Why Quantum Machine Learning?

We aim to make near-term quantum computers useful for hard tasks of practical relevance in machine learning and optimization.

Back

Our approach to QML is to focus scarce quantum resources on hard subtasks in machine learning algorithms

Quantum Machine Learning

Better models for prediction and reasoning

Machine learning, in essence, can be broken into three parts: data, models, and training algorithms. A training algorithm learns from data in order to perform and improve on a task. Typical tasks include prediction, classification, decision making, data generation, and anomaly detection.

Augmenting complex models or costly data generation with quantum computing resources could lead to greater accuracy in downstream tasks. The intuition is that quantum computers naturally excel at certain key ingredients that define machine learning — modelling complex probability distributions and generating hard-to-simulate data.

Many traditional machine learning models can be understood as probabilistic models of observable data and unobservable features or variables. For example, a trader may want to base their buy/sell decisions on the state of the equity market: is it bull or bear? While they cannot observe this abstract state directly, they can observe equity returns from traded assets. Based on their domain knowledge they build a probabilistic model that describes how the market state influences equity returns. Given this model and the return data, they can infer and reason about the unobserved market state using statistical methods and machine learning. Performing this inference exactly and quickly becomes intractable for complex models.

Near-term quantum computers are naturally suited for representing such complex probability distributions. Once trained, the quantum computer can be used in downstream tasks that rely on sampling complex data. The example of the trader performing inference is one such downstream task.

A pragmatic, near-term results-driven approach

Our approach to quantum machine learning is to focus scarce quantum resources on hard subtasks in machine learning algorithms. Consequently, our pioneering research centers around hybrid quantum-classical probabilistic methods for unsupervised learning, generative models, and inference.

A common approach to unsupervised learning is to learn a probabilistic model of data. These models can be trained on unlabeled datasets. Generative modeling is one example of unsupervised learning wherein the model automatically learns the patterns of input data in such a way that it can be used to generate new data samples that would seemingly match the statistical distribution of the original dataset.

Quantum circuit Born machines are an example of generative models and are based on parameterized quantum circuits. They can represent probabilistic models and sample complex probability distributions more efficiently than classical computers for certain tasks or datasets. We can train these Born machines for a diverse range of downstream tasks such as reasoning under uncertainty, data augmentation, and anomaly detection. Generative models can also make machine learning more transparent by incorporating domain expertise.

We implement quantum machine learning methods in a pragmatic and hardware-inclusive way. This allows us to implement small-scale models on today’s limited quantum hardware and improve the models as quantum hardware matures.

Learn more about OML at Cambridge Quantum