Real-Time Adaptive Information-Theoretic Optimization of Neurophysiology Experiments

Abstract

Adaptively optimizing experiments can significantly reduce the number of trials needed to characterize neural responses using parametric statistical models. However, the potential for these methods has been limited to date by severe computational challenges: choosing the stimulus which will provide the most information about the (typically high-dimensional) model parameters requires evaluating a high-dimensional integration and optimization in near-real time. Here we present a fast algorithm for choosing the optimal (most informative) stimulus based on a Fisher approximation of the Shannon information and specialized numerical linear algebra techniques. This algorithm requires only low-rank matrix manipulations and a one-dimensional linesearch to choose the stimulus and is therefore efficient even for high-dimensional stimulus and parameter spaces; for example, we require just 15 milliseconds on a desktop computer to optimize a 100-dimensional stimulus. Our algorithm therefore makes real-time adaptive experimental design feasible. Simulation results show that model parameters can be estimated much more efficiently using these adaptive techniques than by using random (nonadaptive) stimuli. Finally, we generalize the algorithm to efficiently handle both fast adaptation due to spike-history effects and slow, non-systematic drifts in the model parameters. Maximizing the efficiency of data collection is important in any experimental setting. In neurophysiology experiments, minimizing the number of trials needed to characterize a neural system is essential for maintaining the viability of a preparation and ensuring robust results. As a result, various approaches have been developed to optimize neurophysiology experiments online in order to choose the "best" stimuli given prior knowledge of the system and the observed history of the cell's responses. The "best" stimulus can be defined a number of different ways depending on the experimental objectives. One reasonable choice, if we are interested in finding a neuron's "preferred stimulus," is the stimulus which maximizes the firing rate of the neuron [1, 2, 3, 4]. Alternatively, when investigating the coding properties of sensory cells it makes sense to define the optimal stimulus in terms of the mutual information between the stimulus and response [5]. Here we take a system identification approach: we define the optimal stimulus as the one which tells us the most about how a neural system responds to its inputs [6, 7]. We consider neural systems in