Internship Supervisor: Dr. Graham Taylor, Vector Faculty Member
Current: B.Sc student at UofT

What did you work on and what did you learn during your internship at Vector?

During my research internship, I had the opportunity to dive deep into an incredibly engaging project that spanned across hardware acceleration, probabilistic machine learning, and bioinformatics, focusing on scalable algorithms for DNA barcoding. One of the most exciting aspects of this internship was writing custom CUDA kernels for GPU acceleration, which was an entirely new experience for me, and it exposed me to the intricacies of parallel computing and optimizing machine learning algorithms for GPUs. While the initial learning curve and challenges were daunting at times, the rewards in terms of achieving state-of-the-art performance on the task was immensely satisfying.

How did you discover you were interested in AI?

My interest towards Artificial Intelligence wasn’t an immediate leap, but rather a gradual transition that stemmed from my original background in biology. When I began my first year in university, I was driven by a fascination with the complexities of biological systems, and I challenged myself to implement a neural network completely from scratch. Though my initial experiments were rudimentary, I was hooked by the prospect of models that could learn from data, and their potential in fields such as medicine or genomics. This sparked my curiosity and motivated me to delve deeper into the field of machine learning.

Internship Supervisor: Dr. Juan Felipe Carrasquilla, Vector Faculty Member
Current: Quantum Software Engineer,  Zapata Computing
Previous: MSc Student, Physics, University of Milan

What did you work on and what did you learn during your internship at Vector?

My research project was focused on the study of Quantum Spin Systems by leveraging the computational power of Neural Networks. In condensed matter theory, quantum spin systems are simplified models devised in order to gain insights into the microscopic mechanism of magnetism, which intrinsically involves quantum many-body phenomena. In spite of their simplicity, such models are extremely challenging to study theoretically, thus computational approaches have been predominant over the past several decades. Among those, a powerful option is the Variational Monte Carlo method. During my internship, I investigated if and how introducing a Neural Network as a component of this algorithm might increase its performance and provide a better estimation of the main properties of the fundamental state of the spin system. Specifically, my research was focused on utilizing a Restricted Boltzmann Machine: the AI component was extremely interesting to learn about, and play around with!  

What are some of the highlights of your experience at Vector?

The internship at Vector and the collaboration with my supervisor Juan Carrasquilla, was extremely insightful and valuable. It gave me the opportunity to deepen my knowledge about Physics and more importantly, it allowed me to enter the fascinating world of Machine Learning. Connecting with other Vector Researchers in the Vector space exposed me to countless Artificial Intelligence techniques and use cases, and allowed me to get in touch with cutting-edge research and brilliant people!

During my internship at Vector, I also had the opportunity to enter into contact with the industry environment, something that I didn’t have the chance to experience earlier on. This was fundamental in changing my career path from academia to industry.

Internship Supervisor: Dr. Chris J. Maddison, Vector Faculty Member
Current: PhD, Stanford University

What did you work on during your internship at Vector?

Our main project was about data compression for downstream machine learning rather than human perception. Indeed, standard compressors (e.g. JPEG) are optimized such that reconstructions look similar to humans. Nowadays, most data is not seen by humans but rather processed by task-specific algorithms, and storing data in human-readable formats can be prohibitive. On the theoretical side, we formalized compression for downstream machine learning by characterizing the best achievable bit-rate to ensure high performance on all desired tasks. On the practical side, we showed how to train such compressors in an unsupervised manner to get more than 1000x compression gain on standard images compared to JPEG. In the long-term, we hope that this type of compression will enable individuals to process data at scales that are currently only possible at large institutions, and our society to take advantage of large data sources in a more sustainable way.

Read the paper Lossy Compression for Lossless Prediction that was accepted at NeurIPS 2021.

What are some of the highlights of your experience at Vector?

Near the end of the internship, I was invited to give a long talk about our compression work at a Vector’s Thursday talk seminar. That was a great opportunity for me to learn how to present my work, let others know about our results, and interact with other researchers who had interesting feedback. I also had other valuable opportunities to present my work in Chris Maddison’s lab and Dan Roy’s lab.

I was also shown how to maximize the impact of one’s research and collaborate more effectively. Post internship, I have continued to collaborate with Chris’ lab, and really hope to continue doing so in the foreseeable future. Chris has been a true mentor that has continued to give me helpful career and life advice even after leaving Vector.