The Case for Quantum Machine Learning

Quantum machine learning is the use of quantum computing for the computation of machine learning algorithms.

Classical machine learning algorithm

We’ve learned that machine learning algorithms contain three components: representation, evaluation, and optimization.

• When we look at the representation, current machine learning algorithms such as the GPT-3 network, published in 2020, come to mind. GPT-3 produces human-like text, but it has 175 billion parameters. In comparison, the IBM Q quantum computer has 27 quantum bits only. So even though quantum bits store a lot more information than a classical bit, since it is not 0 or 1, quantum computers are far from advancing machine learning for their representation ability.

• During the evaluation, the machine learning algorithm tries to predict the label of a thing. Classically, this involves measuring and transforming data points. For instance, neural networks rely on matrix multiplications. These are tasks classical computers are good at. However, if we have 175 billion parameters, calculating the resulting prediction takes many matrix multiplications.

• Finally, the algorithm needs to improve the parameters in a meaningful way. The problem is to find a set of parameter values that result in better performance. With 175 billion parameters, the number of combinations is endless.

Classical machine learning employs heuristics that exploit the structure of the problem to converge to an acceptable solution within a reasonable time. However, despite advanced heuristics, training the GPT-3 would require 355 years on a single GPU, and cost $4.6 million.