Buu Phan

I am a PhD student at the University of Toronto, where my current research focuses on information theory and statistical machine learning. My advisor is Prof. Ashish Khisti. I recently spent my Summer 2024 as a Research Scientist Intern at FAIR Lab/META with Dr. Karen Ullrich, working on tokenization (or string theory!?!?) and LLMs.

Previously, I was an AI Research Engineer at LG Electronics, working on applied ML for automated checkout and situated reasoning. Before that, I was a computer vision researcher at Algolux, where I worked with Felix Heide and Fahim Mannan on Adversarial Robustness and NLOS Imaging for self-driving cars.

I obtained my Master’s Degree from the University of Waterloo, where I was advised by Prof. Krzysztof Czarnecki and Dr. Rick Salay. My thesis was about Bayesian deep learning, distillation and computer vision. I also spent some time working on anomaly detection and requirement engineering.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

I'm interested in probability, optimization and deep learning. Currently, I am interested in topics intersection between information theory, sampling and deep learning. My long-term research goal is develop reliable AI.

We provide an algorithmic process to convert any tokenized LLMs to token-free (byte-level) LLMs. This naturally solves the odd generation problem in LLM when the prompt ends with mid-token or white space. We also demonstrate the application in model ensemble where each LLM has a distinct tokenization.

We provide a channel simulation method based on Poisson Matching Lemma (Li and Anatharam 2021) for a fixed number of proposals. We also demonstrated its application in distributed compression and vertical federated learning setup, focusing on the scenario where the correlated side information is available only at the decoder.

We provide a information-theoretic analysis on perceptual video compression. Surprisingly, our results show that preserving temporal consistency comes with a cost of higher distortion!

We provide a information-theoretic analysis on the conditional perceptual metric for image compression.

We proposed a new kind of adversarial attack by targeting the ISP and camera optics. Our method produces real-world adversarial scenes without relying on image augmentation.

We show how to scale up NLOS imaging technique in real-world using radar. The method can reliably detect and track NLOS pedestrian and cyclist.

We show that OOD images can be detected by applying the one-class SVM on early-layer's embeddings. The results outperforms several baseline approaches on MNIST, CIFAR and Tiny-Imagenet.

We emperically show that Bayesian deep networks captures some notions of uncertainty that are similar to human's view of uncertainty. Contrary to expectation, we also find that the estimated aleatoric uncertainty from Bayesian deep models can be reduced with more training data.

We provide a tool for synthetic self-driving dataset generation. Our tool can simulate adverse environmental condition such as rain and cloud. We also provide some analysis regarding distribution shift on this dataset.

We show some theoretical disadvantages of using classifier for out of distribution detection. We show that training a classifier for "unseen" class can leverage the issue.

Reconstruction autoencoder was one of the go-to approach for out-of-distribution detection. The idea is to use the reconstruction error for OOD score. Nevertheless, in many cases, we observe that OOD images are perfectly reconstructed. This paper demystifies the above phenomenon from the view of PCA.

We observe that Bayesian Deep Networks do not provide well-calibrated regression uncertainty. In this paper, we address this problem for the single-object localization task by adapting an existing technique for calibrating regression models.

We proposes an approach and architectural design for achieving maximum functionality in the case of system failures.