Theoretical neuroscience:
Single neuron dynamics and computation
STAT 42510 - CPNS 35510
Instructor: Nicolas Brunel [email protected]
TA: Max Gillett [email protected] Practical details
First course in a 3-course series:
• Single neuron dynamics and computation (Fall, STAT 42510 - CPNS 35510) Instructor: Nicolas Brunel
• Network dynamics and computation (Winter, STAT 42520 - CPNS 35520) Instructor: Nicolas Brunel
• Statistics and information theory (Spring, STAT 42600 - CPNS 35600) Instructor: Stephanie Palmer Practical details
• Problem sets will be issued every Tuesday from Week 2 to Week 9 (total of 8), due the next Tuesday in class.
• A take-home final will be issued on December 1, due on December 15 in my office (Jones 116A)
• Grades: Problem sets will account for 50% of the grade; The take-home final for the remaining 50%.
• TA: Max Gillett ([email protected])
• Office hours: Thursday 5-6pm, Jones 308 (Max Gillett); Friday 4-5pm, Jones 116A (Nicolas Brunel) Introduction
• Brains, scales, structure, dynamics
• What is computational/theoretical neuroscience?
• Brief history of computational/theoretical neuroscience
• Outline of the course The human brain: a network of 1011 neurons connected by 1015 synapses Brain networks: from C Elegans to Humans
Animal # neurons # synapses
C Elegans 302 5,000
Fruit fly 105 107
Honey bee 106 109
Mouse 108 1012
Cat 109 1013
Human 1011 1015 Brain networks: from sensory to motor Spatial scales
∼10cm Whole brain
∼1cm Brain structure/cortical area
100µm- 1mm Local network
10µm- 1mm Neuron
100nm- 1µm Sub-cellular compartment
∼1-10nm Molecule Experimental tools The whole brain level: structures Cerebral cortex: network of areas Area level
• Areas are interconnected networks of local networks (‘columns’) Local network level
• Size ∼ cubic mm • Total number of synapses ∼ 109 • Total number of cells ∼ 100,000 (10,000 per neuron)
• Types of cells: • Cells connect potentially to all other cell – pyramidal cells - excitatory (80%) types (E→ E, E→ I, I→ E, I→ I) – interneurons - inhibitory (20%) • Connection probability ∼ 10% Neuron level
• Neuron = complex tree-like structures with many compartments (e.g. dendritic spines) Subcellular compartment level
• Subcellular compartments (e.g. dendritic spines) contain a huge diversity of molecules (in particular protein kinases and phosphatases) whose interactions define complex networks Temporal scales
Days-Years Long-term memory
Seconds-Minutes Short-term (working) memory
100ms - 1s Behavioral time scales/Reaction times
∼ 10ms Single neuron/synaptic time scales
∼ 1ms Action potential duration; local propagation delays
1ms Channel opening/closing Submillisecond
• Molecular time scales (channel opening/closing; diffusion of neurotransmitter in synaptic cleft; etc) Millisecond
• Width of action potentials; axonal delays in local networks Tens of ms
• Synaptic decay time constants; membrane time constant of neurons; axonal delays for long-range connections Hundreds of ms
• Behavioral time scales (e.g. motor response to a stimulus) Seconds-minutes
• Short-term memory Working memory Days-years
• Long-term memory Questions in theoretical/computational neuroscience
What? Describe in a mathematically com- pact form a set of experimental obser- vations.
How? Understand how a neural system produces a given behavior.
Why? Understand why a neural system performs the way it does, using e.g. tools from information theory. Mathematical tools
• Single neuron/synapse models, rate models: systems of coupled differential equations. Tools of dynamical systems (linear stability analysis, bifurcation theory)
• Networks: Graph theory, linear algebra. Large networks: tools of statistical physics
• Noise (ubiquitous at all levels of the nervous system): Statistics, probability theory, stochastic processes.
