An increasingly efficient narrowband object-recognition channel along the ventral stream.

How does the visual system recognize objects in natural environments? Here, we investigate the neural computations underlying robust object recognition along the ventral stream. We adapted psychophysical critical-band masking to fMRI, measuring BOLD responses along the ventral stream to natural images perturbed by bandpass noise. We found that, along this stream, the overall BOLD response to noise alone is sensitive to an increasingly wide range of stimulus spatial frequency, from 2 octaves in V1 to 5 octaves in ventral temporal cortex (VTC). However, when we assess the effect of the same noise on the accuracy of decoding scene images from the BOLD response, we obtain a different result: Recognition bandwidth is conserved along the ventral stream at about 2 octaves, close to the 1.5-octave behavioral channel. Though the recognition band is conserved, its noise to lerance increases steadily along the ventral stream, approaching behavioral level sin VTC. These findings suggest that V1 sets the bandwidth of the object-recognition channel, while downstream areas progressively denoise the signal, establishing the channel's noise tolerance. This biological architecture-early channel isolation followed by successive denoising-may explain why human vision remains more robust than current machine vision.