The Senior Project is an independent, student-led culmination of our high school experience. After three years of academic preparation, our seniors are ready to spend the last trimester of their high school careers applying the skills and knowledge they have gained to develop a project that is insightful, academically rigorous, and professional in nature. This year, we are proud to showcase a senior from one of our neighboring campuses, BASIS Independent Fremont, Patrick Z.

Week 3: My Qubits Can Count, Just Not to Ten

Last week, I hit a major computational roadblock. I had to wait half of an entire day for the quantum simulations to finish. Oftentimes, they did not even work. However, I found my breakthrough with Amazon SageMaker. While I spent the final part of last week struggling with Google Colab’s limitations, this week I decided to port my notebook to Amazon’s machine learning platform: SageMaker. This gave me greater access to more powerful computational abilities. I had been spending hours training models on Colab, but now I could train models in only a fraction of that time. I could experiment, tweak, and retrain much more quickly than before, which is necessary for model development. With SageMaker, I could finally achieve what had been my goal for weeks: training all three models and getting preliminary accuracy numbers on the board.

I ran initial training jobs on all three of my models using the clean MNIST dataset from Keras. For the first time, I had actual figures to refer to. The full-resolution CNN was the strongest out of the three models, and this was honestly expected because it had the full 28×28 pixel images. To give you an idea of what this looks like, here’s an example of the input the full CNN receives.

The fair CNN, my MLP running on the same 4×4 binarized input as the QNN, was less performant but still showed the ability of a simple classical neural network to squeeze a decent amount of data out of the compressed input. Here’s what the compressed input looks like.

The QNN also produced its first accuracy figures. While the accuracy was nowhere near as good as the full CNN’s, seeing the quantum circuit learning and improving its accuracy was still exciting. For the first time, my project was more like an actual experiment than a debugging exercise.

But then I got greedy. Feeling good about having working models, I wanted to try to push the limits. I tried to have the QNN classify all ten digits rather than just the simple subset the initial version was trained on. So, I reworked the output layer as well as the loss function and started the training. It was so painfully slow. While the new hardware Amazon SageMaker provides is great, every additional output class of the QNN means more parameters in the quantum circuit, more calculations using the parameter shift rule, and more simulated quantum operations. These simulated quantum operations stacked on top of one another and made the program extremely slow. I tried different learning rates and tweaked the number of entangling layers, but it was just too slow. By the time I realized the ten-class approach was not going to work, I had already wasted the better part of two days on it with very few results to show.

However, I am not discouraged by this setback because I believe that the preliminary results I achieved prior to the ten-digit experiment are promising. In the future, I’m planning to work on optimizing the quantum circuit architecture itself and exploring different combinations of quantum gates beyond just XX and ZZ. I want to see if reorganizing the quantum circuit can help increase the classification power of the same 16 qubits. Specifically, I aim to determine the sweet spot where I can confidently mitigate noise, which is the whole purpose of the project. Beyond that, I am also interested in exploring how the models can be applied to more meaningful datasets than just MNIST. While MNIST is a great benchmark, classifying handwritten digits does not fully capture the challenges of the noisy data that these systems would encounter in more practical applications, such as medical imaging and autonomous driving.

Week 4: The Dataset Dilemma

Last week ended on a pretty high note for me. After many days of frustration, I was finally able to get the ten-digit QNN classification working with Amazon SageMaker. I did this by optimizing batch sizes and being more aggressive with my learning rate schedule to make sure that my quantum circuit was able to converge before my patience gave in. Seeing all ten digit classes separate out in my predictions seemed like a small miracle to me. So, I was looking forward to continuing with more datasets in Week 4. But then, my datasets caused quite a lot of trouble.

After having MNIST in the bag, I was feeling quite confident with my project, so I decided to try to apply my project to some real-world problems beyond just recognizing handwritten digits. So, I decided to test my models out with something more challenging than MNIST. During the first half of this week, I was looking into using the Fashion MNIST dataset, which contains images of various clothing items such as shirts and shoes. I felt like replacing my MNIST data with this new, more complex set of visual data meant more.

The results were a disaster. The full CNN performed reasonably well with the Fashion-MNIST dataset and its full resolution images. However, the fair CNN and QNN plummeted. This was because compressing a t-shirt and a pullover into 16 pixels makes them almost indistinguishable. The loss of information was fine for the digit dataset but disastrous for clothing items with only subtle visual differences. My QNN was basically guessing.

I tried different binarization thresholds and only used visually distinct classes, such as distinguishing between shoes and bags. However, even these simple two-class problems were not reliable. After two days of failed experiments, I gave up and accepted that my 4×4 input resolution was a hard constraint dictated by the limits of quantum simulation. It was simply not capable of capturing enough information to classify more complex images. MNIST worked because of the simplicity of the digit images. Fashion-MNIST did not.

So, I made a decision. I’m temporarily abandoning the Fashion-MNIST dataset and going back to MNIST. But, I might search for some more datasets of traffic lights to experiment with. Looking back, my experiment was never about Fashion-MNIST anyways, it was more about determining whether quantum computing’s properties provide noise resilience. I can still do this experiment with sufficient rigor using other datasets.

Next week, I’m also ready for more in-depth noise injection. Stay tuned.

BASIS Independent Dublin is a Grades 6 – 12 private school, providing students with an internationally benchmarked liberal arts and sciences curriculum, with advanced STEM offerings. Considering joining the BASIS Independent Dublin community? To join our interest list for the next school year and receive admissions updates and more, please click here.