I will work on creating AI systems for biomedical applications. Longer-term projects will include trying to understand deep learning for more reliable applications, and brainstorming new ideas for AI inspired by biology.

Hope to collaborate with peers with common interests. I’m also hiring PhD students. Please feel free to DM me.

3. Long term: create next generation of AI with world models and intrinsic goal-chasing ability, leveraging our current understandings in biology instead of computer science (definitely not language models). I call it Science-Inspired AI.
🧵

2. Mid-term: debunk the AI hype by mechanistically understanding current deep-learning systems. I call it Science for AI, or "aronscience", meaning artificial neuroscience.
🧵

Coming from biomedical engineering, I'm transitioning into computational science and AI in general. My visions are:

1. Short term: create AI systems (mechanistic and/or data-driven) for biomedical engineering and informatics. I would call it AI for human, not the hyped "AI for Science".
🧵

We showed on a head phantom that individually optimized montage gives stronger TI stimulation than literature and random montages. 🧵END

The non-linear physics of TI gives rise to its higher sensitivity to random changes in the stimulation montages. Interestingly, at shallow brain locations, HD-TES is more sensitive than 2-pair TI

This is even more obvious in 2-pair TI: using a same montage across individuals significantly drops the focality

Individually optimized 2-array TI gives more focal stimulation than using a same montage optimized from a standard head template

2-array TI gives more focal stimulation than conventional HD-TES, at most targets we studied

There is great variability in the optimal montages targeting a same brain location, for all the modalities we studied: conventional high-definition TES, TI using 2 arrays of electrodes, and the popular TI using 2 pairs of electrodes