What is HTM and how is it different from conventional AI?
HTM (hierarchical temporal memory) is a theory developed by Jeff Hawkins to explain how the human brain works. His company, Numenta, has developed algorithms to replicate these processes in a software platform on which applications can be built.
Although technologies like neural networks have attempted to replicate the structure of the brain, there have not been commercial successes modeled on how the brain works. Instead, AI tends to apply rules-based programming, since computers excel at rapid execution. But this rigidity struggles in real-world problems that humans find trivial, such as finding objects in pictures. By adding computing power to the flexibility of human-like thinking, HTM can potentially solve problems an order of magnitude more complex than "narrow" AI can address.
How the brain works
HTM theory describes intelligence as a learning-based memory system. Humans build a model of the world by taking sensory input over time and finding patterns. These patterns are stored in hierarchies. For example, when you look at a picture, you see lines, curves and corners, and learn combinations that you remember as shapes. You learn that certain combinations of shapes are leaves, bark and branches. Combinations of those objects form trees, and combinations of trees form the concept of a forest.
Once you learn the abstract concept of a forest, it becomes easy to identify pictures of forests. You don't have to compare the exact pixels in the image with all of the forest images in your memory. Even when looking at a new image, if your eyes recognize some combination of trees, you think "forest." It's efficient because you don't have to store every image of a forest you've seen. And since you are thinking about high-level concepts, you are not confused if some leaves and branches are obscured. That degree of flexibility far exceeds that of rules in a program.
What this means
Imagine combining this robust flexibility with computing power, and inputs that are not limited to five senses. Sensors on an automobile could monitor visual, audio, speed, vibration and other inputs to warn drivers of dangerous situations. The same processes could be applied to translate languages or find patterns in consumer behavior and match them to relevant information. Any form of data or sensory input that generates patterns presents an opportunity to learn and process with HTM. And in the long run, the ability to integrate different HTM programs will enable an exponential growth in capabilities that has not been seen in conventional AI despite decades of development.
For more information
This is only a simplified summary of HTM. For technical details, see www.numenta.com.