Artificial Intelligence: Let's Think About Learning
summary
Argues that learning in biology is driven by homeostasis-as-error rather than backpropagated loss, that signals propagate neuron-to-neuron over milliseconds to seconds and training happens continuously as actions change sensors, and shows engineered systems using brain‑inspired networks with zero-initialized weights and structured random connectivity. Demonstration video (March) claims a real‑world quadruped learns to stand and balance without thousands of sims. Tools: neuroscience collaboration, chemical/electrical neuron models, on‑robot learning; result: robot succeeds in real world.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:50Defining homeostasis
- 7:53Single-celled organisms
- 12:53From single cells to multicellular life
- 17:36Neurons in signal transmission
- 21:05Homeostasis vs deep learning
- 26:31Biological vs artificial learning
- 31:55Biological inspiration for AI
- 33:02Q&A
related talks
32mHow Neuroscience Makes AI Intelligent
24mFrom Neural Networks to Digital Brains: The Next Leap in AI
51mThe Brain as an Efficient, Robust Adaptive Learner
29mDeep Learning, Neuroscience and the future of AI
49mML: Teaching machine to learn from its own mistakes (...)
07mRobots, houses & the IoT - Microsoft 's Brandon Rohrer on making things intelligent
09mLightning Talk: Intellectual Cognitive Development in Artificial Intelligence
25mBrain-Inspired Natural AI: Unlocking Intelligence