
Why did our #PhD students choose TTIC? In our new video, hear our students discuss their #computerscience research, their community, and their experience at the institute: buff.ly/8YsOWeC

Challenging the Foundation of Computer Science. Bluesky: TTICconnect

Why did our #PhD students choose TTIC? In our new video, hear our students discuss their #computerscience research, their community, and their experience at the institute: buff.ly/8YsOWeC

Join a fully-funded, highly active research environment with a 4:1 faculty-to-student ratio, strong mentorship, and access to the University of Chicago campus and facilities. Learn more about our Ph.D. program in #computerscience and apply (FREE) by Dec 9: buff.ly/xo0DVMS

6/6 "The challenge with both of these settings is that our environment doesn't look nearly as clean, and it's structure is not nearly as uniform, as the ones in which [the policies are trained]," said TTIC's Matthew Walter.

5/6 The second dominant approaches uses reinforcement learning (RL). In RL, "we don't need demonstrations, but instead we need some reward specification," said TTIC's Matthew Walter.

4/6 The first of two dominant approaches to train robot policies in the closed-loop setting is imitation learning, using human-annotated data. "This [turns] the problem of finding an optimal policy into a supervised learning problem," said TTIC's Matthew Walter.

3/6 "We can make this a little bit more sophisticated, and we could close the loop," said TTIC's Matthew Walter. In this case, the robot decides what to do based on the current state of the world. "Now, the agent is reacting to changes in the environment."

2/6 For decades, robot's have had their actions, or policies, hard coded (aka open-loop control). "This works well in factory automation ... because the environment doesn't really change, at least not enough to violate these policies," said TTIC's Matthew Walter.

1/6 "We want a robot that doesn’t only work in factory settings, [but] ideally would work in our homes," said TTIC's Matthew Walter, in his talk on "Learning from Visual Experience: Toward Generalizable Robot Manipulation." Video: youtube.com/watch?v=WKGIFAGu…

Friday, November 20th at 11:00AM: Research at TTIC presents Kaylene Stocking with a talk titled "Building the field of physical AI safety." Please join us in Room 530, 5th floor.

Thank you to Hyde Park Academy students for visiting TTIC! Professor Karen Livescu, Professor Brian DuSell, and student Jose Cruzado presented their ongoing research. Learn more about our community outreach: buff.ly/5EJd7Oc

1/4 Out-of-distribution generalization is "one of the biggest bottlenecks...[to]...real deployment of machine learning in many practical scenarios, such as in autonomous driving, chip design, drug design, etc.," said Yusu Wang of @UCSD at TTIC. Video: youtube.com/watch?v=toRAey5V…

4/4 "Ultimately, with all this understanding, we want to be able to design new frameworks...that can help solve hard problems that we cannot handle very well currently," said Yusu Wang of @UCSD at TTIC.

3/4 The question of OOD generalization can be broken down into non-trivial subquestions, such as: can neural networks naturally learn algorithms?; how can one probe what the NN has learned?

2/4 For OOD generalization, the overarching question is: "When and how can neural architectures learn generalizable 'procedures,' rather than fitting labels in input data," said Yusu Wang of @UCSD at TTIC's Learning-driven Algorithms and Machine-aided Proofs (LAMP) workshop.

Friday, November 13th at 11:00: Research at TTIC presents Will Merrill (@lambdaviking) with a talk titled "From Theory to Practice and Back." Please join us in Room 530, 5th floor.

Presenting at #FOCS2026: “Learning Confidence Ellipsoids and Applications to Robust Subspace Recovery” (Chao Gao, Liren Shan, Vaidehi Srinivas, Aravindan Vijayaraghavan). Wednesday, Nov. 11, 4:56PM 9C, Rooms 905/907