If the proper study of mankind is man, then the proper study of Jeff Hawkins is the brain. It’s always fun to watch Jeff Hawkins’ brain work; it will be seen in action when he speaks at PC Forum
. It is Hawkins’ brain that produced the PalmPilot at Palm, the product that revived the handheld device market nearly ten years ago, and more recently Handspring’s Treo, the product that has defined the emerging smartphone space. As he says himself, “Sometimes I see things ahead of other people.”
In August 2002, Hawkins founded the Redwood Neuroscience Institute
, where he has worked in a small office over Kepler’s Books in Menlo Park, while still serving part-time as CTO of palmOne, Inc., the company resulting from the merger of Palm and Handspring, itself founded by Hawkins and longtime business partner after they had left Palm. Hawkins is now working on creating software from the puzzle he’s been teasing at for the last few years and indeed for the last few decades: How do we think? His conclusion (as outlined in his new book, On Intelligence, which will be distributed at PC Forum) is that it’s turtles all the way down. That is, intelligence is the ability to recognize and predict things by analogy; more intelligence is the ability to recognize and predict increasingly complex or abstract patterns, more quickly. Most people can understand, say, the relationship of a rectangle to a parallelogram or a square; Einstein could “understand” relativity in the same familiar way.
In practical terms, Hawkins spent years studying neuroscience – the brain and its structure of neurons and synapses. It’s pretty much the same material everywhere, yet the brain’s intricate structures comprise the five senses, language (several for some people) and huge amounts of learned knowledge. “Different parts of the brain do different things depending on what they are connected to,” says Hawkins, “whether it’s sensory organs or just other parts of the brain.” Much of the brain is organized in hierarchical structures; details at the bottom, generalizations at the top. An apple, for example, at the bottom layer of the hierarchy is color, taste, shape, and size, while at a higher level of the hierarchy is generalized as “apple,” and at a higher level yet is generalized as “fruit.” On the other hand, an apple can also be one of several “objects in a still life.” The hierarchies overlap.
The neuronal system that recognizes apples – and everything else – is basically a belief propagation network, says Hawkins. You see a red object in the right shape and size, and your brain starts to predict “apple.” If this prediction is supported by new data, such as the smell and feel of the apple, the brain’s hypothesis is confirmed, and the belief is sent back through the hierarchy that “apple” has been identified.
Working with a colleague at the Institute, Dileep George, Hawkins is figuring out how to apply the concepts of his brain theory to software, and how to represent the theories in working algorithms. There are substantial challenges, but substantial potential benefits and applications, if they can get it right.
One of the more intriguing questions raised by Hawkins’ work as applied to computers is how to think about the senses. As humans, we have five senses that send digital streams of input into our brains. A computer can have some of the same senses – such as vision – but could have more, such as radar or sonar or air pressure or temperature sensors or just various kinds of data. These inputs also could stream into a hierarchy. At the lowest level of the hierarchy, they would appear as fragments of data – mere measurements. But, with the right software and algorithms and context, these data streams could be recognized at the highest levels of the hierarchy as broad patterns that allow us to understand complex problems in a way we haven’t been able to do before – whether it’s weather, stock markets, the spread of cancer through a body or the adoption curve of a new technology.