Без заголовка

Еще на счет хардваре. Вот Yoshua Bengio обновил работу про e-prop (equilibrium propagation). И пишет: We lay down the basis for a new framework for machine learning. In this framework the quantities of interest (the prediction and the objective function) are defined, in terms of the data and the parameters of the model, implicitly through an energy function, rather than explicitly (like in a feedforward net). This implicit definition involves the minimization of an energy function. This makes applications on digital hardware (such as GPUs) less practical as it needs long inference phases involving numerical optimization. But we expect that this framework could be extremely efficient if implemented by analog circuits, as suggested by Hertz et al. (1997). The framework presented in this section is deterministic, but a natural extension to the stochastic case is presented in the Appendix B. https://arxiv.org/abs/1602.05179 Вот и думаешь, что надо программировать на вязальных машинках, которые будут вязать аналоговые дендриты/нейроны Изображение — открыть источник F1.Schematic of biological NN and an ONW ST (organic nanowires synaptic transistor) that emulates a biological synapse http://advances.sciencemag.org/content/2/6/e1501326.full