Abstract
The ventral stream of the primate's visual system, involved in object recognition, is mostly hierarchically organised. Along the hierarchy (from V1, to V2, V4, and IT) the complexity of the preferred stimulus of the neurons increases, while, at the same time, responses are more and more invariant to shift, scale, and finally viewpoint. Several feedforward networks have been proposed to model this hierarchy by alternating simple cells, which increase selectivity, with complex cells, which increase invariance (Fukushima 1980; Le Cun & Bengio 1998; Riesenhuber & Poggio 1999; Serre et al 2005). The issue of learning is perhaps the least well understood, and many authors use hard-wired connectivity and/or weight-sharing. Several algorithms have been proposed for complex cell learning based on a trace rule to exploit the temporal continuity of the world (for e.g., Foldiak 1991; Wallis & Rolls, 1997; Wiskott & Sejnowski, 2002; Einhaüser et al 2002; Spratling 2005), but very few can learn from natural cluttered image sequences.
Here we propose a new variant of the trace rule that only reinforces the synapses between the most active cells, and therefore can handle cluttered environments. The algorithm has so far been developed and tested through the level of V1-like simple and complex cells: we showed how Gabor-like simple cell selectivity could emerge from competitive hebbian learning, and how the modified trace rule allow the subsequent complex cells to pool over simple cells with the same preferred orientation, but with shifted receptive fields. Development of the V2, V4, and IT layers is ongoing.