Conventional and unconventional computing systems are compared in a historical trace reflecting the author's forty years' experience in the former and ten years' experience in the latter. It will be demonstrated that much has changed and little has changed. Then specific and general instances of ergodicity as a potential new paradigm for learning systems, are advanced. In particular, the widely used Backpropagation neural network algorithm is shown to contain ergodic learning regimes. Recent experiments and issues in human versus machine learning are discussed. A general philosophy that learning should be ergodic until sufficient context is established, is proposed.
展开▼
