Michael Chang, a PhD researcher at the University of California, Berkeley, presents his work at the GRASP Lab Spring 2023 SFI on neural abstractions for modeling and controlling systems, and discusses future human-computer interfaces at DeepMind.
Chang compares task-specific neural networks with the specialized circuits used in early computing. He frames retrieval-augmented transformers as a modern counterpart to the von Neumann architecture, combining a language model with stored information. His broader goal is software that learns to build and act on its own representations of systems, rather than relying on people to define every variable and procedure.
Chang presents slot-based scene representations as a way to separate a visual scene into object-like components. He links stable discrete representations to error correction: iterative updates can pull noisy guesses toward consistent fixed points. For Slot Attention, implicit differentiation improved optimization; on the reported reconstruction task, mean squared error was nearly seven times lower than with the vanilla method.
Chang argues that variables work because the same operation can apply across different values, and uses permutation-equivariant processing to give scene slots that shared treatment. In the OP3 world model, slots represent entities whose dynamics support planning; the model built block towers with three times the accuracy of SAVP, despite training on fewer objects and without tower-building examples. He then describes a second approach that factors object transitions from object type and context, allowing learned transformations to be reused across rearrangement tasks.
Chang describes problem abstraction as coordinating local decision-makers so their equilibrium achieves a global objective. In his market-based setup, options compete through bids, and a second-price auction encourages bids to reflect their value. Each agent learns with local credit assignment, while their transactions collectively select actions for the larger system.
Chang says people often mediate between computers and the physical world, and argues that machines should learn to model and manipulate systems directly. He also wants computers to better understand human communication through speech, diagrams, and gestures, citing work that converts hand-drawn diagrams into HTML. He describes this direction as his next focus as a research scientist at DeepMind, with the aim of supporting human creativity.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Karen Liu, an associate professor of computer science at Stanford University, discusses physics-based human motion simulation and its use in robotics at an MIT Robotics talk.
Dennis Hong, robotics professor and founding director at Virginia Tech, explains why humanoid robots suit human environments and how his lab developed soccer-playing robots and firefighting prototypes.
A ROBOTIS presenter discusses a Darwin humanoid robot that plays Dance Dance Revolution and the vision-processing and motion constraints behind the demonstration.
Alex, role and affiliation unknown, presents the Darwin Drives project with King at an ICRA 2012 challenge; they show a humanoid robot operating an iRobot Create and discuss possible uses for humanoid-driven vehicles.
Students from Purdue University describe their team’s Darwin-OP robotics project at an ICRA 2012 challenge, explaining its sensor-based reactions and demonstrating them.
Dr. Saltz, a digital media and theater educator, and Mike, a programmer, present ROBOTIS’s Darwin animation tools at the ICRA 2012 challenge, including a planned commedia dell’arte performance with robot character Zebzob.