• Coding: Information theory A few historical landmarks
Before mid-1980s: a few isolated pioneers
• 1907: Louis Lapicque (leaky integrate-and-fire neuron)
• 1940s: Warren McCulloch, Walter Pitts (binary neurons, neural networks)
• 1940s: Donald Hebb (neuronal assemblies, synaptic plasticity)
• 1940s-50s: Alan Hodgkin, Andrew Huxley (model for AP generation)
• 1950s: Wilfrid Rall (cable theory)
• 1950s: Frank Rosenblatt (perceptron)
• 1960s: Horace Barlow (coding in sensory systems)
• 1960s-70s: David Marr (systems-level models of cerebellum/neocortex/hippocampus)
• 1970s: Hugh Wilson, Jack Cowan, Shun-Ishi Amari (‘rate models’)
• 1980s: John Hopfield, Daniel Amit, Hanoch Gutfreund, Haim Sompolinsky (associative memory models) A few historical landmarks
From the 1990s: establishment of a field
• First annual computational neuroscience conference in 1992 (now two main conferences/year, CNS and Cosyne, each around 500 participants per year)
• Many specialized journals created from the beginning of the 1990s (Neural Computation; Network; Journal of Computational Neuroscience; Frontiers in Computational Neuroscience; etc)
• Annual computational neuroscience summerschools since mid-1990s (Woodshole, ACCN in Europe, Okinawa, CSHL China)
• Birth of major Computational Neuroscience centers in the US (1990s, Sloan-Swartz), Israel (HU, 1990s) and Europe (Gatsby Unit, Bernstein Centers, 2000s)
• Now widely accepted as an integral part of neuroscience (theory papers routinely published in major neuroscience journals) Week 1 Sep 27 Introduction Sep 29 Biophysics, Hodgkin-Huxley (HH) model
Week 2 Oct 4 2D models for spike generation Oct 6 Ionic channels and firing behaviors
Week 3 Oct 11 From HH to leaky integrate-and-fire (LIF) neurons Oct 13 Stochastic dynamics of neurons
Week 4 Oct 18 From LIF to ‘firing rate’ and binary neurons Oct 20 No class
Week 5 Oct 25 Coding I Oct 27 Coding II
Week 6 Nov 1 Axons Nov 3 Dendrites
Week 7 Nov 8 Synapses I Nov 10 Synapses II
Week 8 Nov 15 Synaptic plasticity I Nov 17 Synaptic plasticity II
Week 9 Nov 22 Unsupervised learning Nov 24 No class
Week 10 Dec 29 Supervised learning Dec 1 Reinforcement learning Spatial scales of the brain
∼10cm Whole brain
∼1cm Brain structure/cortical area
100µm- 1mm Local network
10µm- 1mm Neuron
100nm- 1µm Sub-cellular compartment
∼ 1-10nm Molecule Single neuron models
• Hodgkin-Huxley neuron dV X C = −IL(V )− Ia(V, ma, ha)+Isyn(t) dt a
IL(V ) = gL(V − VL)
xa ya Ia(V, ma, ha) = ga(V − Va)ma ha
dma ∞ τm (V ) = −ma + m (V ) a dt a dha ∞ τ (V ) = −ha + h (V ) ha dt a • Integrate-and-fire neuron dV C = −gL(V − VL) + Isyn(t) dt
Fixed threshold Vt, reset Vr, refractory period; • How do single neurons work?
• What are the mechanisms of action potential generation?
• How do neurons transform synaptic inputs into a train of action potentials? Coding
Tools from information theory:
• Shannon mutual information
• Fisher information
• What contains information in single neurons spike trains?
• How much information is contained in single neurons’s response?
• What is the optimal way of transmitting information about a given input, given its statistics and other constraints? Are neural systems optimal/close to optimal at transmitting information? Single synapse models
• ‘Conductance-based’
X X k Ii,syn(t) = (V − Va) ga,i,j sa(t − tj ) a=E,I j,k
τas˙a = ...
• ‘Current-based’
X X k Ii,syn(t) = Ja,i,j sa(t − tj ) a=E,I j,k
Popular choices of s – Delayed difference of exponential; – Delayed delta function
• How do synapses work?
• What are the mechanisms of synaptic plasticity on short time scales? Synaptic plasticity
Synaptic efficacy can be modified in various ways: • Spike timing • Firing rate • Post-synaptic V (STDP experiments) (BCM, Sjostrom et al, etc) (pairing, etc)
• Can we capture this experimental data using simplified ‘plasticity rules’?
• What are the mechanisms of induction of synaptic plasticity? Learning
input patterns • N synapses/inputs; synaptic A B parallel weights fibres • sums linearly its inputs; w 0 1 1 1 • emits a spike if inputs > θ ∼ 10 w 1 0 2 2 mV (threshold) w 0 1 3 3 w 1 0 • has to learn a set of p ≡ αN in- 4 4 put/output associations
Purkinje cell • by appropriate choice of its synap- threshold B tic weights wi ≥ 0. (θ ± κ) 0 in a robust way (as measured by ); output • κ A 1 • can be achieved using simple learning algorithms • How much information can a single neuron/a population of neurons store?
• What kind of learning algorithms/rules allow to reach optimal storage? Useful books
• Tuckwell, “Introduction to Theoretical Neurobiology”, Vols. I & II (Cambridge U. Press, 1988)
• Amit, “Modeling brain function: The world of attractor neural networks” (Cambridge U. Press, 1989)
• Hertz, Krogh, and Palmer, “Introduction to the Theory of Neural Computation” (Addison-Wesley, 1991 - now from: Perseus Book Group and Westview Press)
• Rieke, Warland, de Ruyter van Steveninck, and Bialek, “Spikes: Exploring the Neural Code” (MIT Press, 1997)
• Koch, “Biophysics of Computation: Information Processing in Single Neurons” (Oxford U. Press, 1999)
• Dayan and Abbott, “Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems” (MIT Press, 2001)
• Izhikevich, “Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting” (MIT Press, 2007)
• Ermentrout and Terman, “Mathematical Foundations of Neuroscience” (Springer, 2010)
• Gerstner, Kistler, Naud and Paninski “Neuronal Dynamics: From single neurons to networks and models of cognition” (Cambridge U. Press, 2014)