7. Concluding remarks

Neural networks, and in particular deep neural networks, have taken pride of place in artificial intelligence. The fields of application are numerous: image recognition, machine translation, time series, etc.

It's a rapidly evolving field, more so than conventional computers based on the Von Neumann model. However, the two models are different, complementary rather than competing. Neural networks are specialized operators. Although different software environments exist to "program" them, and numerous specific hardware supports exist, they retain an empirical aspect: for a given application, no heuristic can determine in advance the type of network to use, the number of layers, the neuron activation function - in other words, all the parameters that will determine the precision obtained, the learning time and power consumption. Techniques addressing this problem...