AI Object Recognition

Over the holidays I found myself reading great amounts of non-fiction. Among the books I read was ‘Physics of the Impossible’ by Michio Kaku. From that book I discovered that scientists are having trouble designing a robot which can manouver around a room due to difficulties in recognizing complex shapes. This is because the computer looks at the room as a series of straight lines and has great trouble with curves, or anything complex. An example the book gave of how slow progress is would be to compare the number of neurons required for a fly to do incredible loop-the-loops around a room to the huge amount of processing it takes for a robot move around on the floor of a furnished room. Clearly, although the fly processes far fewer bits of information, it is better suited for movement.

Why is this? Here’s my reasoning: the fly recognizes far less than scientists are assuming. Let’s move up the food chain a bit for an example. Birds will avoid eating moths with eyespots on their wings.

Consider that. You or I will look at a moth and will know that it is a moth, regardless of what spots it might have on its wings. If it is camouflaged into a tree or something else, we may not see it, but that is a seperate issue. Yet a bird will see the spots and fly away, avoiding the moth entirely. Why is this? The bird sees the spots and believes that the ‘eyes’ belong to a predator. Thus they avoid the moth. This is caused by a very simple object identification system.

Back to the fly. A fly will land on almost anything, even people, unless it is moving fast. It avoids anything which is moving fast- e.i. that could squash it. They’ve even been known to land on waiting frogs. They also don’t fly straight into any object, they slow and land on them, or go around.

This should be applicable to robot AI. It would be simple enough to scan the area to determine where obstacles are. There is no need to determine what they are, or even their exact shape. From there, as needed, the AI could be programmed to recognize certain things, such as chairs or coffee mugs, and be given instructions other than ‘avoid.’

Thoughts? Hopefully this is an example of where cross-disciplinary thinking can get you, but there may be a flaw in my reasoning.


3 thoughts on “AI Object Recognition

  1. To identify random patterns we need a good application of Fuzzy logic and Sel-Learning state machine which coincide a bit with your reason.

  2. supposedly Darwin witnessed ‘object learning’ in the Galapagos. Galapagos wrens would land fearlessly next to humans. Darwin saw a young lad picking up those that landed near him, smashed their little heads on a rock, and threw them aside. yet, still, the wrens landed nearby. Darwin theorized that it takes generation upon generation for a species with some thought capability (flies aside) to recognize and learn to react to threats.

    so, the ‘simple object identification’ may not be so simplistic. with AI, you have the advantage of, by virtue of speed of calculation, replicating hundreds of generations of learning quickly. should work, but maybe not as quickly as you expect…

  3. Pingback: Pages tagged "lucid"

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s