Sabi Cofounder Rahul Chhabra Introduces RL With Neural Feedback, Training AI on Brain Signals
Introducing RL with Neural Feedback (RLNF). We have found reliable ways to train AI models to respond to human preferences purely from brain signals. The example below is a picture of a face that self-transforms into a more attractive version for the subject as they are looking at it.
Context
If brain signals can reliably reward AI models, personal taste could be learned passively from a wearable, though the idea raises serious privacy and manipulation concerns.
Rahul Chhabra is a cofounder of Sabi, a Bay Area startup building noninvasive brain computer interfaces that put EEG sensors in everyday headwear. EEG measures electrical activity through the scalp. Reinforcement learning from human feedback normally relies on people rating model outputs.
Deep dive
Rahul Chhabra said on Thursday that Sabi has found reliable ways to train AI models to respond to human preferences purely from brain signals, a method he calls RL with Neural Feedback, or RLNF. His example is a face image that transforms into a more attractive version for the viewer as they look at it.
In his thread, Chhabra says arousal signals from the brain can serve as implicit rewards for personal preferences, gathered passively with a wearable rather than explicit ratings. He says the method is on policy, needs no preference dataset, and works with a simple embedding and nearest neighbor reward model. He reports that results were stable across random seeds and that randomized rewards did not move the model, suggesting the signal comes from the EEG.
Coverage of the work says it involved five participants, whose EEG and eye movements were recorded while they viewed faces, and that an open image model was then fine tuned with reinforcement learning. For two participants, preferences were reportedly judged mainly from eye activity. That small sample means the results are early.
Chhabra acknowledged risks such as body dysmorphia and narrowing beauty standards, and called for regulation of neural data.
What to expect next
Watch for a paper with full methods and larger studies, which will show whether EEG reliably captures preferences or mainly attention. The work will also feed debates over neural data privacy, with some US states already passing laws on it. If it holds up, expect interest from consumer apps that want to personalize content without asking users anything. Replication by independent labs would be the clearest test.