Online learning with random representations

摘要: We consider the requirements of online learning|learning which must be done incrementally and in realtime, with results learning available soon after each new example is acquired. Despite abundance methods for from examples, there are few that can used eectively learning, e.g., as components reinforcement systems. Most these few, including radial basis functions, CMACs, Kohonen’s self-organizing maps, those developed this paper, share same structure. All expand original input representation into a higher dimensional an unsupervised way, then map to nal answer using relatively simple supervised learner, such perceptron or LMS rule. Such structures learn very rapidly reliably, but have been thought either scale poorly require extensive domain knowledge. To contrary, some researchers (Rosenblatt, 1962; Gallant & Smith, 1987; Kanerva, 1988; Prager Fallside, 1988) argued expanded chosen largely at random good results. The main contribution paper develop test hypothesis. show random-representation perform well nearest-neighbor (while being more suited learning), signicantly better than backpropagation. nd size does increase dimensionality problem, not unreasonably so, required reduced substantially unsupervisedlearning techniques. Our suggest randomness has useful role play constructive induction. 1. Online Learning Applications divided two types: oine.