School: LAS CRUCES HIGH
Area of Science: Computers
Interim: We started this project because one of our team members read a book , over the summer, called On Intelligence written by Jeff Hawkins. The book describes how the brain works, but it isn't a typical book on the brain. In this book, Hawkins proposed a new theory of how the brain works. It proposed a new way of looking at how the brain uses memory and feedback. After reading his book the team member was able to form a new idea about how to create artificial intelligence.
When Jeff Hawkins was as little boy he became fascinated with brains. He wanted to figure out how brains worked, but he couldn't find anything that really explained it. A few months after Jeff Hawkins graduated from Cornell University in June, 1979 with a degree in electrical engineering, he read a magazine called Scientific American, that was dedicated totally to the brain. It made him remember his childhood and his fascination with brains.
When he was working for Intel, he wrote a letter Intel's chairman Gordon Moore. In his letter he explained that he wanted to start a group dedicated to trying to find out how the brain works, but his proposal was rejected. So because his letter was rejected, he sent his proposal to M.I.T. In 1982, he moved to Silicon Valley, California, and started a company called Grid System. Then he went to the University of California and Berkeley in 1986. In 1987 he created a hand writing recognition program called Graffiti. He read a book called Astonishing Hypothesis written by Francis Crick. Crick's hypothesis was that the underlying structure of the brain is the same in every region, that all regions of the brain perform the same function except with different input. He currently holds the positions of Chief Technical Officer at Palm Inc. and co-founder of Numenta.
The field of Artificial Intelligence (A.I.) began in the 1950's. People used to think that to be intelligent; you had to manipulate abstract symbols. This was what people observed when people acted intelligent. Another idea from early A.I. pioneers was that behavior defined intelligence. During World War II a man named Allen Turing came up with the Turing Test. In this test if a computer could converse with a person on a computer and convince that person it was a human, it would pass the A.I. test.
Early computer programs were based on expert systems. An expert system was a series of rules and facts that work together to give an answer to limited set of questions. In 1980, a man called John Searle wanted to show that computers could not be intelligent, so he came up with an imaginary situation called the Chinese room. In the room is a man that has a rulebook. A Chinese man writes a question in Chinese and gives it to the man through a slot in the door. The man looks in the rulebook and writes the answer back in Chinese. According to Searle, neither the book nor the book understands Chinese, the man was just manipulating meaningless symbols.
During the 80's, computer A.I. was not living up to its expectations. But this was the time when neural networks started to become popular. Neural networks loosely simulated how neurons in the brain work. A neural network is trained to give certain answers for a given input. It does this by connecting the input with what the correct answer should be. One form of neural networks is called auto-associative network. An auto-associative network works by giving the correct output when only part of the input is given.
The part of the brain where intelligence comes from is called the cortex. The cortex has six layers, each two millimeters thick. There are approximately thirty billion neurons in the brain. The brain receives both spatial and temporal patterns. Spatial patterns comes from multiple sensor input at the same time. Temporal patterns happens over time. The brain follows a rule called the hundred step rule. Every second input that the brain receives can only go through one hundred brain cells. In order to accomplish a task, like recognize a cat, only one hundred steps would be involved. Even a super computer cannot do this, because a super computer is fundamentally different from the way a brain works. The way the brain solves a problem is comparable to asking someone what is five plus five. They would say "ten" because they remember it. But if you asked a five year old to find out the answer he would count on his fingers which is how a computer works. The brain also remembers things in a way called invariant representations. That means if you change the angle of a picture the brain will still be able to recognize it. If you move a picture over five pixels on a computer it would not be able to recognize it, but the brain can recognize it even if it is far away, close up or even upside down. The brain doesn't remember everything that it sees. It only remembers the important things and not the details. The brain can only remember things in the sequence in which it saw or heard it. You can remember all of the words of a song forwards but cannot remember them backwards.
The brain uses memory to predict what will happen next. This is how it can only remember the important things. If something unexpected happens then it will remember it but if it can predict what will happen and that thing happens then it will forget. The brain is organized in a hierarchy. Different parts of the brain report to other parts of the brain. They report to each other by feed forward input and feedback. There are at least ten times more feedback connections then there are feed forward connections. In the lower levels of the visual cortex, where the input comes from, only very basic things are recognized, such as lines. In higher regions, objects are recognized.
What our project is how the brain uses memory to predict sequences of events. What we have done to accomplish this, is to make a simple model of four-hundred brain cells, in twenty layers to represent the hierarchy. Each brain cell will send feedback to the brain cells below them. The brain cells will use this feedback to decide whether or not they were able to predict it correctly. If they did not predict it correctly then they will send it up the hierarchy until one of them does.
Sponsoring Teacher: Gregory Marez