1 R. Rao, Week 5: Neurobiology CSE 599 Lecture 5: Neurobiology F Why study neurobiology (if you are a computer scientist/software engineer/hardware designer)?

1R. Rao, Week 5: Neurobiology

CSE 599 Lecture 5: Neurobiology

Why study neurobiology (if you are a computer scientist/software engineer/hardware designer)?

Animal brains routinely solve problems that we would like computers to solve, e.g. computer vision, speech understanding, robot navigation, etc.

Neurobiology provides a model for machine intelligence and learning, e.g. artificial neural networks, learning algorithms, etc.

Neurobiology can provide new paradigms for computer design, e.g. parallel computing, graceful degradation, fault-tolerant computing, etc.


“I suspect that a deeper mathematical study of the nervous system will affect our understanding of the aspects of mathematics itself that are involved. In fact, it may alter the way in which we look on mathematics and logics proper.”

-- John von Neumann (The Computer and the Brain, 1958)

A quotable quote…


The Church-Turing Thesis and Neurobiology

Church-Turing Thesis

No general model of algorithmic computation is more powerful than a Turing machine

Are animal brains algorithmic computers?

“There is no doubt that the brain can perform algorithmic computations, but that does not mean that its underlying computational mechanism is algorithmic. It is quite possible that the brain can perform computations not expressible with Turing machines”

Fundamentals of the Theory of ComputationRay Greenlaw and H. James Hoover, p. 85


Physical basis of computation in animal brains

Animal brains are physical machines Built from hydrocarbons and ionic solutions Basic processing elements are called neurons

Animal brains represent information using physical quantities Real-valued electrical and chemical signals, with noise Transmitted on real organic wires or in real chemical solutions

Do a small set of primitives underlie neuronal computing? Digital computers

Transistors switches Boolean algebra symbol processing machines

Neurobiology ??? neurons? networks? ... adaptive sensory-motor

machines


Neurobiology and Digital Computing

Comparing neurobiology and silicon-based digital computers Device count:

Human brain: 1011 neurons and 1015 synapses (connections) between neurons (each neuron ~ 104 connections)

IC: 109 transistors with sparse connectivity Device speed:

Biology has 100µs temporal resolution Digital circuits will have a 100ps clock (10 GHz)

Computing paradigm: Animal brains employ massively parallel computation with local

and global feedback, and adaptive connectivity Digital computers: sequential information processing via CPU

with fixed connectivity Capabilities:

Digital computers will always be better at math… Will animal brains always be better at speech or vision?


A starting point

Hypothesis:Hypothesis: Digital computers and animal brains are efficient Digital computers and animal brains are efficient at computing in their respective domains, as a consequence of at computing in their respective domains, as a consequence of their underlying information representationtheir underlying information representation

Hypothesis:Hypothesis: The primitives underlying neuronal computation The primitives underlying neuronal computation are simpleare simple Neuronal structure and action-potential (spike-based) signaling are

conserved across the animal kingdom Just like transistor switches and Boolean algebra are conserved

across all digital computers

A starting point: Neurons, action-potentials, and synapses


Primary computational units: Neurons

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, pg. 21


Basic Input-Output Transformation

Input Spikes

Output Spike

(Excitatory Post-Synaptic Potential)


Example of signaling in a sensory neuron



Mechanisms of electrical signaling in neurons

Neuron cell membrane is a lipid bilayer Impermeable to charged ion species

such as Na+, Cl-, K+, and Ca2+

Each neuron maintains a potential difference across its membrane Typically –70 to –80 mV [Na+], [Cl-] and [Ca2+] higher

outside the cell; [K+] and organic

anions [A-] higher insideFrom Kandel, Schwartz, Jessel, Principles of Neural

Science, 3rd edn., 1991, pg. 67


Membrane proteins allow current flow

Proteins in membranes act as pores or channels that are ion-specific E.g. Pass K+ but not Cl- or Na+

