A NIPS*2000 Workshop

---

Summary

• An anatomical perspective:  Does a common denominator, a repeating microcircuit element, exist across the neocortex?  What evidence is there for and against the existence of such an anatomical entity? Edward Jones, Jenny Lund, Javier DeFelipe
• Although there are many obvious instances of columnar structure and organization in cortex, the functional significance of these anatomical regularities remains unclear.
• Common elements of cortical anatomy include:
• focused thalamic input, spreading 500-600um
• vertical connections, e.g. mediated by spiny stellate cells and double bouquet cells
• horizontal connections, which tend to be smallest in layer IV and largest in layer III
• strong inhibition onto the soma and axons of pyramidal cells
• The 500-600um wide columnar structures defined by thalamocortical arbors encompass smaller mini- or micro-columns, each approximately 20-30um wide.  These mini-columns might be engaged differentially depending on the temporal structure of inputs to the column.
• There are anatomical differences in the composition of cortical circuits across mammalian species (e.g., double bouquet cells in monkeys but not rats).  These anatomical differences might reflect species-specific functional specializations in cortex; alternatively, the differences in anatomical form may be superficial, and cortical circuitry may be functionally equivalent across species.

• A cellular and synaptic perspective:  How do the cell types differ in the form of their inputs, the patterning of their outputs, and the transformations they perform?  How does the diversity of cell types and synapses in cortex contribute to computation within a column? Henry Markram, Alex Reyes, Edward Callaway, Harvey Swadlow
• Cortical cells encompass an incredible diversity of different anatomical shapes, electrophysiological signatures, molecular characteristics, and connectivity patterns.  In addition, cortical synapses are highly dynamic, providing additional diversity over time.
• This diversity might imply that the neocortical column should be viewed as an extremely complex, highly adaptable, multidimensional filter machine.
• Alternatively, the computational algorithms performed by neocortical columns might be relatively simple, despite the cellular and synaptic heterogeneity.  Diversity in implementation does not necessarily imply diversity in principles of operation.
• Molecular and genetic techniques (e.g., cell-type specific promoters, viral infection with transgenes, genetically encoded neural tracers) provide powerful new methods for teasing apart the roles of various cell classes in cortical computation.
• In-vivo electrophysiology can also be used to identify functional roles of cortical cell types.  For example, fast-spiking interneurons seem to form a synchronous inhibitory network, shaping a brief window of cortical excitability in response to afferent input.

• A network perspective:  Six layers, extensively recurrently interconnected.  What are the computational advantages of this layered neocortical structure?  To what extent do the layers interact, and is this interaction crucial to the operation of the column?  Does the column work as a unit or can layers operate independently?  What is the computational advantage of the immense recurrent circuitry? Rodney Douglas, Klaus Pawelzik, Terry Sejnowski, Rajeev Raizada
• Recurrent connections amplify cortical responses, and must be contained by inhibition to prevent instability.  The delicate balance of excitation and inhibition in cortex may cause the cortical network to act like a digital state machine with analog gain control, switching from stable point to stable point in the space of synaptic weights.
• Coding strategies in the cortex may depend on the dynamics of the input.
• Spike-timing dependent plasticity might facilitate self-organization among the diverse cell types and synapses in neocortex, ordering recurrent connections to increase response selectivity.
• Laminar structure in cortex may help to subserve attention and grouping.  In particular, projections from layer VI to layer IV may enhance representations of attended stimuli and suppress representations of ignored stimuli, while horizontal connections in layers II/III may integrate information for perceptual grouping.

• A comparative perspective:  Are there common themes for columnar function in cortex?  How do different cortical areas (e.g., visual, auditory, and somatosensory cortex) in different species compare in terms of columnar organization and distribution of response properties across cortical layers?  John Kaas, Terry Sejnowski, Edward Jones, Rodney Douglas
• Cortical columns and other modular structures in cortex represent discontinuities in the receptor sheet or multiplexed representations of decorrelated inputs.  While the factors responsible for creation of cortical modules may be genetically specified, the modules themselves may be an emergent property of cortex, specified only by the statistical structure of cortical inputs.
• Alternatively, columnar structure could be an intrinsic feature of cortical networks, and might arise even in the absence of a peripheral boundary or a computation requiring decorrelation of cortical inputs.

