Jeff hawkins

Jeff Hawkins (June 1, 1957, Long Island, New York) is a computer engineer, inventor of the Palm Pilot and the Treo smartphone, founder of the Palm and Handspring companies. He has also worked in the field of neuroscience and is president of the Redwood Neuroscience Institute, founded by him in 2002. Together with Donna Dubinsky and Dileep George, he founded the company Numenta, with the aim of developing a new type of memory based in the functioning of the human brain.

Memory-prediction framework

According to Hawkins in his book On Intelligence, the brain works on the basis of memorization and pattern recognition, so that the task performed by the brain (or at least the part called the cortex) is the prediction, it is what the author calls the memory-prediction framework. According to the author, "the role of any region of the cortex is to find out what relationship there is between its inputs, memorize it, and use that memory to predict how the inputs will behave in the future."

According to the author, the algorithms used by the brain are general enough so that it can be recognized, imagined, created and learned: «Every step from raw information to abstract idea is based on the same algorithm. It's the only computation the cortex knows how to do, but it's so versatile that it can explain all the amazing properties of the mind."

Hawkins believes his theory could become the unifying standard. He claims that he had a revelation in the mid-1980s while walking to his office door in his Mountain View, California, home. At that precise moment, he wondered: What would happen if the door changed? Why would I know? And he answered himself: I would notice the change because my brain was already predicting what the door would look like. "Brains are not like computers, where you input symbols and then something different comes out," says Hawkins. "As far as the brain is concerned, it's all patterns of information."

Hawkins' theory of pattern-completing memory is a variant of autoassociative memory, that is, the belief that memory drives itself.

Theory development

The cortex is a thin sheet about the size of a dinner napkin, and it's as thick as six business cards stacked on top of each other, that's just 6mm thick. This thin layer is very important because it was defined years ago as the place where intelligence is located. Language, orientation, music, art and everything that identifies humans is also located in this place. This is where all the things that are proper to a high level of reasoning are believed to reside. The key to understanding intelligence is understanding the cortex.

So if you want to build intelligent machines, you shouldn't build them on the old brain. Its construction should be based on the rational part of human experience. Fortunately, the cortex is an extremely uniform structure.

Our brains work on a completely different principle than computers. This does not mean that a brain cannot be emulated on a computer, but one must first understand precisely what the brain is doing. The flaws in artificial intelligence come from the idea that you have several inputs and then you get a series of outputs. That is, you feed the system with information and the results or outputs determine the success of the system. Jobs in artificial intelligence have no concept of what thought or perception is or what it means to understand something. The biggest conceptual difference between computers and brains is the ability to predict. Brains have their inputs and outputs: they are the internal predictive mechanisms. It's basically saying, “Hey, before I act, before I do anything, I need to check. Do I really understand what I am going to do? Success is not in whether you produce the right behavior; success depends on whether you really compute (with the future in mind) what is going to happen next.

There are some things that are still not understood, but for the most part this developed theory fits the bill perfectly. Its completion is not many years in the future. Jeff Hawkins thinks that starting in 2011, products based on these ideas will begin to be marketed. The study of brain function is important because it represents the future of computers.

The Numenta software that Hawkins develops has nothing to do with the field of artificial intelligence, although its purpose is the same. What Hawkins has in mind is more elegant. AI makes use of the brute force of a computer to make the computer appear intelligent through its behavior. This idea comes from the preconceived idea of Alan Turing, who implemented his Turing test to find out, through behavior, if a machine is intelligent or not. When IBM's Deep Blue machine beat Gary Kasparov a decade ago, it wasn't because it was smarter than him, but simply because it was faster. Even today computers have no intuition. They have problems with language recognition and are not good at dealing with ambiguous information. Humans don't have those problems. We are intelligent and those machines are not.

Numenta's software takes a totally different approach. Computers with this software are not programmed like regular computers. Numenta's software makes these machines learn from observation, just as a baby or child would when observing the world around them.

Numenta is developing a new computer memory system that they say can remember patterns of the world presented to it by its senses and use these patterns, just as humans do, to make analogies and then draw conclusions. If this work works out, you could anticipate some impressive business opportunities. The variety of applications could range from the mundane, helping airport security with scanning or interpreting images or x-rays to tornado forecasting, stock market forecasting, improving factor efficiency in a factory or process., development of physical laws, etc. Jeff Hawkins says it has to work because that's how the brain works.

Jeff Hawkins sees the neocortex as a memory system that constantly adapts and rearranges its connections to create a pattern. "Every time you have a new experience, every moment of your life is like a pattern or previous experience plus a prediction of what will happen next." That is, according to Jeff, the present or the consciousness of this is the sum of the past or a past experience plus a prediction of the future. Different regions of the neocortex do much the same thing, according to Hawkins. These regions store spatial and temporal patterns that can represent things like language, music, and vision. From Hawkins's point of view, all the senses of the human being work in the same way: the information arrives from the outside world in the form of patterns or neurons sending signals, these signals sent by the sensory neurons are stored in the form of [engram]s from other neurons, and each piece of new information is paired with a pre-existing sequence of patterns. In other words: there is a general algorithm that recognizes and interprets all this set of patterns in the brain.

Many neuroscientists think that Hawkins' idea is a gross oversimplification. Hawkins also admits that he hasn't done any neuroscientific research yet. The theory sounds a bit sparse on specific details and there is no experimental evidence either.

Hawkins was stuck on the idea of how to turn his idea into something a computer could understand until he met George Dileep in 2003. At the time, George was just a graduate student at Stanford. George was also an electrical engineer who had gone back to college to learn about the brain. After several student neuroscientists did not take him seriously for suggesting that it would be possible to build a model of the neocortex, George was drawn to Hawkins's ideas. George did something that surprised even Hawkins: he took Hawkins's ideas and informally turned them into algorithms.

Algorithms were informal and dirty at first. Even most of the neuroscientists at the Hawkins Institute criticized George for being too simplistic. But Hawkins saw a kindred spirit. "This is great; George is taking my ideas seriously," Hawkins said. George continued to refine the math and the software got better and better. Later, they set up a typical visual recognition problem and entered images into the system.

While a simple child can identify and draw a cat or dog the first time they see it, computers by contrast find this simple task almost impossible. George did a couple of line drawings. He drew a dog, a cat, a sheep, and a helicopter. He uploaded the algorithm and trained the computer by digitally animating the drawings. When the computer was trained it was able to recognize objects. George then began to introduce variations into the drawings that the machine had never seen before.

Slowly the machine began to place the images in the correct categories and even to estimate the probability of success or certainty with which it did so. Many machines would find this problem unsolvable.

Numenta was founded very soon after this test or demonstration to investors. The company is developing what Hawkins calls a "hierarchical temporary memory system." The software runs on Linux computers, but over time it could be implemented on FPGA hardware. The system mimics the structure of the neocortex. Numenta's Hierarchical Temporal Memory (HTM) software has to learn from data the same way we do as children, explains Subutai Ahmad, Vice President of Numenta.

Posts

• A thousand brains, a new theory of intelligence, Jeff Hawkings, Ed. Basic Books, 2022.
• On intelligence, Jeff Hawkings and Sandra Blakeslee, Ed. Spass, 2005.
• Numenta - HTM Concepts.