Networks for vision

How do neurons in cortical networks react to input from outside, and how does this behavior contribute to sensory processing in the brain? Which mechanisms on the level of cells and their networks are underlying vision in particular? Answers to these long-standing questions may come from a joint experimental and theoretical approach that dissects the neuronal mechanisms through which neuronal responses are “tuned” to certain features of a visual stimulus, for example the direction of movement or the orientation of an edge. Sadra Sadeh and Stefan Rotter from the Cluster of Excellence BrainLinks-BrainTools and the Bernstein Center at the University of Freiburg provide another piece of this puzzle through a new study, published in the journal PLOS Computational Biology.

Previous theoretical studies had already provided evidence that orientation selectivity – an important feature of basic visual processing – can emerge in randomly connected networks that could, regarding their architecture, also serve many other purposes. The new study now presents several additional insights. First, it shows that contrast-invariant orientation selectivity arises in many different regimes of brain dynamics, e.g. in fluctuation and mean driven as well as in synchronous and asynchronous regimes of network activity. Moreover, the detailed computational analysis of network responses allowed the authors to disentangle the exact mechanisms that underlie orientation processing in model networks. Of special interest are the different roles of linear mechanisms vs. nonlinear ones, like input rectification (i.e. the fact that neurons cannot respond to input below a certain threshold) or supralinear responses (i.e. cell responses that are larger than a linear summation of input). The authors show that linear mechanisms are in principle sufficient for strongly tuned and contrast-invariant neuronal output to emerge, whereas nonlinear mechanisms are responsible for the “fine-tuning”, which makes the neuronal responses even more specific. In order to understand how the circuits underlying orientation selectivity really work, this assignment of neuronal mechanisms to known features of visual processing is essential, and it was conspicuously lacking in previous theoretical studies.

In addition to these insights, the new approach provides a computational framework to study the relation between functional properties of single neurons and the topology of the network in which they are embedded. In this respect, the present analysis is going beyond purely statistical approaches that rely on the assumption of simple connectivity. Heterogeneity in network connectivity in combination with the non-linear mechanisms outlined above affect the firing rates of individual neurons; as a result, neuronal tuning curves can only be understood correctly if the precise flow of signals in the underlying network is taken into consideration, as it is done now in this computational study. Accounting for the precise micro-scale anatomy of neuronal circuits in this fashion has a great potential to explain also other aspects of brain function, going beyond the process of vision.

Original publication:

Sadeh S, Rotter S (2015) Orientation Selectivity in Inhibition-Dominated Networks of Spiking Neurons: Effect of Single Neuron Properties and Network Dynamics. PLoS Comput Biol 11(1): e1004045. doi:10.1371/journal.pcbi.1004045

Figure caption:

Knowing the network topology, individual output tuning curves can be computed in neuronal networks with different neuron models and network parameters: A network of perfect integrate-and-fire neurons in (A), the same network with more inhibition dominance in (B), and a network of leaky integrate-and-fire (LIF) neurons in (C). Output tuning curves of 24 sample neurons (excitatory neurons are drawn in red, inhibitory neurons in blue) are extracted from numerical simulations (dots) and compared to the corresponding analytical predictions (solid lines): In (A) a linear theory was used, in (B) the linear rectified model, and in (C) the full nonlinear input-output relation for LIF neurons.