Transcript Real Neurons for Engineers

Real Neurons for Engineers
(Lecture 2)
Harry R. Erwin, PhD
COMM2E
University of Sunderland
Take-Home Message
• What real neurons compute:
–
–
–
–
Phasic activity
Tonic activity
Interneural interaction
Plasticity
• How they compute it:
–
–
–
–
–
The role of channels
The role of neurotransmitters
The role of specialized synapses
The role of neuronal topology
The role of axonal delays
Resources
• Shepherd, G., ed., 2004, The Synaptic Organization of
the Brain, 5th edition, Oxford University Press.
• Nicholls et al.
• Kandel et al.
• Koch, 2004, Biophysics of Computation, OUP.
• Bower and Beeman, 1998, The Book of Genesis,
second edition, TELOS, ISBN: 0-387-94938-0
• Rieke et al, 1999, Spikes: Exploring the Neural Code,
Bradford Books.
• Churchland and Sejnowski, 1994, The Computational
Brain, Bradford Books.
What is a Neuron?
• A neuron is an ‘excitable cell’, like a muscle cell.
• Neurons are very primitive—found in most
animals.
• Neurons operate by allowing ions to pass through
their membranes. This changes ion concentrations
and the potential across their membrane. The ions
then function in various ways to cause changes in
the neuron.
• Bob will teach this. I will show you how to model
it.
Ions and Cells
Sodium (Na+)—outside
Potassium (K+)—inside
Magnesium (Mg++)—blocks NMDA receptors
Chlorine (Cl-)—plays various roles
Calcium (Ca++)—important in intercellular
communication.
• Most negative charges within neurons are bound
to proteins and respond to membrane potential
changes by moving a small distance.
•
•
•
•
•
Phasic activity
• A neuron is called ‘phasic’ if it responds to
synaptic input by generating one or more action
potentials in a short time.
• In the extreme, a phasic neuron can serve as a
coincidence detector. Such neurons tend to have
very negative resting potentials and short time
constants so that multiple synchronized inputs are
needed to trigger them.
Tonic activity
• A neuron is called ‘tonic’ if it responds to
activation by generating an extended sequence of
action potentials or enters a state where action
potentials are generated continuously.
• Tonic neurons tend to have resting potentials near
threshold and long time constants.
• Neuromodulation (dopamine, 5HT, etc.) controls
the state of these neurons.
Interneural Interaction
• Neurons can interact via chemical and
electrical synapses (gap junctions).
• Interaction via electrical synapses allows a
small network of neurons to respond to the
activation of one neuron.
• This can also prepare nearby neurons for
follow-on activation.
Electrical synapses
• Rare in the cortex
• Common in the retina and in the basal
ganglia
• Unknown presence in the auditory system
• Generally involve GABAergic cells in the
cortex
Plasticity
• Plasticity (a form of learning) involves changes in
synaptic weights, either short-term or long-term.
• Short-term plasticity tends to involve tonic
neurons and neuromodulation. It can also involve
recurrent signaling within a small network.
• Long-term plasticity is believed to involve
changes in receptor densities on the post-synaptic
side and vesicle densities on the pre-synaptic side.
Memory
• Short-term memory mechanisms
– Changes in vesicle count
– Slow time constant channel dynamics
– Changes in receptor counts
• Long-term memory mechanisms
– Changes in channel count
– Formation of new synapses/activation of silent synapses
• Associative memory
– Local recurrent networks in cortex
– Coincidence detectors in the basal ganglia
– Interacting areas—possibly chaotic in the hippocampus and olfactory
system
Caveat
• However, there is new evidence that synapses
have discrete synaptic states.
• See Montgomery and Madison, 2004, “Discrete
synaptic states define a major mechanism of
synapse plasticity,” Trends in Neurosciences,
27(12):744-750, December 2004.
• Glutaminergic synapses can be active (normal),
potentiated (increased AMPA receptor count),
depressed (reduced AMPA receptor count), silent
(no AMPA receptors expressed), and recently
silent (potentiated).
Transitions
• NMDA receptors for Glutamate facilitate the
transition from active to potentiated.
• mGlu receptors facilitate the transition to active
from potentiated.
• Depressed synapses may form a continuum.
• Silent synapses cannot be potentiated directly to
an active or potentiated state. They pass through
the recently silent state first.
• Recently silent synapses cannot be depressed.
The role of channels
• Potassium channels return the cell to a resting
state. They often control the overall time constant.
• Chloride channels may be inhibitory, shunting
(desensitizing) and even facilitatory. They tend to
have longer time constants.
• Sodium channels are typically depolarizing. Short
time constants.
• Calcium channels are typically depolarizing. Long
time constants. Used in signaling.
The role of neurotransmitters
• Glutamate is excitatory (AMPA, NMDA, Kainate)
and neuromodulatory.
• Aspartate is similar.
• Acetylcholine is excitatory and neuromodulatory.
• GABA and Glycine are often treated as inhibitory,
but they have other roles as well.
• Epinephrine, 5HT, and dopamine are
neuromodulatory. Many more, too.
The role of neuronal topology
• Pyramidal cells have multiple compartments:
–
–
–
–
Soma
Axon
Apical dendrites
Basalar dendrites
• The apical dendrite apparently communicates with
the soma using calcium spikes (I.e., active
conductances).
• Multiplicative interactions among synapses are
important.
The role of delay tuning
• For cells that a coincidence detectors,
tuning of axonal delays may play a role in
their computation.
• For example, azimuth is estimated in man
from relative arrival times of action
potentials.
• Echo delays can also be measured this way.
Real Neurons and What They Do
• Principal neurons (fast)
– Long-range transmission of signals, using action
potentials along a long axon
– Usually have local collaterals
• Interneurons (slower)
– Local processing
• Signal sharpening
• Stabilization of network activity
– Some have axons, some not. Some are binary,
forwarding signals to a dendrite down their axon.
• Neuromodulatory neurons (slowest)
– General control of activity over a large area
Some Neurotransmitters
(Bob covers)
• Glutaminergic neurons
– Excitatory
• GABAergic neurons
– Inhibitory, shunting, or facilitive
• Cholinergic neurons
– Excitatory
• Dopaminergic neurons
– Neuromodulatory
Persistent Activity
•
•
•
•
•
•
Not well-understood
Hot research area
Underlies short-term memory.
Related to learning.
Important in the cortex
Neurons with persistent activity are rarely
used in artificial neural networks, but are
important in producing behaviour.
Mechanisms of Persistent
Activity
• Network recurrence
– Neurons excite each other
– Inhibitory recurrence needed for stability
• Neuron-level persistent activity
– Long time-constant channels
– Causes the soma to depolarize repetitively
– Can produce bursting or periodic signals
Interneurons
• Functions
– Signal normalization/sharpening
– Network stabilization
– Motion sensitivity
• Amacrine type cells in the retina
• Longer-range interactions
Take-Home Message
• Neurons are complex and the brain uses that
complexity to do wonderful things.
• Don’t be afraid to make assumptions about how
neurons might do complex things if it allows your
model to do what you need it to do. It’s likely
you’re right.
• Write your own MATLAB models—what the
Neural Network Toolbox gives you is very
limited.
• GENESIS is intended to allow you to study neural
models before you simplify them for MATLAB.