CSAIL 20/60 Anniversary Celebration, Prof. David Bau - Big Questions and Terrifying Tea

Summary

In this engaging talk, David B, an assistant professor at Northeastern University, shares his journey from industry to academia, focusing on his research in human-computer interaction and machine learning. He reflects on his experiences with programming education, the challenges of re-entering academia after two decades, and the intellectual influence of mentors like Jerry Susman. David discusses his work on improving understanding and control over complex systems like deep learning models and neural networks, highlighting the importance of curiosity, optimism, and embracing unexpected discoveries in research. His insights emphasize the transformative power of fundamental questions and collaborative exploration in academia.

Takeaways

• 😀 David B, an assistant professor at Northeastern University, introduces his journey in academia and his research in human-computer interaction and machine learning.
• 😀 David discusses how his initial career involved working at Google and Microsoft, with a strong focus on creating tools for making computers more accessible to people.
• 😀 He emphasizes the radical nature of teaching young people how to program, drawing inspiration from educational pioneers like Cynthia Solomon and Hal Abson.
• 😀 David reflects on the intimidating experience of returning to academia after 20 years and the challenge of taking his first PhD exam surrounded by much younger students.
• 😀 He shares advice from Rob Miller on selecting research problems that are either known to be hard and important or ones that people don't even know are problems yet.
• 😀 At MIT, David encountered experts like Yan LeCun, who scolded him for asking about understanding deep learning models, pushing him to focus on HCI and education.
• 😀 Through conversations with key figures like Jerry Susman, David realized the significance of understanding and editing machine learning systems and deep learning models.
• 😀 After a period of fear and uncertainty, David embraces Jerry’s advice about solving big, impactful problems in machine learning and develops methods for understanding neural networks.
• 😀 A key breakthrough involved finding causal effects in unsupervised neural networks, which could perform tasks like simulating light transport in a room without labeled data.
• 😀 David discusses how collaboration and openness from mentors like Antonio led him to realize the power of optimism and perseverance in research, especially when exploring unknown territory.
• 😀 The importance of embracing seemingly random discoveries, as illustrated by David's work on editing the knowledge base of language models, shows how unexpected paths can lead to valuable innovations.

Q & A

What motivated David B to return to academia after working for 20 years in the tech industry?

David B was motivated by the idealism he experienced while learning programming as a child. He believed that people should program computers, rather than being controlled by them, which led him to pursue a PhD and later a career in academia.

What was the core research focus of David B's PhD?

David B's PhD research focused on building tools that make it more accessible and understandable for people to control complex computer systems, a response to the growing gap between human users and machine learning systems.

What challenge did David B face when returning to academia after 20 years away?

David B found it intimidating to return to school after a long career in industry. He particularly felt out of place during his first midterm in 20 years, surrounded by younger students who were well-prepared.

How did David B feel about his interaction with Yan LeCun during his first year at MIT?

David B felt scolded by Yan LeCun after asking a question about understanding deep learning networks. LeCun dismissed the idea of focusing on understanding how deep learning systems work, which left David B feeling unsure about his direction.

How did Jerry Susman influence David B's research trajectory?

Jerry Susman gave David B the advice that solving the problem of understanding neural networks would have a tremendous impact, which led David B to reconsider his research focus and eventually pursue the issue of understanding deep learning systems.

What specific breakthrough did David B achieve with his research on neural networks?

David B and his team discovered that neural networks could be understood and manipulated in ways not previously thought possible, such as identifying meaningful neurons and understanding their causal effects in generating images.

What was the significance of the 'exponential graph' drawn by Jerry Susman?

The 'exponential graph' drawn by Jerry Susman was meant to illustrate the potential impact of understanding and manipulating neural networks. It was a turning point for David B, convincing him that tackling this issue could lead to significant advances in the field.

What was David B's approach to solving the problem of identifying meaningful neurons in neural networks?

David B and his team developed techniques to find causal relationships within neural networks by testing neurons for their direct effects on output. This led to the discovery that even unsupervised networks could learn complex features like light transport in images.

How did David B's research on neural networks evolve to impact large language models (LLMs)?

David B applied his research on understanding neural networks to LLMs by developing techniques to edit the knowledge within these models. This included modifying factual knowledge within LLMs, such as changing the city a landmark is located in, and observing the resulting behavior in the model.

What key lessons did David B learn from his time at MIT's CSAIL?

David B learned three key lessons: the importance of asking bold, terrifying questions in research; the essential role of optimism and support from colleagues; and the value of perseverance and openness to serendipitous discoveries in scientific work.