Maximilian Du

When At First You Don't Succeed: Knowing When to Try Again in Novel Test-Time Scenarios

Maximilian Du, Sasha Khazatsky, Tobias Gerstenberg, Chelsea Finn

Ongoing Project

When faced with a difficult task, a human may try different learned strategies until they discover a successful approach. In this project, we try to create a similar try-retry paradigm for robots. First, I train the robot to remember a set of strategies from expert demonstrations. Then, I train a counterfactual model that compares the robot to a hypothetical expert. If the robot does too poorly, we query it for another strategy. We've seen promising results and we are currently scaling it to harder tasks.

Behavior Retrieval: Few-Shot Imitation Learning by Querying Unlabeled Datasets

Maximilian Du, Suraj Nair, Dorsa Sadigh, Chelsea Finn

Robotics: Science and Systems 2023

Often, robot data isn't shared across projects because of different task / environment conventions. To get past this problem, we create a data filter that selects points that are relevant for a particular task. To create this data filter, we train a state-action embedding that captures a meaningful notion of task similarity. Then, we select points based on closeness in the embedding space. On a real robot, we were able to leverage a precollected dataset (from another institution) to boost performance by more than 35%.

Website  /  Paper  /  Code

Play it by Ear: Learning Skills amidst Occlusion through Audio-Visual Imitation Learning

Maximilian Du*, Olivia Lee*, Suraj Nair, Chelsea Finn

Robotics: Science and Systems 2022

For certain robot applications like search & rescue, we need more than vision. Humans and animals consult a combination of senses for decision-making. In this project, we make an analogous robot system that reasons across audio and vision. We learn policies end-to-end from RGB vision and audio from a gripper-mounted microphone. The learned policies can accomplish difficult tasks, like extracting keys from a bag when the keys are not initially visible.

Website  /  Paper  /  Code

Improving LSTM Neural Networks for Better Short-Term Wind Power Predictions

Maximilian Du

IEEE Renewable Energy and Power Engineering (REPE) 2019

Wind power is a crucial renewable energy source, but it is less stable than conventional energy sources. In this project, I created an autoregressive wind power predictor that can help power grids proactively balance grid loads. I gave the model context from weather stations, which doubled the accuracy of the wind power predictions.

Paper  /  Code

Can I Train my Robot like my Dog?

Reinforcement learning in AI has origins in animal training in the 1900's. While much has been abstracted away since then, I take another look at this original connection between CS and psychology. In this 90-minute talk, I explore the interplay between challenges faced in robot labs and dolphin stadiums. We explore behavior cloning, reinforcement techniques, models of learning, and more.

I give this talk twice a year at Stanford Splash, an educational outreach program that serves students in the Bay area, especially students from underrepresented and disadvantaged backgrounds.

Slide Deck Sample  /  Stanford Splash Program

Section Leading for CS 106A/B

I was a section leader for Stanford's introductory CS106A and intermediate CS106B computer science courses. As a section leader, I hosted weekly sessions to review lecture content in a small group setting. I also hosted office hours, graded homeworks, provided interactive feedback, and helped graded exams.

Section Leading Program

Mentor for Deep Learning Portal

I will be a mentor for the Deep Learning Portal outreach program starting Fall 2023. The Portal is a program created from the IRIS lab and the Stanford CS department, designed to help disadvantaged students learn AI by providing access to existing online courses and hosting live office hours. I will be helping with office hours, where I will be debugging code and answering conceptual questions.

Deep Learning Portal Program

Looking Under the Hood of DetectGPT

CS224N Final Project 2023

As Large Language Models (LLMs) become more advanced, there's a strong push to create effective detection algorithms. In this work, we look at a new detection algorithm originally proposed by Mitchell et al: DetectGPT. We proposed and tested ways of improving DetectGPT by focusing on parts of speech. We also demonstrated that DetectGPT can be partially fooled by an adversarial prompt. Finally, we tested DetectGPT on ChatGPT.

Paper

Sixteen Pixels is (Almost) All You Need: Crafting Parameterized Image Uncrumpling Models

Winning CS231N Final Project 2022

Phone document scans can suffer from distortions from a crinkled document. We created a decrumpling model that will take in an image of a crumpled document and smooth it out. We find that an adversarial paradigm with a small PatchGAN yields the most realistic results with the best quantitative scores as well.

Paper  /  Code

Thompson Sampling Simulator

CS 109 Winning Final Project 2021

Thompson sampling is one approach to exploration, where you sample from belief distributions and test the samples in the real world. It is one way of balancing exploration and exploitation. In this simulation of Thompson sampling, we visualize ants finding a good location for a nest. We also implement Tandem Running, which allows ants to "persuade" other ants, resulting in faster convergence.

Code

Media Annotator

This simple Python-based program allows you to use keyboard shortcuts to annotate audio, video, and live events with timestamped comments. The program will export your annotations to copy-and-paste text that you can add to any literature review notes. I rely heavily on this tool to review hours of videos for my book.

Recently, I've also made a book annotator that allows you to make page-specific notes. This allows you digitize your annotations for physical books.

Code

Coding Basics

As a researcher, I often find that I use the same code many times in different applications. To help, I've been working on a large repository of code basics. It's a collection of code snippets that help with model debugging, plot making, basic PyTorch models, and more.

Code

Course Notes

I learn the best when I take notes on things. Through my four years at Stanford, I've taken hundreds of pages of notes on topics that range from probability to psychology. They cover all levels of knowledge, from undergraduate introduction to cutting-edge AI algorithms. I'm working on making all of these notes public, and i hope that they can be helpful to some people.

Notes (actively updating)