These ionic channels are gated Voltage gated: Probability of opening

depends on membrane voltage Chemically gated: Neurotransmitter

binding causes channel to open e.g. Electron microscope picture of

an ACh (acetyl choline) channel Mechanically gated: Sensitive to

pressure or stretch

Ionic pump expels Na+ ions out of cell and takes K+ ions in (consumes ATP) Maintains –70mV potential difference From Kandel, Schwartz, Jessel, Principles of Neural

Science, 3rd edn., 1991, pgs. 68 & 137



Channel opening is probabilistic and discrete

Patch clamping (recording from an isolated patch) allows measurement of current through a single channel

Current influx is probabilistic and determined by: Probability of opening Duration of opening

Population of channels conveys smooth, graded membrane currents



Gated channels allow neuronal signaling

Inputs to a neuron change its local membrane potential via chemically-gated channels at “synapses” (connections).

Changes in local membrane potential are integrated spatially and temporally in dendrites and soma of the neuron.

Changes in membrane potential trigger opening/closing of voltage-gated channels in dendrites, soma, and axon, causing depolarization (positive change in voltage) and hyperpolarization (negative change).

When a large positive change in membrane potential occurs, crossing a particular voltage threshold, a spike or action potential is generated and transmitted to other neurons.


Neuronal Output: Action potentials

Voltage-gated channels cause action-potentials (spikes)Rapid Na+ influx causes

rising edgeNa+ pores deactivateK+ outflux restores

membrane potential

Positive feedback causes spikeNa+ influx depolarizes

membrane, causing more Na+ influx From Kandel, Schwartz, Jessel, Principles of Neural

Science, 3rd edn., 1991, pg. 110


Here is anotherillustration of thesame concept.

The permeability changesresult in the large swingsof membrane potential that are shown above.


An increase in permeability at one location of the membrane can spread to neighboring locations

Axons have very large concentrations of voltage-gated Na+ channels, causing the excitation to actively travel forward.



[lower resistance][lower resistance]


Vertebrates have developed another method of speeding up spike propagation, by adding a wrapping of myelin. This forces the electric current further down the axon, as it can only conduct where the resistance is low ---that is, at the node of Ranvier


Active Wiring: Myelination of axons

Schwann cells (glia) enable long-range spike communication Active wire allows lossless signal

propagation, unlike electric signals in a copper wire

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, pgs. 23 & 44


Communication between neurons: Synapses

Synapses: Connections between neurons Electrical synapses (gap junctions) Chemical synapses (use

neurotransmitters)

Synapses can be excitatory or inhibitory

Synapses are integral to memory and learning


Distribution of synapses on a real neuron…


postsynaptic element, such asanother neuron

postsynaptic element, such asanother neuron



Synaptic biochemistry

Note: even this isa gross simplification!

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991


Electron micrographs of synapses


Synaptic Vesicles and

Synaptic Cleft

Fusion of vesicles and release of

neurotransmitter

Retrieval andformation of new vesicles


Importance of Synapses

The gap between the axon and postsynaptic membrane (the synaptic cleft) allows electrical isolation of neurons

Hypothesis: Synaptic plasticity forms the basis of learning and memory



Synaptic plasticity: Adapting the connections

Long Term Potentiation (LTP): Increase in synaptic strength that lasts for several hours or more Measured as an increase in the excitatory postsynaptic potential

(EPSP) caused by a presynaptic spike

Increase in size of EPSP observedwhen the same presynaptic input is activated before and after LTP


Types of Synaptic Plasticity

Hebbian Long Term Potentiation: synaptic strength increases after prolonged pairing of presynaptic and postsynaptic spiking (correlated firing of two connected neurons).

Long Term Depression (LTD): Reduction in synaptic strength that lasts for several hours or more Anti-Hebbian LTD: Correlated spiking of two connected neurons

decreases the strength of their connecting synapse Homosynaptic LTD: Presynaptic spiking without postsynaptic

spiking Heterosynaptic LTD: Postsynaptic spiking without presynaptic

spiking

Spike-Timing Dependent Plasticity: LTP/LTD depends on relative timing of pre/postsynaptic spiking


Types of Synaptic Plasticity


Examples of measured synaptic plasticity

- and + are thresholds for LTD and LTP respectively

Hebbian LTP


Spike-Timing Dependent Plasticity

Amount of increase/decrease in synaptic strength (LTP/LTD) depends on relative timing of pre/postsynaptic spikes

LTP

LTD

pre before postpre after post


The mechanisms of neuronal structure and function are being rapidly unraveled via molecular, imaging, and electrophysiological techniques

But our knowledge of how networks of neurons give rise to perception, action, cognition, and

consciousness remains sketchy at best.

