The ability of humans to identify and categorize objects in complex natural scenes has long been thought to be beyond the capacities of artificial vision systems. However, recent progress in Deep Learning and Convolutional Neural Networks has demonstrated that simple feed-forward processing architectures composed of less than 10 layers of neurons can achieve human levels of performance in object recognition tasks. It is interesting to note that such processing architectures have a very similar structure to the primate visual system. Could it be that we are close to understanding how our brains recognize stimuli? I will argue that the main problem with the current state of the art in computer vision is that the learning procedures used are totally unrealistic. Essentially, building such a system requires hundreds of millions of training cycles of supervised learning. By contrast, our own visual systems can learn new stimuli in a few tens of presentations. I will suggest that more biologically realistic learning mechanisms based on spike-based processing and Spike Time Dependent Plasticity (STDP) may be much closer to the way our own visual systems operate, and allow our visual systems to learn about objects in the visual world on the basis of experience.
Mot(s) clés libre(s) : neurosciences, information visuelle, extraction valeur des images