• A physiological perspective - visual cortex:  What information reaches a column in primary visual cortex from the thalamus, and what information is sent to the column from other areas?  How is this information integrated and transformed within the column?  Judith Hirsch, Yves Fregnac, Ralph Freeman, Adam Sillito
• Layer IV simple cells, representing the first stage of thalamocortical processing in visual cortex, may function primarily as integrators of thalamic input.  Neurons in the superficial layers, representing a later stage of thalamocortical processing, are much more selective and strongly gated than simple cells, and may function primarily as extractors of specific visual features.
• Contrast-independent orientation tuning in simple cells may arise because input from inhibitory interneurons offsets the contrast-dependent portion of thalamic input.
• Simple cells receive inhibition and excitation over the entire receptive field, but excitation develops faster than inhibition in the on-field, while inhibition develops faster in the off-field.  In general, receptive fields in visual cortex arise from time-dependent interactions between excitation and inhibition, and might best be viewed as dynamic structures or synaptic integration fields.
• Neighboring cells within visual cortex tend to be similar in some receptive field parameters (e.g., orientation tuning, spatial frequency tuning, etc.), but different in other receptive field parameters (e.g., spatial phase sensitivity).  Computation within visual cortical columns is likely to occur primarily along the dimensions of the receptive field which are represented differently by cells within the column.
• Receptive fields in primary visual cortex are shaped by feedback from higher areas.  The dynamics of computation in visual cortex may therefore be very dependent on the dynamics of cortical feedback.

• A theoretical perspective - visual cortex:  What components of the cortical column, and what rules of columnar organization, are sufficient to generate visual feature selectivity?  Kenneth Miller, Misha Tsodyks, Klaus Obermayer, Harel Shouval
• Steady-state properties of cortical cells in layer IV of cat visual cortex (e.g., contrast-independent orientation tuning) can be reproduced in a simple cortical circuit which might emerge through application of a simple Hebbian rule.  Important features of this basic visual cortex circuit include:
• strong thalamocortical input to excitatory and inhibitory cells
• opponency between inhibitory and excitatory cells with anti-correlated receptive fields
• dominant feedforward inhibition
• recurrent connectivity
• Neurons in layers II/III of visual cortex should be insensitive to stimulus features represented by opponent pairs of neurons in layer IV.
• Dynamic properties of visual cortex may be exploited to achieve efficient encoding of visual information.  Coding strategies in visual cortex may depend on the temporal statistics of the visual input.
• Correlations in spontaneous activity reveal functional connectivity in visual cortex.  This connectivity, which reflects orientiation maps, could arise from a simple cortical network with biologically realistic horizontal interactions.
• The very large number and diversity of visual cortex cells per degree of visual space suggests that the visual cortex may have the computational power to function as a general purpose visual learning machine.

• A physiological perspective - barrel cortex:  How do the different layers and columns in barrel cortex interact during tactile perception, and how are those interactions affected by learning?  Mathew Diamond, Martin Deschenes, Asaf Keller, Stanislaw Glazewski
• Whisker responses recorded in barrel cortex first emerge within the barrel of the stimulated whisker, then appear in neighboring barrels.  The functional connectivity revealed by this spread of activity is reflected in transfer of learning between whiskers (fastest for neighboring whiskers).
• Lateral spread of activity is also observed in the somatosensory thalamus, where cells extend projections across multiple neighboring barreloids.
• Thalamic afferents to barrel cortex have multi-whisker receptive fields.   Circuitry within the barrel cortex transforms this multi-whisker input into highly selective, single-whisker receptive fields.  Input from inhibitory interneurons may offset the multi-whisker portion of thalamic input to shape the single-whisker receptive fields.
• There is a continuing debate about whether connectivity between barrels involves horizontal connections in cortex, or only feedforward connections.

• A theoretical perspective - barrel cortex:  What does the example of barrel cortex reveal about how cortical networks extract information from their inputs?  What is a cortical network, how many neurons make a network, and how independent are different networks?  Daniel Simons, David Pinto, Kevin Fox, Ford Ebner
• Barrels transform thalamic multi-whisker receptive fields with weak between-whisker inhibition into single-whisker receptive fields with strong between-whisker inhibition.  The computational function of the barrel may be to enhance contrast between synchronous and asynchronous inputs.
• A simple model circuit can reproduce many of the features of barrel cortex responses.  Important features of this basic barrel cortex circuit include:
• strong thalamocortical input to excitatory and inhibitory cells
• reciprocal connectivity between excitatory and inhibitory cells
• dominant inhibition
• recurrent connectivity
• Parameter setttings in the model, optimized to match biological observations, suggest that barrels are cortical dampers rather than cortical amplifiers.  Temporally synchronous thalamic inputs briefly engage cortex before being damped down by inhibtion.  The ratio of inhibitory to excitatory thalamic input is critical to the operating range of the circuit, as changes in this ratio transform the damping circuit into an amplifying circuit.
• Damping circuits like barrel cortex are essentially high-pass filters, most sensitive to the timing of changes in the input.  In contrast, amplifying circuits would be most sensitive to input magnitude.  
• According to the model, each barrel acts as an independent processing unit (although there may be microdomains within a barrel forming multiple parallel networks).  A contrasting view posits transfer of information along horizontal connections between barrels.
• Plasticity in barrel cortex occurs fastest in the supragranular layers, and last in the middle layers, suggesting feedback-mediated dynamics in barrel computation.

Last revised:  02/21/01