Abstract
Neocortical pyramidal neurons receive a large fraction of their
excitatory inputs from neurons in other cortical columns. These inputs
innervate basal dendritic segments that are capable of initiating active
dendritic spikes. The function of these lateral connections remains
unclear. In this study, we propose that cross-column connections
implement an important neural mechanism that contributes to robust
pattern recognition. We consider the problem of object recognition with
a set of cortical columns, where each cortical column contains thousands
of neurons. An object consists of a set of component features. Each
feature corresponds to sensory input at a particular location on the
object, which are represented by sparse distributed representations
(SDR) in an upstream neural population. A cortical column learns an
object by forming feedforward connections from its component feature
SDRs to a sparse set of neurons. After learning, sensation of a sequence
of object features leads to activations of the corresponding neural
population. Since features can be shared among multiple objects,
information received by a single cortical column can be ambiguous. It
takes many sensations before the network converges to the correct
representation. We find that the performance can be dramatically
improved by simultaneously considering a set of cortical columns with
lateral connections, where each column learns feedforward connections
independently and learns cross-column lateral connections according to
Hebbian rules. The lateral inputs target onto distal dendritic segments
and act as contextual signals. Although they are not strong enough to
directly activate a neuron, neurons with both lateral inputs and
feedforward inputs will fire faster and prevent other neurons from
responding. We show that objects can be recognized faster with
multiple columns because the lateral connections help to disambiguate
the inputs, and that each cortical column can store more objects without
being confused with cross-cortical column connections.
Workshop
Subutai Ahmad, VP Research, will deliver a workshop on Feb. 28
titled “Why Do Neurons Have Thousands of Synapses? A Theory of Sequence Learning in Neocortex.” This workshop will be based on the content in our peer-reviewed paper by the same name.
Related links:
About Cosyne
The annual Cosyne meeting provides an inclusive forum for the exchange of
experimental and theoretical/computational approaches to problems in systems neuroscience.
To encourage interdisciplinary interactions, the main meeting is arranged in a single track. A set of invited talks are selected by the Executive Committee and Organizing Committee, and additional talks and posters are selected by the Program Committee, based on submitted abstracts.
Cosyne topics include (but are not limited to): neural coding, natural scene
statistics, dendritic computation, neural basis of persistent activity, nonlinear receptive field mapping, representations of time and sequence, reward systems, decision-making, synaptic plasticity, map formation and plasticity, population coding, attention, and computation with spiking networks. Participants include pure experimentalists, pure theorists, and everything in between.