Why Data Scientists Should Know About Quantum Annealing
D-Wave’s quantum computing strategy avoids the challenges of logic gates in favor of prioritizing streamlined optimization problem speed-ups
If you're in the data science and engineering field like I am, you're probably motivated by a deep sense of curiosity and desire to understand how complex technologies work. You know that computing power is a limiting factor, and that quantum computing is promised to provide “exponential” speedups for many applications. However, there are also many credible skeptics who say it’s all hype and no substance. I’m not here to answer the question that every tech magazine attempts: “What is quantum computing and how will it change the world?” but rather to show how one specific enterprise is innovating in this space, and why I think this is specifically relevant for data science and machine learning use cases.
Wait, how did we trick a rock to think?
The metaphor of a computing chip as a ‘rock that can think’ is one that has stuck with me. Most scientists and engineers have built a basic circuit, but the jump from a potato and a lightbulb to a calculator and a computer screen is a chasm indeed.
Computer chips must manipulate natural phenomena (electrical charge or quantum particle) in a way that mirrors the problem we are trying to solve and therefore allows us to ‘calculate’ the answer using that natural phenomena.
For a math equation, we can map complex problems into simple (but often tedious) addition. Thankfully, bits are able to be added together using some basic logic that can be physically built into a circuit using “Logic gates”. Logic gates take input, manipulate it, then give an output reading. But for quantum circuits, we use quantum particles instead of an electrical charge. However, maintaining the flow or absence of electricity is much easier than maintaining the state of a quantum particle long enough to calculate something with it. This is the problem of coherence — how long can you keep the qubit stable?
Gate-based quantum computing is often graded on these factors: how many coherent qubits were created, how long were they coherent for, and then how relevant is the chosen algorithm for solving this problem?
What if we instead of using logic gates to manipulate qubits, we instead simply prepared a quantum system then let the system guide us to the answer within microseconds?
The Science Behind Annealing
Annealing is a concept that is commonly used in metallurgy or materials science. It involves heating and then slowly cooling a metal to increase its toughness and hardness for better longevity. The connection between these two concepts lies in the adiabatic theorem of quantum mechanics. The adiabatic theorem postulates that a quantum system in its ground state will stay in its ground state given that the environment changes are sufficiently minimal (this means minimal external influences from temperature fluctuations or radiation). This quality of “staying in its ground state” allows us to set up our qubits in a way that reflects our problem, then see what the ground state of that set up is.
Get Lauren Phillips (Cotton)’s stories in your inbox
Join Medium for free to get updates from this writer.
Returning to a ground state is easily analogous to optimization problems, where the goal is to minimize a combination of factors or find the best balance. By building a quantum system and framing our problem correctly, we can let quantum particles show us the answer.
Building a Quantum Annealer
The qubit’s state in D-Wave’s quantum annealer is indicated by the direction of a circulating current, such as clockwise for the binary value 0 and counterclockwise for the binary value 1.
Now it becomes important to remember a key visual from every introductory physics course: the Energy Diagram
In this representation, we can imagine the particle like a ball thrown into a well. It wants to live at its ground state energy (the lowest point in the curve), but in that well its states of CW or CCW circulating current are in a superposition. A random 50/50 chance for CW or CCW is not useful to us — that would make it a coin toss — but how can we influence the probability that it will be either 0 or 1 to encode details of the problem we want to solve?
Qubits are useful because they represent probabilities that collapse into one answer. By applying an external magnetic field, we can bias the probabilities (such as shifting the odds to 75/25 for collapsing to the 1 state rather than 50/50) and encode information about our problem. Now let’s say we have biased all of our qubits, what’s next?
Entanglement is the Key
When two particles become entangled, they operate as one unit. The binary states of 0 and 1 increase exponentially from 2 distinct states to 4. Where before each particle was either 1 or 0, now the two entangled particles could be 10, 01, 11, or 00.
Below shows a visualization of the probabilities of these entangled particles.
Imagine this premise with thousands of qubits. Each encoded to properly represent the problem at hand. The intuition becomes clearer as to why this method would be immensely useful for optimization problems. Once entangled, the system wants to reach that lowest energy well. Now imagine all of the tiny calculations that would be needed to evaluate a system of this size in a classical computer!
This method harnesses the value of quantum particles’ natural behavior (explained by the adiabatic theorem) to determine the most optimal answer to a problem. It does this by framing the answer as the ground state energy of the quantum system.
How do we encode our problems into this chip and what are the limitations? Essentially, you must determine the weights (how much to bias each qubit) by converting whatever you want to solve for into a specific optimization problem format. These problems are called “Quadratic Unconstrained Binary Optimization Problems”. Some of the key things we can optimize are Feature Selection (optimizes for a balance between descriptive features and reducing redundancy) as well as Clustering (optimizes for both representative and unique clusters) among a litany of other optimization tasks.
D-Wave has also shown a prioritization towards machine learning applications by developing a quantum-ai toolkit that “provides direct integration between D-Wave’s quantum computers and PyTorch”
After spending a month researching how quantum computers are actually able to deliver useful and novel results, quantum annealing stood out for its ability to side-step the headaches associated with gate-based quantum computers by utilizing a unique strategy in the market. It’s important to understand the ground level of these technologies as they increase in scale and effectiveness — they may soon be the go-to tools that data scientists use to increase compute power and speed for some of our toughest problems.
No AI was used in the creation of this article. Ai-generated writing is an important cause to me, and I am passionate about writing things myself rather than outsourcing. Thank you for your time.
Disclaimer: I do not own the diagrams shown in this article, I am not financially profiting off their use, and I post them with credit to the original source.
