Abstract

D. J. Berg, S. E. Boehnke, P. F. Baldi, D. P. Munoz, L. Itti, Modeling adaptation responses in the superior colliculus using a Bayesian theory of surprise, In: Proc. Society for Neuroscience Annual Meeting (SFN'07), Nov 2007. (Cited by 3)

Abstract: A fundamental question in visual neuroscience is the role of adaptation and habituation in sensory processing. Here we demonstrate that a simple mathematical model of Bayesian surprise can explain the adaptation responses of visual neurons in the Superior Colliculus. We previously proposed a Bayesian surprise model (Itti and Baldi, '06) to quantify information contained in a piece of data by measuring the effect this data has on an observer whether the observer be a single neuron or an organism. We call this new kind of information 'surprise'. This is fundamentally different from Shannon information (Shannon, '48) which only takes into account the probability of the data. Surprise transforms the observer's prior beliefs into posterior beliefs, according to Bayes theorem. Information can now be measured in a natural way by the distance (relative entropy) between the prior and posterior distributions of the observer over the available space of hypotheses (i.e., beliefs about the data). Surprise is important in situations where new data changes beliefs about sensory information. This mechanism is particularly important for phenomena such as adaptation and habituation. To test this computational theory we modeled recordings from neurons in superficial (SCs, n=18) and intermediate (SCi, n=36) layers of the Superior Colliculus of two awake behaving monkeys (Macaca mulatta). A simple paradigm was used where 7 flashes of light (55 ms in duration) were repeatedly presented in a cells receptive field. We introduced surprising changes by altering the intensity of the 4th flash on rare trials (30%). This oddball stimulus was of brighter intensity (10%), dimmer intensity (10%), or absent (10%). Additionally, inter-stimulus interval was varied from 75-800 ms to assess the time course of habituation or adaptation. Firing rates to the oddball stimuli in superficial neurons represented sensory adaptation, while responses in SCi neurons showed both adaptation and habituation (see Boehnke et al.). In the control condition, neural firing rate decreased significantly after the first presentation of the stimulus and then stabilized. This can be modeled with a single time constant representing the decay of beliefs about the information content of the stimulus. We find that surprise tracks adaptation effects seen on the peak magnitude of the response for SCs in all conditions. Thus we can quantify neuronal responses in the SCs as surprise units.

Themes: Computational Modeling, Bayesian Theory of Surprise, Monkey Electrophysiology