Quantum Machine Learning — what does the future of ML look like?
Quantum Machine Learning (QML) is a rapidly growing field that combines two of the (arguably) most exciting areas of science today: quantum computing and machine learning. QML is a promising new approach to solving complex optimization problems that are difficult or impossible to solve with classical computers. In this article, we will explore the future of QML and what it holds for the field of machine learning.
In order to understand QML, let’s look at quantum computing first. Quantum computing is a type of computing that uses the principles of quantum physics to perform computations. In contrast to classical computing, which uses bits that can only represent one of two states (0 or 1), quantum computing uses quantum bits (qubits) that can represent multiple states simultaneously. As opposed to a single bit being either 0 or 1, a qubit, using the principle of superposition, can be both 0 and 1 at a given time. Going into an explanation of superposition is beyond the scope of this article, so please refer to this excellent explanation on superposition and other fundamentals of quantum mechanics.
For this article, it is enough to understand that taking advantage of some quantum phenomena allows us to perform computations much faster than what is fundamentally possible with a classical computer. The use of the word ‘fundamentally’ is critical in that sentence, because taking advantage of these quantum phenomena allows for capabilities that would be theoretically impossible with a classical algorithm. For example, Grover’s search algorithm allows for searching in an unsorted list in O(1) time. If you know anything about classical computation, you know that searching in a list has to take O(n) time, the best you can hope for is if the list is sorted, in which case, O(log(n)) can be achieved. Thereby, the idea of searching on a list in O(1) time, which means that the time taken for the search operation is independent of the size of the list, might seem ludicrous. However, the phenomenon of superposition and entanglement allows for impossible-seeming feats such as this.
Many different quantum algorithms such as this can speed up various classical algorithms. In this article, we take a look at how these advancements can be used for supercharging Machine Learning.
Quantum Algorithms to speed up base operations
A basic use-case of QML is not to dramatically rethink how we do ML, but to speed-up specific operations within ML. In this paradigm, we use a combination of a classical chip as well as a quantum chip to do the full set of operations. For example, we set up the problem to achieve ML exactly how we do it now, but certain operations like matrix-matrix multiplication, matrix inversion, Singular Value Decomposition (SVD) etc can be offloaded to a quantum chip for much faster operation.
In practice, we would set up a classical neural network where the input data, the intermediate network weights, activation functions, etc are set up in the normal bit representation, and we train the network with Stochastic Gradient Descent running on backpropagation. However, instead of the matrix-matrix multiplication being performed on Graphical Processing Units (GPUs), it will be performed on Quantum Processing Units (QPUs). This has the potential to be orders of magnitude faster than what is possible with GPUs, meaning we can achieve more number of epochs of network training in less time.
Processes like this can also be completely game-changing for certain ML applications like time-series analysis. For many time-series analysis operations, SVD is capable of providing the best theoretically optimum results, but the data is usually too large for it to be done efficiently. Which is why we have to fall back on modern techniques like Recurrent Neural Nets, Transformer Nets, Attention Mechanism etc. These are computationally feasible if trained correctly and with enough data and can adequately approximate the theoretically optimal results. However, SVD might still be capable of producing better results if it was computationally feasible on large datasets with a classical computer. Hence, having a quantum system capable of fast and efficient SVD might be able to replace modern ML algorithms for certain applications. This is especially true in the case of finance, where being able to SVD on large sequence of share price data might solve a lot of problems in finance.
Using purely quantum systems
In addition to being able to support classical techniques, there is another class of quantum computing which aims to (almost) replace classical chips altogether. Although at the moment, these techniques are only capable of only solving a very specific class useful of problems. That is to say, these hardware will not be able to do general computations, but only solve very specific computational problems. For example, it is possible to build a quantum computer that can compute very fast solutions to the Ising model problem.
The Ising model is a mathematical model used to study the behavior of interacting particles in a system, such as magnetic spins in a material. The model consists of a lattice of particles, where each particle can have a spin of either up or down. The spins of neighboring particles interact with each other, and the total energy of the system depends on the configuration of the spins. The Ising model problem involves finding the lowest-energy configuration of spins that minimizes the energy of the system. It is difficult to find an exact solution to this problem, particularly for large systems, due to the exponential growth in the number of possible configurations as the system size increases.
The solution to this problem has applications in areas such as image segmentation and community detection in social networks. It is possible to create a quantum computer that solves just the Ising model. Although this chip will not be able to solve anything else, it can come up with solutions to the Ising model much faster than any classical system can.
A fair question to ask at this point is if a system like this can be practical. After all, how can you expect to sell a chip if it can only ever solve a single problem? However, companies might do a business model where they lend out their chip online, much like how large cloud service providers have on-demand access to GPUs now. Any analyst wanting to solve a social network or a large image segmentation problem might upload the problem to their quantum servers and get a much faster/more accurate response than waiting around to use a machine learning algorithm on their classical computer.
Similar to this, there are problems in drug discovery and biology research which can theoretically be solved by classical algorithms, but are intractable for the magnitude of any useful input space. For example, a very impactful area where QML can help is the simulation of molecular systems.
The simulation of molecular systems on quantum computers involves encoding the electronic structure of the molecule as a quantum state and performing quantum operations to simulate the behavior of the electrons. This can enable researchers to study the properties of molecules that are difficult to measure experimentally, such as the behavior of excited states, and to design new materials and drugs with improved properties.
Final thoughts
In conclusion, the future of QML is very promising, and it has the potential to revolutionize many areas of science and industry. QML has the potential to accelerate the development of new drugs, improve supply chain management, and solve many other complex optimization problems. However, many challenges still need to be overcome, and it will likely take several years before QML becomes widely adopted. Nonetheless, QML is an exciting new field that holds great promise for the future of machine learning.
Written by: Sertis Vision Lab
Originally published at https://www.sertiscorp.com/sertis-vision-lab
