NIPS*99 Spike Timing Workshop Schedule and Abstracts
7:30AM Larry Abbott and Paul
Munro
Welcome and introduction
8:15AM William B. Levy
The Time Span of Associative Synaptic Modification is Substantially
Amplified by Neural Interactions
Abstract:
The asymmetric time span of associative modification is useful for
storing information that creates predictions into the future.
Unfortunately, biophysical limitations preclude a sufficiently long time
span for many behaviorally-relevant predictions. Neuronal circuitry and
network dynamics can produce a 10 to 100-fold increase in this time
span.
8:40AM Mu-Ming Poo
Correlated Activity and Long-term Synaptic Modifications - Hebb's
Postulate Revisited
Abstract:
In search of a cellular mechanism for associative learning, Hebb (1949)
suggested that correlated excitation of pre- and postsynaptic neurons
may lead to strengthening of the synaptic connection between them.
Rigorous examination of the role of temporal patterns of pre- and
postsynaptic spiking activity in synaptic modifications has only
recently been carried out. I will summarize our efforts in studying the
role of spike timing in activity-induced long-term potentiation (LTP)
and long-term depression (LTD) in artificial networks of cultured rat
hippocampal neurons and in the developing frog tectum. In both systems,
we found that for repetitive spiking activities to induce LTP and LTD,
they must appear in the pre- and postsynaptic cells within a critical
time window. Synaptic inputs that are repetitively activated within 20
ms before the spiking of the postsynaptic neuron become potentiated,
while inputs activated within 20 ms after the postsynaptic spiking
become depressed. Both potentiation and depression depend on activation
of NMDA-subtype of glutamate receptors and can be readily reversed by
subsequent spiking activity of appropriate patterns. Thus correlated
pre- and postsynaptic spiking are required for both strengthening and
weakening of the synapse, but with an opposite requirement in the
temporal sequence of spiking.
Input specificity is another inherent concept in the Hebb's postulate:
Only synapses along the activated pathway become modified. I will also
summarize our recent studies on the issue of synapse specificity in
activity-induced synaptic modifications in defined networks of cultured
hippocampal neurons. Our results showed that induction of LTP and LTD
by repetitive stimulation of one connection in a defined network leads
to a "propagation" of potentiation and depression selectively to other
connections within the networks. Furthermore, we found evidence for
induction of LTP and LTD at selective "remote" synaptic sites distant
from the neurons that are repetitively stimulated. These remote
modifications can be accounted for by synaptic modification due to
positively- or negatively-correlated spiking at remote synaptic sites
where pre- and postsynaptic cells are activated through separate
pathways with transmission delays that happen to match
interpulse-intervals associated with the repetitive stimuli. Through
such a "delay-line" mechanism, temporal information coded in the timing
of individual spikes can be converted and stored into spatially
distributed patterns of persistent synaptic modifications in a neural
network. Taken together, our results argue that specific rules of input
specificity must be considered for activity-dependent synaptic
modifications at the network level.
9:05AM
BREAK!
BREAK!
9:20AM Richard Kempter
Intrinsic Rate Normalization by Hebbian Rules: Spike-Based
vs. Rate-Based Learning
Abstract:
Over a broad parameter regime, Hebbian learning leads to an
intrinsic normalization of the mean firing rate of the
postsynaptic neuron. Subtractive normalization of the
weights is guaranteed if, in addition, the mean input rates
are identical at all synapses. In a rate description,
intrinsic rate normalization is most easily achieved with a
negative factor in front of the correlation term in the
learning rule, often called `anti-Hebbian' learning. It is
shown that, for spike based learning, a distinction between
Hebbian and `anti-Hebbian' rules is no longer possible.
Learning is driven by correlations on the time scale of the
learning window which may be positive even though the
integral over the learning window is negative.
9:45AM Dan Feldman
Long-term plasticity induced by action potential-EPSP timing at layer
II/III pyramidal cells in rat somatosensory cortex.
Abstract:
Long-term potentiation and depression (LTP and LTD) induced by the
relative timing of EPSPs and postsynaptic spikes may contribute to
experience-dependent plasticity of neuronal response properties in vivo.
In the rat's primary somatosensory cortex (S1), plucking one whisker but
sparing its neighbor causes neuronal responses to the deprived whisker
in layer II/III to become rapidly depressed. This depression appears to
be driven in part by activation of the spared whisker and can be readily
explained by a form of timing-based, associative long-term plasticity
recently observed in these cells in vitro. An unusual feature of this
plasticity is that the range of spike-EPSP delays that induces LTD is
much longer than the range of delays that induces LTP. This observation
predicts that synapses that elicit subthreshold EPSPs in a manner
uncorrelated with postsynaptic spiking will, over time, become
depressed. In vivo, spontaneous spiking of inputs representing plucked
whiskers will therefore drive depression of these same inputs. This
mechanism may contribute to experience-dependent depression of sensory
responses in other cortical areas as well.
10:10AM Sen Song
Spike-Timing Dependent Synaptic Plasticity Enables Groups of
Correlated Neurons to Supervise Network Learning
Abstract:
Recent experiments have shown that the timing of pre- and post-synaptic
action potentials plays a key role in the modification of synaptic
efficacy.
One of the striking features of such a form of plasticity is the asymmetry
between pre- and post-synaptic neurons, namely the sharp transition from
potentiation to depression depending on the order of spiking. In a recurrent
network, this leads to the strengthening of the synapses formed by a highly
correlated cluster of neurons onto other cells. These synapses can carry an
instructional signal. A computational model is presented to illustrate this
phenomenon. A layer of recurrently connected neurons receives inputs from a
layer of neurons with Gaussian tuning curves. We present stimuli at random
positions for periods chosen from an exponential distribution. This leads to
the development of receptive fields at different random locations if the
recurrents are turned off. We next seed a small group of neurons in the
recurrent layer with feedforward connections tuned to the same location.
