Interesting, but not sure not having a lot of memory is transformative.
It is. For this reason:
We're not just talking about "a lot of" memory, we're talking about FAST memory.
When you train an artificial neural network for machine vision, it takes hundreds (sometimes thousands) of repeated presentations of examples of the same image, before the network learns it.
It's kind of the same situation as training a ChatGPT, it takes thousands of nVidia GPU's and many days, and millions of training iterations.
Which is DIFFERENT from human memory. YOU can recognize a face after one presentation.
Google regularly runs machine learning contests, and one of the standard benchmarks is the "handwritten digit training set". You have digits from 0-9, hand written by hundreds of people. The benchmark measures how long it takes your network to achieve 99% recognition accuracy.
The digits vary along many axes, some slant left, some slant right, some are scribbled and some look almost like a typeface. Think about discriminating a 3 from an 8, under those conditions.
So if you have a robot on an assembly line, and suddenly you change a component, how long does it take the robot to learn what the new component looks like, and then how long does it take the robot to learn how to pick it up and how to install it?
What if the component can only go in one way, like a diode, and the only identifying mark is a tiny + sign on one end, and there is also a part number and a manufacturing number obscuring the identifier?
These are common real world situations, and they're reasons why many manufacturers have chosen to forego robotics.
