In this chapter it’s all about developing concepts and recognise the patterns through experiences. Something like Aristotle's Empirical evidence.
Aristotle believed that knowledge comes from the senses.
So it goes on like this:
- A vehicle set out to experience the world, The chapter 7 kind with
mneomotrix. It identifies aBounded universeby noticing the arrangement of the objects at the edge of it’s environment. - Kinda like it tosses the coin and sees/discovers it has two face. Hence, forms the concept of
Two-faced coin.
Also he(Braitenberg) goes on to talk about the process of Forming ideas has Strength and also Limitations. How?
- Overgeneralisation: It can obscure individual details. That also leads to loss in flexibility in classification and decision-making as a whole. Or we can coin this term as
Overfitting Probolem. It does well on the training data set. ButEhhin the test dataset. - Closed chain of objects: A vehicle focusing too much on the concept of a closed chain objets loses track of the serial order of the individual elements.
He went on to explain these two terms with several example, like: Garden explorer vehicle that sees the flower and develops concept. And also combines this concept development with evolution theory. And with sufficient time and resource, a vehicle is likely to develop concept.
In the lecture, professor talked about concept based learning. How a transformer is unlikely to get general intelligences because, it doesn’t have any concept. But now people are using LLM with Agent. That kind of guides the answer of the model. I wouldn’t say, this gives the LLM concept. But, it does seem like a good fit for reward function right? Through iterative process of this Agentic model, we can achieve some sort of concept. But definitely it is not that much of efficient.
It’s like firing a canon to kill a fly. And the fly is incredibly fast at evading. Why? because, he has just about 100K neurone. Lesser neurone means lesser time to process the data. And that is why it makes such blazing fast decision.
So, perhaps the way to solve intelligent machine is to go modular. With simpler solution and with concept development from experience.
13.01.25 -Abrar Matrikel-Nr: 5012300