You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
It seems in the AI game, it's not only power, but the time/cost of putting the data in a form that a particular AI can digest the data and Watson is loosing in this arena. How much data needs to be processed before a particular AI can use it will become a key metric. Making sense of diverse data streams will be as critical as the power of the AI itself. I feel there are massive short cuts that AI and data bases can use and strategy/architecture are going to become as key as the AI itself. Sometimes it takes an outsider to bring a whole different take on a subject and we'll see more of this in AI/ML in how it gathers data, which data it gathers and how it's processed. I have already seen one short cut someone used that saved tens of billions and years in time by placing strategy above brute force. As with anything, strategy if faster and cheaper than brute force. Once in a great while, I see an opportunities for outsiders and this is one of those times. AI is just another tool, although very powerful, that we have to learn how to use and in this aspect we are in the very early stages. Not only will AI itself become a large market for semis, but the tools for data acquisition will offer even larger markets for semis and semi/nanotechnologies.
Feature selection is an NP-hard problem, and complex applications are likely to have a large number of features. So I don't really see a way around data preprocessing and manual feature selection and supervision by domain experts.
It's been done, with someone eliminating about 80% of the normal research process and the problem itself. I'll start digging as soon as time permits and I don't want to reveal processes I plan on keeping proprietary.
"AI" is like "transport", it's MANY different things with many different characteristics.
Saying that Watson competes with GPU-based image recognition is like saying that FedEx competes with Maersk. Well, yes they do in some vague theoretical sense, but the set of things that Maersk transports is very different from the set of things that FedEx transports.
Right now deep learning has all the hype, and that's fine --- it's useful technology that can do many useful things. But "many things" is not the same as "all things", and once the hype moves on, Watson will still be there, plugging away at a DIFFERENT set of problem that also need to be solved in order to make computers ever more useful to all of us.
I've been keeping buried in studying AI/ML structures and applications and see a field in its very infancy with what may be one of the fastest and longest growth curves of any change in business or industry. This will be not only a gold rush, but a revolution. I'm not engaging in the war, but investing in the arms merchants.