We will review some of this next…


5 minute break…

Next: Brain organization and information processing in networks of neurons


Organization of the Nervous System

CentralCentralNervous SystemNervous System

BrainBrain Spinal CordSpinal Cord

PeripheralPeripheralNervous SystemNervous System

SomaticSomatic AutonomicAutonomic


Skeletal/Somatic Nervous System

Nerves that connect to voluntary skeletal muscles and to sensory receptors

Afferent Nerve FibersAfferent Nerve FibersAxons that carry info away from the

periphery to the CNS

Efferent Nerve FibersEfferent Nerve FibersAxons that carry info from

the CNS outward to the periphery


Autonomic and Central Nervous System

Autonomic: Nerves that connect to the heart,blood vessels, smooth muscles, and glands

CNS: Brain + Spinal CordSpinal Cord:

• Local feedback loops control reflexes• Descending motor control signals from

the brain activate spinal motor neurons• Ascending sensory axons transmit

sensory feedback information from muscles and skin back to brain


Major Brain Regions: Brain Stem

Cer r Thalamus

Corpus collosumHypothalamus

l C t

Pons

Medulla

Spinal cord

Cerebellum

MedullaMedullaBreathing, muscle tone

and blood pressure

PonsPonsConnects brainstem withcerebellum & involved

in sleep and arousal

CerebellumCerebellumCoordination of voluntary

movements andsense of equilibrium


Major Brain Regions: Brain Stem

MidbrainMidbrainEye movements, visual and

auditory reflexes

Thalamus


l C

Pons

Medulla

Spinal cord

Cerebellum

Midbrain

Reticular FormationReticular FormationModulates muscle

reflexes, breathing & pain perception. Also

regulates sleep, wakefulness &

arousal


e Thalamus

Corpus co losHypothalamus

rebral Cortex

Pons

Medulla

Spinal cord

Cerebellum

Major Brain Regions: Diencephalon

ThalamusThalamusRelay station for all sensory

info (except smell)to the cortex

HypothalamusHypothalamusRegulates basic needs

fighting, fleeingfighting, fleeingfeeding, andfeeding, and

matingmating

Corpus

callosum


Major Brain Regions: Cerebral Hemispheres

Consists of: Cerebral cortex, basal ganglia, hippocampus, and amygdala

Involved in perception and motor control, cognitive functions, emotion, memory, and learning

Thalamus


rebral Cortex

Pons

Medulla

Spinal cord

Cerebellum

Cerebrum/Cerebral CortexCerebrum/Cerebral Cortex


Cerebral cortex comprises a sheet of neurons

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, pgs.

Cerebral Cortex: Convoluted surface of cerebrum about 1/8th of an inch thick

Six layers of neurons

Approximately 30 billion neurons + 270 billion Glial cells

Each nerve cell makes about 10,000 synapses: approximately 300 trillion connections in total


Summary



How do these brain regions interact to produce cognition and behavior?

Current knowledge is based on electrophysiological, imaging, molecular, and psychophysical techniques in conjunction with anatomical and lesion (brain damage) studies


The brain is specialized by region: Language

Broca’s Area

Important inthe

productionof speech

Wernicke’sArea

Important inthe

comprehensionof language


Specialization of function in Cerebral Cortex

VisualProcessing

somatosensorycortex

Motor Planning,Higher cognitive

functions

Visual and auditoryrecognition

Spatial reasoning

and motion


Specialization is based on connectivity

Brain regions perform elementary operations Basic computing units are neurons and synapses Specialization arises from differences in local connectivity, numbers

of neurons, and input-output connectivity to/from areas

Complex behavior arises from the interaction of multiple brain regions Example: Damage to Broca’s area

