Final Report of LDRD Project Number 34693: Building Conscious Machines Based Upon the Architecture of Visual Cortex in the Primate Brain

Our research plan is two-fold: first, we have extended our biological model of bottom-up visual attention with several recently characterized cortical interactions that are known to be responsible for human performance in certain visual tasks, and second, we have used an eyetracking system for collecting human eye movement data, from which we can calibrate the new additions to the model. We acquired an infrared video eyetracking system, which we are using to record observers' eye position with high temporal (120Hz) and spatial ({+-} 0.25 deg visual angle) accuracy. We collected eye movement scan paths from observers as they view computer-generated fractals, rural and urban outdoor scenes, and overhead satellite imagery. We found that, with very high statistical significance (10 to 12 z-scores), the saliency model accurately predicts locations that human observers will find interesting. We adopted our model of short-range interactions among overlapping spatial orientation channels to better predict bottom-up stimulus-driven attention in humans. This enhanced model is even more accurate in its predictions of human observers' eye movements. We are currently incorporating biologically plausible long-range interactions among orientation channels, which will aid in the detection of elongated contours such as rivers, roads, airstrips, and other man-made structures.