|Title||Using ‘read-out’ of object identity to understand object coding in the macaque anterior inferior temporal cortex|
|Publication Type||Conference Proceedings|
|Year of Publication||2005|
|Authors||Hung, CP, Kreiman, G, Quiroga, RQuian, Kraskov, A, Poggio, T, DiCarlo, JJ|
|Conference Name||Computation and Systems Neuroscience (COSYNE)|
|Conference Location||Salt Lake City, Utah, USA|
Recent efforts to develop robust computer vision systems capable of approaching the level of human object recognition performance have shown that models based on neurobiology can outperform non-biological models. To continue the development of such systems, we must understand the codes used by neuronal ensembles to represent object identity--a matter of continued debate in the neuroscience community. Understanding such codes would also allow the development of brain-machine interfaces capable of ‘reading-out’ or ‘writing-in’ information to AIT. One approach to understanding these neuronal codes is to obtain a large sample of neuronal data using a fixed set of objects, and then test the ability of various neuronal population measures (codes) to convey knowledge of object identity – that is, determine the ability of each putative neuronal code to ‘read-out’ object identity information from AIT. To this end, we have now recorded single-unit activity (SUA), multiunit activity (MUA) and local field potentials (LFP), from over 500 sites in the anterior inferior temporal cortex (AIT) of two macaque monkeys while they viewed a fixed set of 77 complex objects. In this paper, we focus on the temporal latency and resolution of AIT neuronal population codes that best convey knowledge of object identity as assessed by object classification. Specifically, the 77 objects were divided before the experiment into eight different classes: toys, foodstuffs, human faces, monkey faces, hands, vehicles, boxes and cats/dogs. To evaluate the performance of each putative neuronal population code, we used support vector machines (SVM) and jackknife cross-validation to avoid over-fitting. Under the assumption of no covariance across spatially separate sites, these data and methods allowed us to examine putative neuronal population codes that differed in their latency (0 to 100 ms from stimulus onset) and their temporal resolution (12.5 to 200 ms wide bins). Results were evaluated for consistency across different sets of recording sites. We found that MUA and SUA spike count codes could perform single trial object classification. Performance improved with the number of sites and reached ~90% when ~128 arbitrary AIT sites were included. For any given number of sites, increasing the temporal resolution beyond ~50 ms did not lead to improved performance and the optimal latency from stimulus onset was ~100 ms. Because the neuronal data were not collected simultaneously, we cannot rule out spatio-temporal codes with a time scale less than ~50 ms. However, these results support the hypothesis that the time scale of the object identity code in AIT is ~50 ms and that downstream read-out of object identity for behavior or further processing may simply integrate over this time scale.