Kato D, Baba M, Sasaki KS, Ohzawa I.
Effects of generalized pooling on binocular disparity selectivity of neurons in the early visual cortex.
Phil. Trans. R. Soc. B 371: 20150266. (2016)
DOI: 10.1098\/rstb.2015.0266<\/p>\n\n\n\n
There is a recent surge of interest in deep-learning neural nets. A typical configuration of such neural nets is a hierachical repetition of layers consisting of filtering and pooling. We have studied theoretically and experimentally, the effects of pooling on neural representation of stereoscopic depth information.<\/p>\n\n\n\n
The key problem of stereoscopic vision is traditionally defined as accurately finding the positional shifts of corresponding object features between left and right images. Here, we demonstrate that the problem must be considered in a four-dimensional parameter space; with respect not only to shifts in space (X, Y), but also spatial frequency (SF) and orientation (OR). The proposed model sums outputs of binocular energy units linearly over the multi-dimensional V1 parameter space (X, Y, SF, OR) as illustrated in the figure.<\/p>\n\n\n\n
Monocularly, pooling generally causes loss of information about visual detail, since it is similar to defocus in a camera, in exchange for reduction of noise and gain in robustness.<\/p>\n\n\n\n
Binocularly, however, the same pooling produces completely oppsite effects. Pooling sharpens tuning for left-right stimulus matches in all of the V1 parameter domains.<\/p>\n\n\n\n
Our research demonstrates theoretically and experimentally that the pooling has the capacity to generate a more precise representation of stereoscopic information. Our finding has a potential to improve accuracy and robustness of stereovision applicatons.<\/p>\n\n\n\n
Figure legend:
(a) Pooling in the space domain (X, Y ) is illustrated schematically. (b) Pooling in V1 parameter space. Intuitively, the degree of pooling is determined by the number of subunits (small spheres) included within the pooling sphere (large sphere).<\/p>\n\n\n\n