2008 Special IssueImprecise correlated activity in self-organizing maps of spiking neurons
Introduction
The pioneering work of Hubel and Wiesel (1962) on cortical activity in the primary visual cortex of cats showed that cortical neurons were more effectively activated when a certain receptive field on the retina was stimulated. They also demonstrated that the position of the receptive fields, the orientation selectivity and the ocular dominance present a global columnar organization in the cortex.
However, there is still some debate on how cortical orientation selectivity is set up (Ferster & Miller, 2000). While diverse physiological (Ferster et al., 1996, Reid and Alonso, 1995) and theoretical (Linsker, 1986, von der Malsburg, 1973) studies support the role of feed-forward connectivity for orientation selectivity (see Miller, Erwin, and Kayser (1999) and Swindale (1996) for revisions), other reputable works (Ben-Yishai et al., 1995, Martin, 2002, Somers, 1995) point to a greater influence of intracortical circuitry and inhibitory connections.
Another controversial issue is how the neurons communicate with each other to form the proper circuitry that will robustly support different features of the central nervous system. While many studies address the possibility of having a sparse neural code based on neural assemblies (Gerstein & Kirkland, 2001), synchronous fire chains (Gewaltig, Diesmann, & Aertsen, 2001), or oscillations (Ritz & Sejnowski, 1997), others claim the existence of neural codes that require a precise inter-neural communication to be arranged (Reich et al., 1997, Reinagel and Reid, 2000).
Neuronal response properties undergo several important transformations from the retina to the primary visual cortex. Receptive fields become more elaborated (Hubel & Wiesel, 1962), average firing rates are reduced and visual responses become more variable (Kara, Reinagel, & Reid, 2000). Studies of cross-correlation analysis suggest an additional transformation in which the correlated firing between presynaptic and postsynaptic neurons becomes less precise. While retinogeniculate connections generate very narrow correlograms with less than 1 ms width at half-amplitude (Usrey, Reppas, & Reid, 1999), the connections from the lateral geniculate nucleus (LGN) to cortical layer 4 and from layer 4 to layers 2 and 3 generate much wider correlograms (Alonso, Usrey, & Reid, 2001).
One of the main factors explaining this reduction of precision of the inter-neural correlated activity is synaptic jitter (Veredas, Vico, & Alonso, 2005). In the pathway between the retina and the visual cortex of mammals, neural synaptic jitter (i.e. the variability of synaptic transmission latencies) increases. For example, in the developing rat neocortex, Markram, Lubke, Frotscher, Roth, and Sakmann (1997) measured fluctuations in excitatory post-synaptic potential latency of 1.5 ms between layer 5 neurons, which is a very large value in comparison with the submillisecond produced by retinogeniculate connections (Cleland, Dubin, & Levick, 1971). Increasing transmission jitter significantly reduces the precision of the correlated firing (Veredas et al., 2005). The question that arises now is how this reduction of the time accuracy of spike transmissions could affect the organization of feature maps.
In this article we analyze the self-organization of receptive fields and orientation selectivity maps in feed-forward networks of integrate-and-fire (IAF) neurons with modifiable connections. Our results show how changes in the precision of the correlated firing among neurons do not prevent the emergence of orientation selectivity maps but affect the features of receptive fields. This means that reducing the exactness of inter-neural communication significantly influences the configuration of receptive fields in this sort of feed-forward network. We modified this accuracy of the correlated activity by introducing incremental and physiologically plausible changes in the synaptic transmission jitter, i.e. increasing the variability of synaptic transmission delays.
We address the problem using a spike-time dependent plasticity (STDP) (Panchev and Wermter, 2004, Rao and Sejnowski, 2001) approach, where modifiable excitatory synapses change during a learning process to converge into a self-organizing structure that presents orientation sensitive receptive fields.
Section snippets
The spiking Neuron model
IAF models are a particular case of simplified neuronal models. The traditional form of an IAF neuron (Stein, 1967) consists of a first order differential equation (Eq. (1)) with a subthreshold integration domain (where the neuron integrates its inputs ) and a threshold potential (not explicitly represented in the equations) for action potential generation. where is the neuronal membrane capacitance, the membrane potential, the membrane resistance,
Precision of correlated activity and synaptic jitter
As revealed by physiological and theoretical studies (see Veredas et al. (2005)), increasing the synaptic transmission jitter–defined as the variability of the synaptic delay–affects the precision of the correlated firing in a monosynaptic connection. When this correlated activity is measured by cross-correlation analysis of the activity of two monosynaptically connected neurons, the increase of transmission jitter is evidenced by significant changes in the monosynaptic peak of the correlogram.
