U of T Robotics Institute Seminar: Aran Nayebi (CMU)

In this talk

1. 1:00

   Why physical interaction matters

   Nayebi argues that photorealistic AI can still fail to model how objects behave: he points to a generated chair that appears from sand, floats, and changes shape. He proposes that learning through interaction could help AI build more physically grounded models. Animal behavior, he says, also offers concrete targets for capabilities such as prediction, memory, and learning over time.

2. 6:24

   Training models to explain brain responses

   Nayebi describes comparing task-trained neural networks with recorded brain activity, using vision as an example: models that perform image recognition better also tend to predict monkey visual responses better. He frames a brain model around four choices: its architecture, objective, training data, and learning method. For comparison, he says models should be judged against the similarity between brains, since biological responses vary even under matching conditions.

3. 19:40

   Perception and prediction across species

   For vision, Nayebi reports that contrastive self-supervised models match mouse visual cortex better than categorization-trained models, and transfer more effectively to navigation tasks. In tactile sensing, his group builds models from simulated whisker forces and torques; recurrent encoder models perform well on recognition and brain-response matching. He also finds that video models trained to predict future latent states better match macaque responses and human judgments than models trained on static images alone.

4. 41:34

   Planning and internally driven behavior

   Nayebi explains that networks trained for path integration can develop a range of activity patterns resembling those found in hippocampal recordings. He then describes a zebrafish-inspired agent whose intrinsic reward tracks progress in its learned world model. In the reported evaluation, the agent swims more naturally and its internal activity nearly explains all measured whole-brain response variance, while also matching fish behavior.

5. 51:10

   Safety and the cost of alignment

   Nayebi argues that specifying too many goals can make AI alignment intractable, and highlights corrigibility, or preserving human ability to intervene, as a small set of goals worth considering. He also presents an economic analysis suggesting AI would need roughly five to seven times current automation efficiency to meet a threshold for funding broad benefits such as basic income. He says alignment costs could raise that threshold, making cheaper operation relevant to deployment.