Oscar Scholin

OSCAR (named by Professor Weiqing Gu) is my first project at Dasion as as Machine Learning Engineer, of which I am senior personnel / lead researcher. OSCAR stands for Ocean Science Cutting-edge Anomaly Research, and the focus is developing an interpetable, differential geometric approach to real-time, energy efficient decision making in underwater autonomous vehicles. Helped co-author a grant proposal and a patent application. Currently, I have tested on the Watkins Marine Mammal acoustic dataset, achieving 99.3% validation accuracy and 98.7% test accuracy on distinguishing 39 different classes of mammals on a max depth 5 50 esimtator eXtreme Gradient Boosting model. In collaboration with researchers at MBARI (Monterey Bay Aquarium Research Institute) to extend this model to other kinds of data.

For my physics senior thesis with Professor Lynn of Harvey Mudd, I have determined the number of bell states with single particles of dimension d an LELM (linear evolution and local measurement) device can distinguish using an analytic argument. The answer is k = 2*d-1 out of d^2. This question is motivated by the audacious assumption that one doesn't possess a quantum computer ready at hand but is interested in performing some algorithm or communication protocol somehow involving a bell state measurement. This extends the work of Pisenti et al. (2011), who solved d = 2^n, and Leslie et al. (2019), who solved d = 3. A manuscript is in progress. I am also attempting to solve the odd dimension case to complete this problem. To see my full thesis, visit here.

Hydraluric acid is an important component of the cell outer envelopes of various anaerobic bacteria. Working with Dr. Lee, Dr. Yu, Xinyi Dong, Yomna Mohamed, Daiki Kawagishi, and Michelle Ho, we are attempting to describe the quantum effects of electron transfer in a single polymer chain of this molecule using the technique of Guassian Boson Sampling, a method of photonic quantum computing. I am developing a qumode reduction algorithm that is a hybrid quantum-classical machine learning algorithm to enable simulation with noise on current photonic devices.

For my math senior thesis with Professor Ami Radunskaya of Pomona College, I implemented the wormhole teleportation protocol of Jafferis et al. (2022), which examines a two cloud system of Majorana fermions as dual to a gravitational system with two black holes, in Qiskit and obtained experimental data from the IBM Kyoto processor on a simplified 3 qubit setup trained to replicate mutual information curves from the Jafferis paper. I also ran a full simulation of the protocol as shown above for 12 qubits, 333 CNOT gates, and 171 U(3) gates, testing a variety of ansatzes for the variational quantum eigensolver (VQE) and mu values (coupling parameter between the two black holes). The only sensible ansatz was the first (should be no mutual information at mu = 0), but the other mutual information curves do not seem to follow the general trend reported by Jafferis et al. To see my full thesis, visit here. We are working on a new training procedure and ML architecture in order to preserve scrambling dynamics while sparsifying the SYK Hamiltonians that describe the evolution of the system.

Wrote a lot of custom code to generate, manipulate, and measure (theoretically and experimentally) 2-qubit quantum states. Trained a variety of machine learning models (eXtreme gradient boosting, neural networks) on 4 million generated states with the goal of predicting the optimal set of next measurements to take in order to most efficiently verify entanglement based on an initial set of projective probabilities, using entanglement witnesses building on those by Riccardi et al. (2019) and previous work by the group. Achieved 4% increase in performance from previous models and successfully applied the models to experimental data. Presented results, "Entanglement Witnessing: a Neural Network Optimization and Experimental Realization", at Southwest Quantum Information and Technology (SQuInT) conference, October 2023. Also experimented with an automatic decomposition of a quantum state into Jones matrices via gradient descent, which achieved up to 99.3% fidelity in our experimental setup.

Independent research project with Larry Liu (Pomona '24), Tom Tang (Pomona '24), Donny Lu (Pomona '24), Song Song (Pomona '24), and Professor Jason Gallicchio of Harvey Mudd College. We focused on examining the current approaches to actual quantum computing, in which we started with linear optical conditional gates in photonic systems, then moved to photonic chips, Gaussian Boson Sampling, superconducting qubits, trapped ion qubits, topological qubits, and free electron qubits. We produced a review paper and an in depth annotated literature review, both of which are aimed at other interested advanced undergraduates. See our review paper here.

Machine learning intern at Dasion (Data-To-Decision) working with Professor Weiqing Gu of Harvey Mudd on problems involving differential geometry. Developed code to classify the MNIST digits without using regression or neural networks, but by understanding the geometry of the digits. Achieved accuracy of 64.2% on all 10 digits; 93% on comparing 0, 1, and 9. Using a 2 hidden layer neural network on image moments, eachieved up 92.6% accuracy on all 10 digits. See Video presentation 1 and Video presentation 2 for more details.

Created a custom backend for the website oscars47.github.io/math-zombies, described in the above graphic. A password protected Google form is used to submit blog posts for verified users, which calls a Google Scripts file to create automatically formatted HTML file for the post as well as a mini description of the post to go on the main post page. This file calls a Javascript file running on an Amazon EC2 server to automatically insert the mini description onto the main page in HTML and then add the full post and image as a separate Git branch. A merge request is then made to the main branch, which is reviewed by me upon the generation of an automatic email. The website is hosted on Github Pages and currently has two other blog posters, Professor Gizem Karaali of Pomona College and Kamden Baer (Pomona '24).

p-stars is a team of 12 undergraduates from Pomona College, Harvey Mudd College, and Pitzer College with the aim of classifying variable stars using unsupervised clustering. We are working primarily with the ASAS-SN catalog of variable stars, which is a collection of ~300,000 variable stars with labels, building off of my work in Fall 2022 detailed in my research report. We are developing a metric for our input space of variability indices, a collection of 36 different statisitical functions designed to quantify the variation of a star's magnitude over time. Moreover, inspired by Valenzuela et al. (2018), we are building a tree structure based on the idea that one can imagine slicing a lightcurve (brightness over time) into subsequences that correspond to tokens in a large language model (LLM). We are also considering the Hubble Catalog of Variables (HCV), which is a set of 84,000 unlabeled objects with much more sparse data per object in order to increase the versatility of the process. Previously my research partner Graham Hirsch and I developed a restrictive search algorithm in the summer of 2022 and found 2 potential tidal disruption events, which if confirmed would be the first experimental observation of a black hole found in the HCV. Pomona College also operates the Table Mountain Observatory, which we can use for follow-up observations.

Inspired by a paper about the strange behavior of neuropeptides in the worm C. elegans by Ripoll-Sánchez et al. (2023), I created this simple quantum circuit in Cirq as a possible hybrid quantum machine learning algorithm that takes classical inputs, converts them into quantum states via phase encoding, entangles all the states together--modeled on the action of the neuropeptides--to adjustable levels, and then performs single qubit rotations before converting back to classical probability output via measurement.

Designed, implemented, and trained a custom long short-term memory (LSTM) recurrent nerual network called "Thinking Parrot" on the works of the ficitous scholar Nasrudin based on the lines, "To save money, I made my donkey go without food. Unfortunately the experiment was interrupted by its death. It died before it got used to having no food at all. People sell talking parrots for huge sums. They never pause to compare the possible value of a thinking parrot.". I wrote two essays of over 120 words total explaining the model to a non-CS audience and intepreting its results literarily.

Worked with Seohyeon Lee (Pomona College '24) and Marwin Bit (Harvey Mudd College '25) in a project to rank top songs. I designed and implemented a automatic pipeline for given a list of songs, extract their metadata via Spotify API, download them locally and convert into spectrograms for input to a convolutional neural network to classify them.