Happy Sunday and welcome to InsideAI weekend. I'm Rob May, Founder/CTO at Dianthus. I'm also a very active angel investor so if you have an early stage AI company that is raising, please reach out.
AI is in an interesting place because the performance of large AI systems is moving forward at a rapid pace, and AI products are showing up everywhere. Despite all of this, in a lot of ways we still seem pretty far away from a truly intelligent machine, and AI training remains expensive and slow compare to what we need it to be.
There are a few emerging trends around where AI research should go next, and the one I find most interesting is the return to neuroscience. There has been an ongoing debate for decades in the AI community about how much we need to understand the human brain in order to generate an artificial one. Some people believe building a chip that with elements that work like neurons and then hooking them together will get us to true AI. Others believe intelligence is an algorithm in the brain that can run on many different types of compute and we just need to figure out what it is. There is also the counterpoint that says we don't really need to understand the brain. After all, we don't fly the way birds do - we invested airplanes. Maybe intelligence can be done in many different ways.
I bring this up because of a recent article I read about what AI can learn from babies. Here's the relevant and interesting part:
In their article "Lessons from infant learning for unsupervised machine learning," Dr Lorijn Zaadnoordijk and Professor Rhodri Cusack, from the Trinity College Institute of Neuroscience, and Dr Tarek R. Besold from TU Eindhoven, the Netherlands, argue that better ways to learn from unstructured data are needed. For the first time, they make concrete proposals about what particular insights from infant learning can be fruitfully applied in machine learning and how exactly to apply these learnings.
Machines, they say, will need in-built preferences to shape their learning from the beginning. They will need to learn from richer datasets that capture how the world is looking, sounding, smelling, tasting and feeling. And, like infants, they will need to have a developmental trajectory, where experiences and networks change as they "grow up."
The idea that we would seed machines with some basic preferences before their learning began is an interesting one, and one at the center of a long debate in cognitive linguistics about how much of human language is based on innate structures.
I don't have a strong opinion on whether this will work, but I do think to get to GAI we are going to need many different approaches. No one knows where the final breakthrough will come from. If you are working on ideas like this, I'd love to hear from you.
Thanks for reading.
@robmay
Comments
Post a Comment