The STDP learning rule
Modifiable synapses are modeled by an STDP rule (see Eqs. (4), (5)). STDP rules model synaptic plasticity by considering in their formulation the precise timing of neuronal activity, so that synaptic efficacies are enhanced or reduced in a way that depends on the time distance of each post- and presynaptic spike. For the network architecture presented in this paper, only excitatory synapses are considered, which are governed by Eqs. (4), (5) (based on the learning rule for excitatory synapses
Simulations and results
The network architecture for simulations consists of three two-dimensional layers of 32×32 IAF neurons with feed-forward modifiable excitatory connections from input to output. There are no lateral connections. Receptive fields of neurons in intermediate and output layers consist of an 11×11 array of afferent synapses from the presynaptic layer, centered at the position of the postsynaptic neuron, assuming that layers are circular matrices for neighboring considerations. Synaptic weights are
Conclusions
We analyzed the effects of increasing the jitter of transmission delay on the self-organization of receptive fields and orientation selectivity maps. For this purpose, we simulated feed-forward networks of IAF excitatory neurons with modifiable synapses–by an STDP learning rule–and determined how a reduction in the precision of the correlated activity affects the final arrangement of neural receptive fields. Our main results show how when the precision of the interneural correlated activity is
Acknowledgments
This research has been technically supported by the Grupo de Estudios en Biomimética of the Universidad de Málaga, Spain, and partially funded by the Consejería de Salud, project PI0197/2007, and the Consejería de Innovación, Ciencia y Empresa, PAI, project P06-TIC01615, Junta de Andalucía, Spain.
References (39)
- et al.
Neural assemblies: Technical issues, analysis, and modeling
Neural Networks
(2001) - et al.
Propagation of cortical synfire activity: Survival probability in single trials and stability in the mean
Neural Networks
(2001) - et al.
Low response variability in simultaneously recorded retinal, thalamic, and cortical neurons
Neuron
(2000) Microcircuits in visual cortex
Current Opinion in Neurobiology
(2002)- et al.
Spike-timing-dependent synaptic plasticity: From single spikes to spike trains
Neurocomputing
(2004) - et al.
Synchronous oscillatory activity in sensory systems: New vistas on mechanisms
Current Opinion in Neurobiology
(1997) - et al.
Rules of connectivity between geniculate cells and simple cells in cat primary visual cortex
Journal of Neuroscience
(2001) - et al.
Combined Hebbian development of geniculocortical and lateral connectivity in a model of primary visual cortex
Biological Cybernetics
(2001) - et al.
Theory of orientation tuning in visual cortex
Proceedings of National Academy of Sciences, USA
(1995) - et al.
Postsynaptic potentials recorded in neurons of the cat’s lateral geniculate nucleus following electrical stimulation of the optic chiasm
Journal of Neurophysiology
(1988)
Anatomy of the cortex
Sustained and transient neurones in the cat’s retina and lateral geniculate nucleus
Journal of Physiology
Quantitative studies of intracellularpostsynaptic potentials in the lateral geniculate nucleus of the cat with respect to optic tract stimulus response latencies
Experimental Brain Research
Orientation selectivity of thalamic input to simple cells of cat visual cortex
Nature
Neural Mechanisms of Orientation Selectivity in the Visual Cortex
Annual Reviews of Neuroscience
Digital image processing
A quantitative description of membrane current and its application to conduction and excitation in nerve
Journal of Physiology
Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex
Journal of Physiology (London)
Spontaneous symmetry-breaking energy functions and the emergence of orientation selective cortical cells
Biological Cybernetics
Cited by (2)
A spiking self-organizing map combining sTDP, oscillations, and continuous learning
2014, IEEE Transactions on Neural Networks and Learning SystemsIntroduction to spiking neural networks: Information processing, learning and applications
2011, Acta Neurobiologiae Experimentalis