| dc.description.abstract | Neural activity from the brain can be recorded at different scales using a variety of electrodes, which vary in their resolution, cortical spread and invasiveness. The electroencephalogram (EEG) is recorded from the scalp, electrocorticogram (ECoG) from the cortical surface, while microelectrodes are inserted into the cortex which record local field potentials (LFP) and spiking activity in animals. These signals have been used to drive Brain Machine Interfaces with varying degrees of success, but an objective comparison of their efficacy has not been performed. A sensory system such as the visual cortex can be used as a model to compare the information available across these scales.
In this work, using a customized hybrid array containing both micro and ECoG electrodes, we recorded simultaneous signals from up to four scales (spikes, LFP, ECoG and EEG) from the visual cortex of two monkeys while they viewed a large array of natural images as well as parametric stimuli such as gratings. Complementary information theoretic and decoding approaches were used to quantify the information content about naturalistic and parametric stimuli at each of the scales. We found that the information content in ECoG exceeded all other measures, including spiking activity. Further, the maximum information content was found in the gamma (30-80 Hz) frequency range of the signals.
Several theories have been proposed to explain a potential role of gamma oscillations in the coding and visual information and its communication across brain areas. We instead tested whether gamma oscillations elicited by natural images could be explained simply based on the local image properties. To do this, first the gamma response for multiple visual features (such as orientation, spatial frequency, size, contrast, hue, saturation etc.) needs to be determined. Though the dependence of gamma on such features has been well studied when presented alone, how these features jointly affect gamma has not been investigated in detail. We found that gamma responses to a pair of features were largely separable in both LFP and ECoG. Based on this, we developed a multiplicative model in which the response to multiple features is simply a scaled product of individual features, and used it to predict the gamma responses to parametric gratings and chromatic patches.
Finally, we built an image computable model to predict gamma responses to complex natural images by extracting simple features from them and incorporating the previously learnt dependencies of gamma response. Our model was able to estimate the gamma responses to both chromatic and grayscale images.
Overall, the comparative study of information across scales can help in designing more accurate and reliable BMIs, while the predictability of responses can be used to increase the precision of BMIs. The prediction of gamma responses based on low level features also offers a simple “null” model based on local image properties, against which more advanced theories of gamma based on predictive coding or selective communication can be tested. | en_US |