These neurons became correlated and form strong connections to other untuned
cells in the network. As a consequence of the recurrent connections, the
other cells also become responsive to this particular location, forming a
column-like structure. Once the proper feedforward connections are formed,
i.e. the instructor has done its job, the recurrent connections weaken.
Thus, in this example, the recurrent connections served as the vehicle for a
small group of "teacher" neurons to supervise the development of other
neurons in the network. The asymmetry in the learning is crucial in insuring
the one-way information flow from the "teacher" to the "student".
10:10AM Kevin Franks
Simulated Dendritic Ca Influx through Ligand- and Voltage-Gated
Channels
Abstract:
Calcium is an important intracellular second messenger in many
biological systems and is, seemingly paradoxically, necessary for the
induction of both LTP and LTD. We have used MCell, a Monte Carlo
simulation package to simulate the influx of Ca2+ into a postsynaptic
dendrite following pre- and postsynaptic activity and to track the
binding of Ca2+ to endogenous proteins and fluorescent sensors. The
order and spacing of the EPSP and back-propagating action potential has
a large effect on the influx of Ca2+, and thus on intracellular
Ca2+ concentration.
Calcium binding proteins can both transduce Ca2+-dependent signaling
cascades and buffer intracellular Ca2+ so as to dampen the
initiation of
such signals. Here we show how a specific, spatially restricted binding
protein, calmodulin, can serve as a differentiator of [Ca2+],
by virtue
of its kinetics. Such a differentiator might serve to transduce a
signal to either potentiate or depress a synapse depending on the
spatio-temporal profile of the Ca2+ signal.
10:35AM DISCUSSION
11:00AM LUNCH & FREE TIME
4:30PM
Mayank R. Mehta and Matthew A. Wilson
Effect of LTP on Spatio-Temporal Structure of Receptive Fields:
Hippocampal Lessons for V1
Abstract:
Recent works have investigated the effect of patterns of neuronal
activity and synaptic plasticity on the size and specificity
of receptive fields. In this work we have investigate the effect
of plasticity on the receptive field shape.
Here we show that within a few traversals of a familiar environment a
large
fraction of hippocampal place fields become asymmetric such that, the
firing rate of a place cell rises slowly as a rat enters a place field,
but the firing rate drops off abruptly at the end of the place field.
A computational model that can explain the results, based on NMDA
dependent LTP, is presented. The model provides a mechanism underlying
phase precession in the
hippocampal neurons and the development of inseparable spatio-temporal
and directionally selective receptive fields in the striate cortex. Such
asymmetric receptive fields can allow an animal to predict an upcoming
event, such as the location of a visual stimulus or the animal's location,
based on previously experienced patterns of neuronal activity.
4:55PM Misha Tsodyks
An Algorithm for Modifying Neurotransmitter Release Probability Based on
Pre- and Post-Synaptic Spike Timing
Abstract:
The precise times of pre- and post-synaptic action potentials play a
key role in the modification of the synaptic efficacy. Based on
stimulation protocols of two synaptically connected neurons, we infer
an algorithm which reproduces the experimental data by modifying the
probability of vesicle discharge as a function of the relative timing
of spikes in the pre- and post-synaptic neurons. The primary feature
of this algorithm is an asymmetry with respect to the direction of
synaptic modification depending on whether the presynaptic spikes
precede or follow the postsynaptic spike. In the case where neurons
fire irregularly with Poisson spike trains at constant mean firing
rates, the probability of discharge converges towards a characteristic
value which is determined by the pre- and post-synaptic firing
rates. On the other hand, if the mean rates of the Poisson spike
trains slowly change with time, our algorithm predicts modifications
in the probability of release which generalize Hebbian and BCM rules.
5:20PM Rajesh Rao
Predictive Learning of Direction Selectivity in Recurrent Cortical
Circuits
(with Margaret Livingstone and Terry Sejnowski)
Abstract:
When a spike is initiated near the soma of a cortical pyramidal
neuron, it may backpropagate up dendrites toward distal synapses,
where strong depolarization can trigger temporally asymmetric Hebbian
plasticity at recently activated synapses. We first show that these
mechanisms can implement a temporal-difference algorithm for sequence
learning. We then show using biophysical simulations that a
population of recurrently connected neurons with this form of synaptic
plasticity can learn to predict spatiotemporal input patterns. In
particular, we demonstrate that a network of cortical neurons,
starting from a weak bias, can develop direction selectivity similar
to that observed in complex cells in alert monkey visual cortex as a
consequence of learning to predict moving stimuli.
5:45PM
David Horn
Distributed Synchrony in a Hebbian Cell Assembly of Spiking
Neurons
Abstract:
We investigate the formation of a Hebbian cell assembly of spiking
neurons, using a temporal synaptic learning curve that is based on
recent experimental findings. It includes potentiation for short time
delays between pre- and post-synaptic neuronal spiking, and depression
for spiking events occurring in the reverse order. The coupling between
the
dynamics of synaptic learning and that of neuronal activation leads to
interesting results. One possible mode of activity is distributed
synchrony, implying spontaneous division of the Hebbian cell assembly
into groups, or subassemblies,
of cells that fire in a cyclic manner. The behavior of distributed
synchrony is investigated both by simulations
and by analytic calculations of the resulting synaptic distributions.
6:25PM Break, Discussion, and Conclusion