An algorithm may clear the choke point in evaluating gradients, paving the way to quantum neural nets.
If quantum AI is the Silk Road that will one day connect the power of quantum computing with solutions to some of the most important problems and biggest opportunities of the day, then evaluating gradients is the daunting Khyber Pass along the route.
Now, a researcher suggests that an algorithm he developed is a promising new approach to evaluating gradients, a key technical challenge for teams hoping to develop quantum AI.
In a paper available now on GitHub, Robert Tucci, a researcher who has a doctorate in theoretical physics, said that his algorithm represents a new and potentially fruitful approach to evaluating gradients of quantum cost functions, which he calls a major choke point in the creation of quantum AI. Gradient evaluation is one of the critical steps that neural networks use to learn.
Just as backpropagation is widely used for gradient evaluation in classical computing, Tucci said that this method could be implemented as a counterpart for backpropagation in quantum AI applications.
The solution may lie in multi-threading, also known as threading, a technique that is already used in classic computation, but is novel as a quantum computing strategy, said Tucci, a pioneering quantum programmer who invented quantum software tools including Qubiter and Quantum Fog.
“What I mean by threading is it is a strategy of partitioning the qubits in a gate model quantum computer into small, disjoint sets — or “islands” — that are uncorrelated from each other and run concurrently,” said Tucci. “The qubits within one of these islands are strongly correlated but qubits from different islands are probabilistically independent.”
After final measurements, each island of qubits, gives a mean value. The mean values of all the islands, then, lead to the gradient.
When the algorithm is in place, Tucci said it begins to look like digital rain, of Matrix fame.
Tucci added that this method is ideal for burgeoning quantum computer systems, such as Noisy Intermediate Scale Quantum and Hybrid Quantum-Classical computers, which are quantum computer designs that are still susceptible to the noise that can affect their performance.
Rigetti, a California-based quantum computer company, uses a hybrid quantum-classical design, for example.
The power of quantum computing lies in the massive potential of qubits to be in multiple states at the same time, unlike classical bits that can either be in a 1 state or an 0 state. As more qubits are linked together — entangled — their ability to process data increases exponentially. However, quantum computers are vulnerable to interference from the environment, or noise.
Scientists and entrepreneurs are already investigating the use of quantum computing machine learning techniques in applications, such as drug discovery and materials research. A quantum AI system, for example, could analyze millions of possible drug compound combinations to find the best formulation for a disease or condition.
Tucci said his algorithm could be a breakthrough. To check it out, this digital rain approach to calculating the gradients of quantum cost functions is currently a part of Tucci’s Qubiter, a full suite of tools for designing and simulating quantum circuits on classical computers. It’s available on GitHub.
“I believe that Qubiter, with its new addition for calculating gradients of quantum cost functions, will become a seminal work in the field,” Tucci writes on his blog, QB-Nets.
Tucci received a bachelor’s degree in mathematics and one in physics from MIT. He received his doctoral degree in theoretical physics from the University of Wisconsin. As an independent researcher, he has produced more than 50 papers.