Neuromorphic Adaptable Ocular Dominance Maps
Time staggered winner-take-all (ts-WTA) is a novel analog CMOS neuron cell [8], that computes ‘sum of weighted inputs” implemented as floating gate pFET ‘synapses’. The cell behavior exhibits competitive learning (WTA) so as to refine its weights in respo
- PDF / 619,664 Bytes
- 8 Pages / 430 x 660 pts Page_size
- 74 Downloads / 218 Views
Abstract. Time staggered winner-take-all (ts-WTA) is a novel analog CMOS neuron cell [8], that computes ‘sum of weighted inputs” implemented as floating gate pFET ‘synapses’. The cell behavior exhibits competitive learning (WTA) so as to refine its weights in response to stimulation by input patterns staggered over time such that at the end of learning, the cell’s response favors one input pattern over others to exhibit feature selectivity. In this paper we study the applicability of this cell to form feature specific clusters and show how an array of these cells when connected through an RC-network, interacts diffusively so as to form clusters similar to those observed in cortical ocular dominance maps. Adaptive feature maps is a mechanism by which nature optimize its resources so as to have greater acuity for more abundant features. Neuromorphic feature maps can help design generic machines that can emulate this adaptive behavior. Keywords: Floating Gate pFET, competitive learning, WTA, Feature maps, ocular dominance.
1 Introduction Interconnectivity patterns between hierarchically organized cortical layers are known to extract different sensory features from a sensory image and map them over the cortical sheet. Higher cortical layers successively extract more complex features from less complex ones represented by lower layers. In fact it has been shown that different sensory cortices are also an outcome of a mechanism by which a generic cortical lobe adapts to the nature of stimulus it receives so as to extract sensory features embedded in it. [1]. Thus feature extraction and hence formations of feature maps are fundamental underlying principles of parallel and distributed organization of information in the cortex. Any effort towards an artificial or neuromorphic realization of cortical structure will have to comprehend these basic principles before any attempt is made to derive full benefit of cognitive algorithms that are active in the brain. Neuromorphic realization of cortical map finds useful application in the area of robotic vision where potential improvement are possible by employing mechanisms observed in a living brain [2]. Dedicated adaptive hardware can help design generic machines that conserve resources by acquiring greater sensory acuity to more abundant features at the cost of others A. Ghosh, R.K. De, and S.K. Pal (Eds.): PReMI 2007, LNCS 4815, pp. 49–56, 2007. © Springer-Verlag Berlin Heidelberg 2007
50
P. Gupta, M. Bansal, and C.M.Markan
depending on the environment they are nurtured. Another emerging area of interest is that of neural prostheses wherein implants such as retinal or cochlear artificially stimulate sensory nerves to overcome blindness or deafness. Such implants are effective only when cortical infrastructure (feature maps) to interpret inputs from these implants is intact [3]. Animals born with defunct sensory transducers find their representative cortical area encroached upon by competing active senses. In such animals or in those who have a damaged or diseased cortex, senso
Data Loading...
