Email
Password
Remember meForgot password?
    Log in with Twitter

article imageOp-Ed: AI learns to identify objects like humans

By Paul Wallis     Dec 24, 2018 in Technology
Stanford - Artificial intelligence, machine learning, and nearly a generation of work are paying off as AI learns to visualise and identify objects like humans. This is a huge, and very necessary, step to AI evolution.
UCLA Samueli School of Engineering and Stanford research has created a computer program with the rather forgettable name of “computer vision” which can independently identify partially seen objects. This goes beyond the usual program/task scenario which limits normal computer behaviours.
This is a super-hot area of AI research, and it’s on of the most important of all AI functions, in terms of making AI operational in the physical world. The sheer scale of research in this field is indicative; just about every major research organization is putting a lot of work in to this field, and this new achievement is a big, very important breakthrough.
Google, in fact, just released information about its custom optical character recognition (OCR) engine AI lens which can identify a billion products, using different tech. You can see where this is going; use visual equipment hooked up to AI, and you basically have an instant reference to anything you’re scanning. That also just so happens to be one of the primary needs for AI in any role.
Critically, the new UCLA/Stanford approach can recognize an object by seeing just part of it. This is a classic human trait. A human brain will see something and instantly recognize it, regardless of visual obstacles or blocked lines of sight. Humans can also predict the position of the rest of the object, based on these partial sights of objects.
The new tech creates an “assembly” methodology so computers can do these things:
1. Partial images, aka “chunks” are read by the computer.
2. The computer learns how to assemble these chunks accurately to identify the object.
3. It assesses other objects around the target object to aid description and identification of the object.
To explain:
* The average environment is a clutter of images. To accurately read the environment, identify specific objects, and do so efficiently is very difficult for computer programs.
*The “read” function, which is how computers execute programs step by step, needs to be extremely efficient in artificial intelligence. The clunky, rather ponderous process of write code/ read program/ list/execute/ on/off, etc. isn’t good enough for AI to act independently. If AI can simply learn for itself, and have that memory ready when required like humans, the entire process becomes much smoother.
Now consider:
Most people do not, in fact, recognize absolutely everything they see instantly, in any environment. There’s always something which needs to be assessed. Even in your own home, you can expect “What’s that doing there?” or “Why is that here, not there?” or “Where’s….?” and the ever popular “What in the name of guided salad tongs is that?” are pretty typical.
You have to look, and make some conscious decisions based on your knowledge, or lack of knowledge, even in a familiar environment. So for a computer to be able to recognize, and equally importantly, recognize things it can’t identify, is a gigantic step.
Consider also your “visual vocabulary”. That’s a door, that’s a wall, that’s a car…. It’s a lot of information which is processed so fast you don’t really have to think about it, simply be aware of it, the location of objects, risks, etc.
Just navigating across a room, knowing what you’re doing, (and in some cases even why) is based on a literal tonnage of information, in terms of computer coding and processing. That’s why this research is so fundamental, and such a vast leap in to the future of AI in just about all possible configurations.
Remember that AI will come in millions, eventually billions, of different forms. Forget the Internet of Things, this is the AI of Things, and it’s likely to be absolutely gigantic in scope and range of applications.
How AI functions, how well it functions, is predicated on how it interacts with the world about it, and it’s looking like that will be a lot more efficient in future. Samueli School of Engineering and Stanford have just created a major asset for AI in the most functional possible sense.
This opinion article was written by an independent writer. The opinions and views expressed herein are those of the author and are not necessarily intended to reflect those of DigitalJournal.com
More about UCLA Samueli School of Engineering, Artificial intelligence, AI partial image recognition, AI of Things, Google AI guided lens
More news from
Latest News
Top News