RIKEN Brain Science Institute
Brain mechanisms of object recognition
Visual object recognition, namely, recognition of objects by their visual images is a key function of the primate brain. This recognition is not a template matching between the input image and stored images, but flexible process which tolerates considerable changes in images due to different illumination, viewing angle, and articulation of the object. In addition, our visual system can deal with images of novel objects, based on previous experience of similar objects. Generalization is an intrinsic property of the primate visual system. The neural mechanisms of visual object recognition have been studied by behavioral studies of brain-damaged patients and lesioned monkeys, and anatomical connection studies and unit-recording studies in monkeys. Recent development of non-invasive measurement techniques from human brain has provided new and powerful tools to this field. In this lecture, I will overview the frontiers first in monkey studies and then human studies.
The visual images sensed at the retina are first conveyed to the primary visual cortex (V1), a cortical area located at the occipital pole (the back end of the brain). The elemental features of the images, such as the orientation and location of contour segments and color of stimulus patches, are extracted in the neuronal circuits up to V1 and represented by the activity of individual neurons in V1. The signals of visual images are then sent to two serial pathways. One pathway is directed toward the inferior temporal lobe, and the other toward the parietal lobe. The lesion studies on monkeys have shown that the occipitotemporal pathway is essential for the visual object recognition, while the occipitoparietal pathway for visuospatial functions including the perception of object location and visual control of movements. Single-neuron recordings have shown that the elemental features of visual images represented by neurons in V1 are integrated gradually along the pathways, and the way of integration is different between the occipitotemporal and occipitoparietal pathways.
Activity of single neurons represent more and more complex object features as the signals go forward along the occipitotemporal pathway. In the inferotemporal cortex, which is the final purely visual stage of the temporal pathway, there are neurons which selectively respond to faces. However, in the case of other objects, inferotemporal neurons respond only to moderately complex features, which are commonly contained in images of multiple, but a subset of objects. Neurons responding to similar features cluster in a small region named "column," which is about 0.5 mm in width. The stimulus selectivity of neurons in a column is similar but not identical. They respond to overlapping, but only partially overlapping, features. It is as if neurons in a column represent a category of features. The inferotemporal cortex of the monkey has about 1000 columns, which represent different categories of features. We also have found that similar, or related, categories of features are represented by neighboring columns in the inferotemporal cortex.
Recent studies on human subjects with non-invasive imaging techniques (PET, fMRI, and MEG) have shown that there are also functional division between the occipitotemporal and occipitoparietal pathways in the human brain. Moreover, there are several lines of evidence that object images are processed serially along the occipitotemporal pathway. However, the spatial resolution of the imaging techniques has been limited to 2-3 mm, and we could hardly go beyond the functional localization in human studies. We have recently developed the technique of fMRI with higher magnetic field (4 Tesla) and succeeded in imaging one columnar system (named the ocular dominance column) in the human V1. Imaging of the temporal cortex is more difficult, but if the column-level fMRI becomes applicable to the human temporal cortex, it will provide a great breakthrough in the study of object recognition mechanisms.