According to the Foxconn Technology Group news, Foxconn Research Institute of Quantum Computing Director Mingxiu Xie and the University of Sydney team, Australia, jointly proposed “Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational The research report “Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational Quantum Circuits” proposes a solution to the long-standing phenomenon of Barren Plateaus in the field, which can be improved by appropriate initial values of given variable parameters.
The research has been accepted by NeurIPS 2022, the world’s leading conference on machine learning and computational neuroscience, and was selected from more than 10,000 submissions worldwide for publication in late November.
The barren plateau refers to the fact that when the number of bits of a quantum computer is large, the current framework of quantum neural networks can easily become ineffective for training, and its objective function can become flat, leading to too long training or training failure.
In response to the research results, Director Mingxiu Xie said that the proposed solution to the barren plateau phenomenon allows quantum learning machines to show real advantages over traditional machines. In conjunction with this solution, the Institute of Quantum Computing presented the results of quantum simulation for battery development at this year’s Foxconn Technology Day (HHTD22), greatly reducing the quantum resources required.
In general, in the process of quantum machine learning, we learn by controlling the adjustable parameters of the logic gates to get the desired quantum circuit model, but in the process of learning, it is often difficult to update the parameters because of the excessive number of logic gates and the deep structure.
The problem that has been plaguing the field of quantum machine learning for a long time has been solved by improving the barren plateau phenomenon with the appropriate initial values of the tunable parameters, and a breakthrough has been achieved in this field, said Ming-Hsiu Hsieh, director of the Institute.