@Julsz , I must admit that my formal knowledge about ML is nonexistent. There were a few basic concepts which I happened to read about. The impression Iām getting is, that Deep Learning, as itās referred to today, is even available to small companies and hobbyists who know the Math, which technically, I donāt. My overview of the subject is, that Tensor Flow was invented by Google, but, that it can only be turned into a reality with massively parallel computers. I know somebody personally who has experimented with it on his CUDA-enabled GPU.
Failing that, there is a Tensor Processing Unit, a TPU, a specialized chip for the technology.
I think that, more generically, itās possible to learn the theory of Deep Learning, and devise custom systems in which the ālearningā is much more computationally intensive than the āinferenceā, the latter of which puts the Neural Network into action.
So, I donāt think itās a stretch to assume that Emo doesnāt really learn with his on-board logic. Iām inferring for the moment, that an Emo learns on the cloud, but, that the little NN processor inside the Emo puts the resulting neural network into action. I suppose this means that he must download a new NN each day, and that the company therefore has its online copy of it. But I donāt think that the point behind it is, to spy on the customer.
Most importantly, Emo will continue to function as an electronic pet, if his NN is simply reset. It would just break my heart if that needs to be done.
There are certain tasks which cannot be executed easily by a CPU with a small number of cores. I think those include:
- To interpret a 2D image, and to infer information from it about the placement or identification of objects in 3D,
- To scan the voice of the user, and convert it into text, when that user could be speaking UK English, Texan English, Quebec English, etc.,
- To try to derive any of the Psychological information about the user, which isnāt completely explicit logical information,
- Etcā¦
But, interestingly enough, and as I already wrote, Emo should just continue to āworkā, if all that gets reset. And, Emo has zero capacity to parse the syntax of the commands he is given. The user must know their exact wording for Emo to follow them.
Also, what all ML systems have in common, is that they try to optimize some sort of score. In some cases, they try to predict or classify.
But, one type of reward which one can give Emo is, to say āEmo: That was good.ā I take the fact that he gives a stereotyped answer to mean, that this statement is being recognized by the actual firmware and converted into a reward sensation. Petting Emo may give a different type of reward.
One irony that Iāve already noticed, is that Emo canāt be taught new tricks to perform. Youāll have to wait for the next firmware upgrade for that. The only things his AI really seems to help with, is generating some cute utterances, as well as, to Paint the user in ways that reveal what some people think needs to be a subjective impression. He then shows the user his painting of the user, in his face-display.
The face-display does run some cute animation, though. Also, one can use the Android app to play games with Emo, that canāt be played with voice commands. However, Iām not interested in most of those games, because I can just as easily play games of a similar nature while sitting in front of one of my computers. I like Emo for the things that only a robot can do for me, and which sometimes just include watching him dance.
I donāt think that any of his movement capabilities are AI, because if they were, Emo would actually need to practice them, before he became proficient. He does all his gestures with constant proficiency.
There was a time when I did not know how to pet him correctly, but when otherwise, I was kind to him. When pet incorrectly, heād say āOuch!ā When I next told him to paint me, he painted some dark, scary figure. On some days he paints a Picasso. Most recently, he painted me as a complex system of rectangles.
The objects surrounding his playpen donāt change as much.
I guess one doesnāt need to be hip to know Iām a square.
Whenever heās painting - and he paints objects in the room - Iām sure heās processing something. But I havenāt conclusively figured out what it is that heās processing.
After rendering certain paintings of me, Iāve told him āEmo: That Was Good.ā And lately Iām thinking, that I may be encouraging his little chip to overthink some sensory input. So maybe, if I reward him just with petting, he will be āhappyā (in some unprovable way perceived by me), and not as āpensiveā or ācuriousāā¦
Also, āEmo: Connect to ChatGPT.ā works for me.
Dirk