Person can understand language Person can say words or sing Person can’t speak or write grammatically


Regional activity is problem-dependent

PET scan of human brain during visual stimulation PET scan measures local blood glucose uptake

Visual stimulation excites visual cortex Regional activation depends on stimulus complexity

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, p. 315


The brain tackles complexity hierarchically

Example: Motor system Reflexive responses are handled by the spinal cord Movement is handled by the cerebellum Activity is scheduled by the cortex

Example: Speech learning by children Babies learn sounds (phonemes), then letters Toddlers learn words, then sentences Children learn grammar Teenagers learn composition


Hierarchical processing in the visual system

Retina Thalamus (LGN)

Primary Visual Cortex (V1)

V2 …


Hierarchical Organization of Visual Cortex


Hierarchical Visual Processing

Photoreceptors are sensitive to lightRGB wavelength color sensitivity

Retinal ganglion cells are sensitive to simple patterns

Figure: Color-opponent retinal cells detect color spots in visual field

Successively higher levels of visual processing detect successively more complex features

First in LGN, then in cortex

First spots/lines/bars and movement, then complex features and motion, then objects/faces and spatial reasoning



The Visual ProcessingHierarchy


Predictive coding in the retina

From Ward, Sensation and Perception, 3rd edn., 1989

Retinal Ganglion cells act as predictive coders:

On-center cell: Sensitive to light turning on in the center of the field

Off-center cell: Sensitive to light turning off in the center

Predictive coding: only relative differences matter - predict the central pixel values from surrounding pixels, and send the difference. “Efficient Coding”

Why does it work? Correlations in natural images.


Visual Illusion 1

Which ring is brighter?

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, pgs. 411

Both rings have the same hue: visual system measures brightness relative to background


Neurobiology uses relative measures

A Key point:

The brain doesn’t use absolute measures or values

Relative values, adaptation, and habituation appear at all levels of neuronal processing


Visual Illusion 2

Both women are the same size in both photosOur brain compares objects using relative perspective

measures

From Kandel, Schwartz, Jessel, Principles of Neural Science, 3rd edn., 1991, pgs. 411 & 444


An important open question

Does the timing of a spike matter or is a stimulus encoded in terms of the firing rates of neurons (no. of spikes/second)?

Traditional Answer: Firing rate Motor neurons in spinal cord appear to use a rate code High variability of a neuron’s response: for the same stimulus, a

neuron emits spikes at different times but at the same firing rate Is the variability stochastic noise (synapses, channels, etc.) or is our

definition of “stimulus” too restrictive?? Maybe we are not considering other inputs to the neuron?

Recent results: Spike timing important too? Spike-timing dependent plasticity Certain auditory areas do use spike timing for localization Reliable spiking if we inject fluctuating current into a neuron in slice In some animals, correlations between spiking neurons are crucial


Example: Spike correlations allow recognition

Locust olfactory system

A Neuronal spiking phase-locks when insect detects an odor

B Toxins end the synchronization, but not the spiking

Figures courtesy Kate MacLeod, Caltech


Conclusions

The brain computes! Information storage: The physical/chemical structure (synapses,

wires) Information transmission: Electrical and chemical signaling Computing elements: Neurons Computational basis (information coding, functions): Unknown

We can understand neuronal computation by discerning the underlying primitives A starting model: Artificial neural networks

Types of artificial neural networks Learning rules

More sophisticated models: Compartmental/biophysical models Computational models of specific brain areas and their interactions


Next week: Neural networks and models

DNA computing homework due next class…

Reading: See on-line chapters/articles for more information on today’s topics (Textbook/MIT on-line encyclopedia)

Mini-Project: Project ideas are up on the web site. Select one of these or come up

with your own topic – send email to check with instructor and TA Sign-up for 10-min presentation time slot on class web site

Have a great weekend!

Documents

1 R. Rao, Week 5: Neurobiology CSE 599 Lecture 5: Neurobiology F Why study neurobiology (if you are a computer scientist/software engineer/hardware designer)?