Building thinking machines. Follow @LaschowskiLab for updates.

Yes, but alignment may require reward inference before reward optimization. Then again, human and AI values may never be fully defined via math or natural language.

Invited talk at @MSFTResearch on building computational models of intelligence. Thanks for having me. @UofT @LaschowskiLab

Proud to lead the @UofT summer research program in computer science for students from Ukraine. @UofTCompSci @VectorInst

Our mission: reverse-engineer intelligence. Follow Machine Intelligence Lab for updates. @UofT @LaschowskiLab

Come learn about our research on reverse-engineering intelligence and the physics of learning and neural computation. @LaschowskiLab

Our new study on bidirectional representational alignment between biological and artificial neural networks. @UofT @LaschowskiLab arxiv.org/abs/2608.18244

Are you a @UofT student in computer science, math, physics, or engineering? Want research experience? We’re recruiting students to explore theory and algorithms for inverse reinforcement learning. Email me your resume and transcript.

I started in neuroscience. Then spent several years building intelligent machines. Now I study general principles of learning and intelligence. @UofT @LaschowskiLab

Proud professor. The next-generation of machine learning researchers from Ukraine. @UofT @LaschowskiLab

Full-circle. When your students go on to supervise their own students. Congrats Dima on the @eccvconf papers. Proud of you.

Introducing our framework for interpreting reward functions recovered by inverse reinforcement learning. biorxiv.org/content/10.64898… @UofT @LaschowskiLab

Applications are now open for the @VectorInst Distinguished Postdoctoral Fellowships. Come join our world-class machine learning research community. Apply by August 31: vectorinstitute.ai/research-…

Thanks for your feedback. We use "domain" in the machine learning sense, where different users, tasks, sessions, and devices correspond to different data distributions. In that sense, we view cross-user transfer as a special case of cross-domain adaptation.

We see these as complementary. Scaling emphasizes data quantity, whereas source-domain adaptation emphasizes data selection. Both are important for transfer performance and data efficiency in neural decoding.

Interesting. Is scaling alone sufficient for cross-domain decoding? Not all source domains are equally informative, and indiscriminate scaling can induce negative transfer. Selective source-domain adaptation can improve both transfer and data efficiency. biorxiv.org/content/10.1101/…

Come join us in Toronto. @UofT is hiring an Associate Professor/Professor and endowed Hinton Chair in Artificial Intelligence. jobs.utoronto.ca/job/Toronto…

New paper: We introduce a framework for studying how intermediate representation choice affects downstream learning and reconstruction, with an initial demonstration in neural speech decoding using diffusion latents. biorxiv.org/content/10.64898… @Comp_NeuroLab @UofT