Modeling spatio-temporal dynamics of network damage and network recovery.

How networks endure damage is a central issue in neural network research. In this paper, we study the slow and fast dynamics of network damage and compare the results for two simple but very different models of recurrent and feed forward neural network. What we find is that a slower degree of network damage leads to a better chance of recovery in both types of network architecture. This is in accord with many experimental findings on the damage inflicted by strokes and by slowly growing tumors. Here, based on simulation results, we explain the seemingly paradoxical observation that disability caused by lesions, affecting large portions of tissue, may be less severe than the disability caused by smaller lesions, depending on the speed of lesion growth.

Prediction of individual differences in risky behavior in young adults via variations in local brain structure.

In recent years the problem of how inter-individual differences play a role in risk-taking behavior has become a much debated issue. We investigated this problem based on the well-known balloon analog risk task (BART) in 48 healthy subjects in which participants inflate a virtual balloon opting for a higher score in the face of a riskier chance of the balloon explosion. In this study, based on a structural Voxel Based Morphometry (VBM) technique we demonstrate a significant positive correlation between BART score and size of the gray matter volume in the anterior insula in riskier subjects. Although the anterior insula is among the candidate brain areas that were involved in the risk taking behavior in fMRI studies, here based on our structural data it is the only area that was significantly related to structural variation among different subjects.

New attractor states for synchronous activity in synfire chains with excitatory and inhibitory coupling.

In a feedforward network of integrate-and-fire neurons, where the firing of each layer is synchronous (synfire chain), the final firing state of the network converges to two attractor states: either a full activation or complete fading of the tailing layers. In this article, we analyze various modes of pattern propagation in a synfire chain with random connection weights and delta-type postsynaptic currents. We predict analytically that when the input is fully synchronized and the network is noise free, varying the characteristics of the weights distribution would result in modes of behavior that are different from those described in the literature. These are convergence to fixed points, limit cycles, multiple periodic, and possibly chaotic dynamics. We checked our analytic results by computer simulation of the network, and showed that the above results can be generalized when the input is asynchronous and neurons are spontaneously active at low rates.