The Sweet Lesson of Neuroscience
9 days ago
- #AI alignment
- #neuroscience
- #brain-inspired AI
- Scientists initially looked to the brain for AI inspiration, but now AI research may inform brain understanding.
- Deep learning's early days were heavily influenced by neuroscience concepts like hippocampal replay and dopamine-based learning.
- The 'bitter lesson' in AI highlighted that scaling general methods with data and compute outperforms brain-inspired designs.
- Modern AI focuses on architectures, learning rules, and training signals, with the latter being the least explored.
- Steve Byrnes proposes the brain consists of a learning subsystem (neocortex, hippocampus) and a steering subsystem (hypothalamus, brainstem).
- The steering subsystem provides innate, evolutionarily hardwired reward signals that guide the learning subsystem.
- Thought Assessors bridge learned concepts in the learning subsystem to innate signals in the steering subsystem.
- Laughter and social instincts are examples of how the steering subsystem shapes behavior through intermediate rewards.
- Byrnes' theory suggests AI alignment could benefit from understanding how the brain aligns learning with innate goals.
- The hypothalamus's complexity supports the idea of a sophisticated steering subsystem with species-specific behaviors.
- Songbirds demonstrate how innate instincts and learned behaviors interact through reward signals.
- Byrnes' framework could inform AI alignment by mimicking the brain's internal teaching mechanisms.
- Psychiatric conditions may stem from misaligned steering subsystems, highlighting the importance of understanding these circuits.
- Future AI might adopt brain-like architectures with separate learning and steering subsystems for alignment.
- The 'sweet lesson' of neuroscience is that brains are not just learners but also architectures of internal teachers.