Untangling emergent cortical dynamics: neurons from networks, noise from chaos

The way in which cortical microcircuit components -- most importantly neurons -- and their connectivity -- the network -- shape and constrain emergent dynamics is a long-standing question in neuroscience. Experimentally observed dynamical properties can often be explained by circuit models with different simplifying assumptions for the underlying neuron models and their network structure, such as deterministic synapse models describing stochastic synapses, or a uniform network structure describing heterogeneous synaptic connectivity. Intrinsic neural variability, for example, can emerge from both stochastic synaptic properties (noise) and deterministic network dynamics (chaos). It is therefore often not clear if models with ad hoc simplifying assumptions for various biological details provide correct explanations for the emergence of cortical dynamics. In this thesis, we set out to advance our understanding of how detailed biological properties of cortical neurons and their network structure shape emergent dynamics by studying a model of a prototypical neocortical microcircuit that was reconstructed using all relevant available biological data. To make our predictions as biologically accurate as possible, we used a "zero tweak" strategy wherein parameters in the model were not adjusted to replicate specific experimentally observed dynamical properties, but instead were constrained by biological data. This allowed us to characterize the effect of two biological properties that are often abstracted away in ad hoc simplifications: stochastic synaptic transmission and a heterogeneous network structure with complex higher-order connectivity. Studying the model, we made several important predictions: (1) Stochastic synaptic transmission, in an interplay with recurrent network dynamics, causes rapid chaotic divergence of spontaneous activity. (2) Synaptic noise overshadows other local cellular noise sources. (3) Amid the noise and chaos, neurons can reliably respond to external inputs with millisecond spike-time precision. (4) This reliable response goes beyond mere feedforward suppression of recurrent dynamics and is driven by the circuit at a near-critical excitation-inhibition balance. (5) An abundance of high-dimensional cliques of all-to-all connected neurons which shape correlations between neurons in a hierarchical manner. (6) This effect is strongly reduced when synaptic connectivity is replaced by a rejected null model with reduced higher-order network structure. We conclude that a detailed representation of cellular noise sources and high-dimensional network structure is imperative to accurately model emergent cortical network dynamics. Models that make ad hoc simplifying assumptions need to carefully justify the exclusion of such details.

Sensory memory in a model of a cortical column

Julien Mayor

Sensory memory is the first step of a complex memorization process. Sensory information is buffered in a "sensory store" for a short period. Then part of it is transferred to working memory, where information can be actively processed and, eventually, through its rehearsal, can be memorized and stored in long-term memory. Sensory memory is characterized by a high capacity and a short temporal extension. In this work we show that such a transient buffering of information is well captured in models of cortical microcircuits. Sensory information stays for a short time in the perturbations it induced to the neural activity. Two fundamental characteristics of cortical columns appear to be useful to this purpose. The first one is an approximate balance of neuronal excitation and inhibition. The second one is the sparseness of cells recurrent connections. Together, these two discriminants ensure a broad set of complex dynamical behaviours of the neural network along with the presence of a phase of sustained non-oscillatory activity with very irregular spike trains in individual neurons. We show that at the vicinity of a phase transition from a quiescent state to a phase of chaotic spiking activity, global stimuli can be buffered in the "macroscopic" neuronal activity. However if the sensory input only excites a subpart of the network, efficient information buffering is made using the detailed spiking activity of every neuron in the network. The "microscopic" structure of the network is therefore needed in order to buffer sensory information. In most situations, addition of noise enhances the buffering performance and the computational power of the network. This effect is known in the physical literature as "stochastic resonance". If the network has to achieve complex transformation of its inputs, stochastic resonance is shown to be an emergent property of the population of neurons. Both the very irregular spiking activity present in vivo and the balanced excitation and inhibition measured in cortical columns have a new functional relevance in being the neuronal substrate of an efficient sensory memory .

EPFL2005

Input prediction and autonomous movement analysis in recurrent circuits of spiking neurons

Henry Markram

Temporal integration of information and prediction of future sensory inputs are assumed to be important computational tasks of generic cortical microcircuits. It has remained open how cortical microcircuits could possibly achieve this, especially since they consist--in contrast to most neural network models--of neurons and synapses with heterogeneous dynamic responses. It turns out, however, that the diversity of computational units increases the capability of microcircuit models for temporal integration. Furthermore the prediction of future input may be rather easy for such circuits since it suffices to train the readouts from such microcircuits. In this article we show that very simple readouts from a generic recurrently connected circuit of integrate-and-fire neurons with diverse dynamic synapses can be trained in an unsupervised manner to predict movements of different objects, that move within an unlimited number of combinations of speed, angle, and offset over a simulated sensory field. The autonomously trained microcircuit model is also able to compute the direction of motion, which is a computationally difficult problem ('aperture problem') since it requires disambiguation of local sensory readings through the context of other sensory readings at the current and preceding moments. Furthermore the same circuit can be trained simultaneously in a supervised manner to also report the shape and velocity of the moving object. Finally it is shown that the trained neural circuit supports novelty detection and the generation of 'imagined movements'. Altogether the results of this article suggest that it is not necessary to construct specific and biologically unrealistic neural circuit models for specific sensory processing tasks, since 'found' generic cortical microcircuit models in combination with very simple perceptron-like readouts can easily be trained to solve such computational tasks

2003

Sensory memory in a model of a cortical column

Julien Mayor

EPFL2005

Input prediction and autonomous movement analysis in recurrent circuits of spiking neurons

Henry Markram

2003