Quantum machine learning: Concepts and Examples

Machine learning has been a hot topic for many years now. There are different types of machine learning algorithms that data scientists and engineers use in their projects, depending on the type of problem they’re trying to solve. Recently, quantum machine learning has emerged as an alternative to classical machine learning techniques. The future of quantum computing holds tremendous possibilities promising exponential speedups over current technology. In this blog post, we’ll explore quantum machine learning (QML), its benefits over traditional machine learning methods, and the common quantum computing concepts it relies on.

What are key concepts related to quantum computing?

Quantum computing takes advantage of the computing power available through superposition and entanglement in addition to other phenomena related to quantum mechanics such as tunneling or interference effects. Quantum algorithms outperform their classical counterparts when it comes to searching large spaces for low-energy solutions (minimizing related functions).

The following are key concepts related to quantum computing:

• Qubits (Quantum bits) are information units that follow two basic rules: they can exist only in two states at any one time – either ON or OFF; once measured, qubit has no way of telling whether it was turned on or off until we measure again later. A qubit can be understood as a generalized concept of the classical bit. Qubits are the basic unit for quantum computation, which is able to process exponentially more information than classical computers can deal with in the same amount of time by exploiting superposition and entanglement effects.
• Quantum Superposition is the key concept related to quantum computing that allows for particles to be in more than one state at once which provides great power and flexibility for solving complex problems through parallel processing. This process can also help with optimization tasks like searching large spaces for low-energy solutions (minimizing related functions).
• Entanglement: A physical phenomenon that occurs when pairs or groups of particles are generated, interact, or share spatial proximity so that their quantum states become correlated (the measured property of each member of an entangled pair will always correlate with the measured property of its partner regardless if they are separated by microns/miles). Entangled photons have been used successfully to carry out quantum communication.
• Tunneling is the quantum mechanical effect where a particle passes through what seems like an energy barrier. Quantum interference effects are used to build highly stable qubits that can hold their states for very long periods of time (years or even longer).
• Quantum gate: It’s an operation acting on some number of input quantum states called basis states, which results in desired output state(s). Gates are represented by unitary matrices that act on a basis vector.

What is quantum machine learning (QML)?

Quantum machine learning (QML) is an emerging field that combines quantum computing with machine learning algorithms to give solutions for problems that are very difficult to solve in a conventional way. Quantum machine learning uses quantum computers for training machine learning models. Quantum computers can process data in a special way using quantum bits (qubits) instead of conventional bits used by classical digital computers. Quantum machine learning has emerged due to the current issues with big data processing. Machine learning models can be trained in a faster way using quantum computing. In addition, using Quantum algorithms, large spaces can be searched for approximating the functions.

Quantum machine learning can be defined as the technology wherein quantum algorithms are applied to machine learning problems. The goal is to find the best possible solution of a given quantum algorithm for a particular problem and apply it in real-world applications like image recognition, language translation, or any other task which can be solved using machine learning. Quantum algorithms could help transform artificial intelligence (AI)/machine learning (ML) use cases by accelerating big data analytics at incredible speeds.

The following are some examples of quantum algorithms for quantum machine learning:

• Quantum annealing is a quantum computing technique, which does quantum search and optimization. It is an optimization technique used to determine the local minima and maxima of a function over a given set of candidate functions. This is a method of discretizing a function with many local minima or maxima in order to determine the observables of the function.
• Quantum Boltzmann machine (QBM) is a type of quantum recurrent neural network that has been applied to model supervised and unsupervised learning tasks. Recall that Boltzmann machines are neural networks used for generative machine learning.  QBM can be considered as an example of the more general class of quantum generative models, which are probabilistic graphical models where coherence plays a central role in information processing. QBM can be understood to have quantum version of energy function, objective function and training data keeping in mind the Boltzmann machine. Recall that when training Boltzmann machine learning model, the algorithm requires starting with initial weights and bias and an energy function which is followed by optimizing an objective function which results in updation of weights and biases and generation of new energy function. Here is a good slide on Quantum Boltzmann machine
• Quantum reinforcement learning (QRL) is aimed to harness the computational advantages provided by quantum computers by designing RL agents that rely on quantum models of computation.Recall that reinforcement learning is machine learning technique for training AI agents with data from their environment and then allowing the agents to autonomously perform tasks rather than providing explicit instructions on what actions should be taken at each step of the trial-and-error process. In quantum reinforcement learning, quantum states are used to store the information about a subset of actions. Quantum reinforcement learning has been demonstrated in some simple cases which shows its scope for applications in AI and robotics. The reinforcement learning algorithms such as policy-gradient and a deep Q-learning can be trained using parametrized quantum circuits (PQCs) or variational quantum circuits (VQCs). PQCs has been widely accepted because they can be efficiently implemented on noisy intermediate-scale quantum (NISQ) processors. Recall that in policy-gradient implementation, an RL policy is trained with a policy-gradient method while in Deep Q-learning models, a Q-function approximator gets trained with deep Q-learning algorithm. Here is tensorflow page on quantum reinforcement learning.
• Quantum deep neural networks (QDNNs) are an extension to artificial feedforward NN. The QDNN models get trained with quantum gates as activation functions in the hidden layer. It can be used to solve classification problems by training it on labeled data sets. QDNN has been proposed to overcome shortcomings of classical ANN such as vanishing gradients and slow convergence rate due to high dimensionality and overfitting/underfitting problem respectively. It has been shown that QDNN models can uniformly approximate any continuous function and are found to have greater representation power than the classical DNN. QDNN does provide all advantages of classical DNN including multi-layer neural network structure, non-linear activation functions, and the backpropagation training algorithm. As like QBM, QDNN makes use of parameterized quantum circuits (PQCs) as the basic building blocks. Here is a great read on quantum deep neural networks
• Quantum walks: Quantum walks are a universal model for quantum computation. These are quantum-enhanced search algorithms that can be used for optimization problems. They offer an alternative approach for implementing a variety of quantum algorithms, including data-base search, graph isomorphism, network analysis and navigation and quantum simulation. Quantum walks are quantum analogs to classical random walks. In quantum walks, randomness comes from quantum superposition over position states. In the case of classical random walks, randomness arises due to the stochastic transition between the states. Neural network architectures based on quantum walks can be termed as Quantum Walk Neural Network (QWNN). Further details on QWNN will be covered in future posts.
• Quantum big data analytics helps to solve Big Data problems by combining quantum computation and machine learning capabilities which result in better performance than that of classical neural networks or statistical models. These quantum algorithms can be used for various use cases like image recognition, pattern detection, language translation, etc.

Quantum Machine Learning is a relatively new but promising field in the world of computer science. It’s based on quantum computers, which are still being developed and refined to have more power than classical ones. Quantum Computing has the potential to solve some complex problems for data scientists with powerful modelling abilities that existing computing cannot achieve (e.g., modeling climate change). Currently it can’t be applied broadly due its limited availability; however many companies working hard towards making this happen. If you are looking to learn more about quantum machine learning topics, check out this space in near future. I will be posting more blogs on quantum machine learning topics.