In RC and ELM, one only needs to adjust the weights of the output connections via, for example, linear regression (see Figure 1), while the rest of the connections can be initialized with random weights and are not optimized. Figure 1 illustrates the three layers typical of RC and ELM, namely an input layer, a hidden layer or substrate, and an output layer. The main advantage of the RC and ELM concepts in the context of artificial NN is their minimal requirements for learning (usually referred as training in the machine learning literature). RC and ELM are machine learning paradigms that exploit the natural dynamics of input-driven randomly connected NN for information processing. In this review article, we concentrate on the potential of quantum devices for reservoir computing (RC) and extreme learning machines (ELM).
The implications of combining machine learning and quantum physics indeed represents a major avenue for research in the coming years. Thanks to recent advances, artificial NN are envisioned to be run even on top of analog quantum computing devices, with the possibility to exploit the advantages of superposition in quantum computing and the parallelism in neural computing. As computing approaches get closer to considerations on their physical substrates, the analog properties of physical systems come into focus.
The potential to build systems that are orders of magnitude more energy efficient than traditional ones is a major key motivation. Machine learning, and in particular the field of artificial neural networks (NN), can similarly benefit from the progress in neuro-inspired computing devices. Neuromorphic, and more generally neuro-inspired, computing is one of such fields where the computational paradigm goes hand in hand with the design of the physical substrate, aiming at approaching the computational power of the human brain. įor the unconventional computing revolution to occur, computational models and computing substrates are to be considered as a whole. One of the driving motivations for these efforts on unconventional computing is to go beyond von Neumann architectures, physically co-locating processing and memory operations. In recent years, we are witnessing an explosion of unconventional computing methods and systems.