells

Network layout

• • • • • •

1x1 mm patch 9 hypercolumns Each hypercolumn

• •

100 minicolumns 100 basket cells 100 patterns stored 29700 neurons 15 million synapses November 15, 2007 Albanova Instrumentation Seminar 20

STOCKHOLM

BRAIN

INSTITUTE

9 hypercolumns

• • • • • •

1x1 mm patch 9 hypercolumns Each hypercolumn

• •

100 minicolumns 100 basket cells 100 patterns stored 29700 neurons 15 million synapses November 15, 2007 Albanova Instrumentation Seminar 21

STOCKHOLM

BRAIN

INSTITUTE

100 hypercolumns

• •

330000 neurons 161 million synapses

4x4 mm

22 November 15, 2007 Albanova Instrumentation Seminar

8 rack BG/L simulation

STOCKHOLM

BRAIN

INSTITUTE • • • • • • • •

22x22 mm cortical patch

22 million cells, 11 billion synapses 8K nodes, co-processor mode

used 360 MB memory/node Setup time = 6927 s Simulation time = 1 s in 5942 s >29000 cpu hours Massive amounts of output data 77 % of linear speedup

Point-point communication slows (?) Currently (inofficial) world record!

Djurfeldt M, Lundqvist M, Johansson C, Rehn M, Ekeberg Ö, and Lansner A (2007): Brain-scale simulation of the neocortex on the Blue Gene/L supercomputer. IBM J R&D (in press) November 15, 2007 Albanova Instrumentation Seminar 23

STOCKHOLM

BRAIN

INSTITUTE

The three different cell types

3 sec simulation Pyramidal RSNP Basket 24 November 15, 2007 Albanova Instrumentation Seminar

STOCKHOLM

BRAIN

INSTITUTE • • •

2000+ neurons 250000+ synapses 5 s = 600 s on PC November 15, 2007 Lundqvist M, Rehn M, Djurfeldt M and Lansner A (2006). Attractor dynamics in a modular network model of the neocortex. Network: Computation in Neural Systems: 17, 253-276 Albanova Instrumentation Seminar 25

STOCKHOLM

BRAIN

INSTITUTE

Perception and associative memory performance

• • • •

Pattern reconstruction

• • •

Figure-background Pattern completion and rivalry 50 – 100 ms Sustained after-activity

• •

150 ms – 2 sec NMDA Ca , K Ca modulation Robust to parameter changes and scaling Cortical long-range recurrent excitation strong enough to support attractor dynamics November 15, 2007 Albanova Instrumentation Seminar 26

STOCKHOLM

BRAIN

INSTITUTE

Attractor dynamics:

Pattern rivalry November 15, 2007 Fast ”decision” <100 ms!

Albanova Instrumentation Seminar 27

STOCKHOLM

BRAIN

INSTITUTE

Bimodal membrane potential

Log(p ISI ) Exponential fit Jeffrey Anderson, Ilan Lampl, Iva Reichova, Matteo Carandini, and David Ferster. Stimulus dependence of two-state fluctuations of membrane potential in cat visual cortex.

Nat. Neurosci.,

3(6):617 – 621, 2000.

November 15, 2007 Albanova Instrumentation Seminar 28

Bistable activity with irregular firing, similar to in vivo recordings

STOCKHOLM

BRAIN

INSTITUTE •

Ground state stable only in larger networks with many patterns stored

Increase in irregularity in active cortical states is a challenges for persistent activity models

This L2/3 network model

• •

displays irregular fluctuation driven low-rate firing operates in a high-conductance regime of balanced excitatory and inhibitory currents

is stable to synchronization even with blocked NMDAR

Details under investigation November 15, 2007 Albanova Instrumentation Seminar 29

STOCKHOLM

BRAIN

INSTITUTE

Attentional blink – effect of GABA

↑ GABA baseline GABA 150%

100 80 60 40 20 0 0 40 80 120 160 200

milliseconds

240 280 320 • •

Attractor activation correlates with percentage of correct probe detections Time scales different but qualitatively similar results November 15, 2007 Albanova Instrumentation Seminar 30

Ongoing work

STOCKHOLM

BRAIN

INSTITUTE • • • • + +

Layer 4

Selective feature detectors

V1 model with

learned orientation map (LISSOM)

• 

patchy horizontal L2/3 connectivity Layer 5

Martinotti cells, local (delayed) inhibition to superficial layers

Pyramidals, cortico-cortical connections Analysing L2/3 dynamics, spiking statistics, conductances, intracellular potentials

Non-orthogonal stored memories Better synthetic VSD, BOLD signals Modelling interacting areas … using parallel NEURON Scalable abstract connectionist cortex model

Cortical area module, on-line learning, network-of networks,… November 15, 2007 Albanova Instrumentation Seminar 31

STOCKHOLM

BRAIN

INSTITUTE

Computing Power

Moore’s law …

1E +1 0 1E +0 9 1E +0 8 1E +0 7 1E +0 6 1E +0 5 10 00 0 10 00 100 10 1 1980 0, 1 100 ops/synapse/ms

IBM BlueGene/L 128K cores

1985 1990 1995 2000 2005 2010 2015 2020 2025 year

?

Next generation supercomputers >1M cores November 15, 2007 Albanova Instrumentation Seminar 32

STOCKHOLM

BRAIN

INSTITUTE

EU/FACETS – analog VLSI

From cortex physiology to VLSI EU/GOSPEL – NoE in Artificial olfaction SSF/Stockholm Brain Institute (SBI) OECD/INCF – International Neuroinformatics Coordinating Facility November 15, 2007 Albanova Instrumentation Seminar 33

Conclusions

STOCKHOLM

BRAIN

INSTITUTE •

Computational models are enabling tools in brain science

• •

Human brain level computing power in 10-15 yrs Brain mysteries likely to be largely uncovered at that time

A principled understanding of brain function will emerge

Great benefits!

Brain-like computing and AI

Consequences for society…?

November 15, 2007 Albanova Instrumentation Seminar 34

STOCKHOLM

BRAIN

INSTITUTE

Collaborators

• • •

Model development

Mikael Lundqvist, PhD student

David Silverstein, Phd student Parallel simulation

Mikael Djurfeldt , PhD student

Örjan Ekeberg, Assoc Prof Data analysis

Martin Rehn , postdoc November 15, 2007 Albanova Instrumentation Seminar 35