I wanted to share my latest write-up exploring an interesting convergence between TBT and Jean Piaget’s observations on infant cognitive development. As someone trained as a developmental psychologist, I’ve always been interested in how intelligence emerges and matures through interaction with the environment and hypothesis stesting, making this parallel particularly intriguing to me. In this piece, I’ve delved deeper into the overlap between TBT and Piaget’s insights on how infants construct knowledge through direct sensorimotor exploration. My aim is to let everyone know about a powerful analogy from developmental psychology that lends more credence to TBT.
I was not familiar with this approach to teaching, thanks for pointing it out. I like that it places a focus on sensorimotor learning, as Piaget or not, it’s a necessary component of learning in humans. I think it’s an interesting idea to think about how individual children might learn differently based on their sensorimotor preferences/strengths and that conceptual development is tied to movement in space. That works for me.
Really nice blog post, I couldn’t have said it better myself!
Some of Piaget’s theories actually strongly inspired the work I did during my PhD, trying to show how learning by interacting with the world is so important and how it is very different from how most current AI learns. Here is an image from my thesis (that I showed in many of my presentations, trying to motivate my experiments)
Back then, I was working with deep reinforcement learning, which let me focus on the sensorimotor aspect but had some fundamental drawbacks that the TBT nicely addresses. Some big ones are reference frames and local learning. DeepRL can be super unstable and hard to train + requires tons of data (even though I got it to work unsupervised, purely curiosity driven). Since it uses deep learning, it incorporates the i.i.d. assumption (that data samples are independent and identically distributed), which fundamentally doesn’t hold when learning from continuous interaction with the world. Learning structured models (using reference frames) that can use local learning rules and don’t suffer from catastrophic forgetting solves all those challenges. So I was very happy to join Numenta and find out that the mechanisms of the brain described in the TBT are a perfect fit for rapid, continuous, and robust learning through largely unsupervised interaction with the world.
Thanks for the insight! When I was in graduate school in the late 90s, Piaget was mostly taught so you could learn all the ways new research showed he was wrong. But I think some of his ideas continue to be relevant, including looking at the notion of the development of reference frames as part of Piaget’s egocentric to allocentric representations. Newborns have plenty of egocentric data to start working on before they can make sense of the external world - developing object permanence, Theory of Mind, and general perspective taking takes the development of allocentric reference frames.
