Time: 2pm
Place: 2N3
Structural Drivers of Function in Neural Network Systems
Numerous experimental and computational studies have identified relationships between architecture and performance in information processing systems ranging from artificial networks to natural neuronal ensembles. Across this wide range of systems, no single architecture has been identified as optimal; instead, inherent tradeoffs between robust and fragile features often limit performance during complex tasks. In this seminar, I will discuss recent progress made toward understanding the structural features that constrain functional activity in small-scale artificial neural networks and in large-scale human brain networks.
Events
Time: Noon
Place: IRCS Conference Room
Neural Basis of Probabilistic Inferences
Time: Noon
Place: IRCS Conference Room
Computational Neuroscience Seminar Series
Time: 12-1
Place: Barchi Library (140 John Morgan)
Cell types, circuits and repair
Time: 12:15 - 1:15
Place: Barchi Library (140 John Morgan)
In search of the holy grail of fly motion vision
Time: 4pm
PLace: DRL A6
Vertical silicon nanowire arrays as a chemical and electrical interface to living cells
Time: 3:30 PM
Place: DRL A6
Decoding the architecture of vascular networks
Time: 9AM
Place: IRCS Seminar Room, 3401 Walnut Street, 400A
Sparse high-order interaction networks underlie learnable neural population codes
Information is carried in the brain by the joint activity patterns of large groups of neurons. Understanding the structure and function of population neural codes is challenging, due to the exponential number of possible activity patterns and dependencies between neurons. Studying groups of 100 retinal neurons responding to natural movies, we found that they are strongly correlated and that pairwise maximum entropy models, which are highly accurate for small networks, are no longer sufficient. We show that because of the sparse nature of the neural code, the higher order interactions can be easily learned with surprisingly high accuracy using a novel pseudo-likelihod model, and that a very sparse interaction network underlies the code of large populations of neurons. Additionally, we show that the interaction network is organized in a hierarchical and modular manner, suggesting scalability of the code. Our results suggest that learnability is a key feature of the neural code.
Time: 3:30
Place: Stiteler Hall B21
Normalization as a canonical neural computation
Time: Noon
Place: IRCS Conference Room
Optimal integration of evidence for decision-making in the rat
Gradual accumulation of evidence over time is thought to be a fundamental component of decision-making, but the mechanisms and properties of the accumulation remain unclear. Although most models assume a noisy evidence accumulation process, the properties of this noise have never been isolated and measured. We developed a novel decision-making task that is particularly amenable to quantitative analyses that reveal properties of the decision process. The task allowed us to measure, for the first time, the magnitude of noise in the evidence accumulator process, separately from the magnitude of noise in sensory processes. Remarkably, we found that accumulator noise magnitude was zero. In addition, we found that the accumulator had very long (~1 sec) time constants. Our results show that rats have near-optimal graded evidence accumulators, characterized by long, noiseless memories.
