Creating PCs that can learn, inventor's goal

Chip designers received an interesting challenge this month from Jeff Hawkins, the founder of Palm and Handspring and an expert on the human brain: If they really want to design something intelligent, they shouldn't be doing processors. They should be making memories.

At the International Solid State Circuits Conference in San Francisco, Hawkins addressed the question, "Why can't a computer be more like a brain?"

He has been preoccupied with that question for decades and has been working on the answers for his startup, Numenta.

Hawkins, inventor of the Palm Pilot, wrote about his research on the brain in 2005, in the top-selling book, "On Intelligence."

Brainlike software

He created the Redwood Neuroscience Institute in 2002 and in 2005 cofounded Numenta with his longtime partner Donna Dubinsky. Their company is working on software that is brainlike in recognizing patterns.

Hawkins says the software may have applications in computer-visual recognition, among other areas.

Hawkins said computers today are fast at solving calculations but terrible at doing things that children can do, like understanding simple stories or recognizing the difference between a cat and a dog. That's because computers are designed with a processor connected to memory.

The brain, by contrast, has a "hierarchical memory system." At the lowest level, its neural connections store memories such as the shape of an object.

At the next level up, the pattern associated with a human form is stored. And at the highest level, the brain recognizes the person as Bill Clinton.

The brain operates on the same formula, or algorithm, in every region, according to Hawkins.

It doesn't have one algorithm for processing vision or a different one for processing sound.

Each part of the brain is also self-training.

You are born with nothing and learn. Each part of the brain learns based on what information is passing through it.

"This is a basic blueprint for how the memory system works," Hawkins said.

All the nodes, or neuron connections — the places where memories are stored — do the same thing. They look at patterns of data coming in. They look for common spatial patterns and sequences of those patterns, and store them.

If a node learns sequences that it sees, then it passes that information upward, like passing a name upward instead of the details.

You have fast-changing patterns at the bottom of the hierarchy and slow-changing patterns at the top. Words and concepts are at the top.

Each node can predict what the next sequence will be and passes that along. You can generate a complex pattern.

"What this does is you build a hierarchical model of the world," Hawkins said.

Time-stamped

The recognition process improves because the memories are essentially time-stamped.

For instance, the brain might have a hard time recognizing a helicopter in a snapshot of a scene. But if the memory stored shows how that helicopter moves within the scene, it is much easier to recognize.

Numenta's software can accomplish this kind of recognition task today. When shown a crude picture of a helicopter, it can recognize it 19 percent of the time.

But when the helicopter is moving and several frames of the memory are stored together, the software can recognize the helicopter 52 percent of the time, Hawkins said.

Numenta released its software development tools about 10 months ago.

The partners are using the software for applications such as video games, auto lane-change prediction, speaker voice identification, visual-object recognition and process control of power networks.

The goal is to create a computer that is efficient at learning to recognize things, improves itself over time, can be generally used across a lot of different industries and has an ability to make predictions.

"The goal is not to build humans, but to solve problems that are much harder" than what humans can solve, Hawkins said.

Dean Takahashi is a

technology columnist at the

San Jose Mercury News.