Internal Representations by Error Propagation , " In

21 at dierent places along the dendrites). [Hasselmo et al.] show that acetylcholine aects the synaptic connection strength of lateral but not feedforward connections in rat piriform cortex. Neuromodulators such as this could provide a means for neurons to give more or less weighting to the \teach" inputs. The neuromodulators or the teach inputs themselves could serve as a gate for learning ensuring that weight modication only occurs during times of appropriate teaching. It should also be noted that occasional misgating would not be disastrous. The teaching input could be considered as input from other sensory modalities (or sub-modalities as in colour helping shape) possibly through the limbic system. Some of these modalities would be receiving nearly constant input and thus could serve as a teacher for others receiving changing stimuli. There is evidence that proprioceptive input is necessary for the development of visually guided reaching [Hein, 1974] and depth perception [Graves et al., 1987]. Alternatively teaching input could come from temporally contiguous views. Miyashita in [Miyashita, 1988] has found neurons in macaque inferotemporal cortex that associated \visually distinct but temporally related" images. The article suggests that the strong limbic connections are important in this association. Kohonen mapping algorithms have been shown to account well for the retinotopic maps observed in striate cortex [Obermayer et al., 1990]. It is possible that the addition of the teach input with recurrent connections may explain how higher areas achieve position invariant responses and object-centered representations [Rolls, 1990]. This theory would also explain the purpose of the prevalent back-projections throughout cortex. [Eckhorn et al., 1988] found that lesioning V2 had little aect on the response properties of cells in V1 indicating that back-projections, or at least the V2-V1 projection has little purpose during regular processing. In this report we have provided some support for the hypothesis that these back-projections are important for the organization of the incoming sensory stimuli during learning. Acknowledgements We would like to thank Steven Nowlan and Leonidas Kontothanassis for their helpful comments and suggestions on earlier drafts. 20 4 Conclusions We have seen that the addition of a teach input with recurrent projections greatly increases the uses of the competitive learning algorithm and may make it suitable for such tasks as position-invariant recognition. Although the algorithm has only been demonstrated on small problems, it was not specialized in any way for the above tasks. In fact given enough hidden layer …