Memory that learns could help tomorrow's intelligent computers

20.01.2016
As researchers try to build more complex computers that get closer to emulating the way the human brain works, one of the areas of focus is memory.

Existing chips, hard disks and tape drives are great at storing large amounts of data, but a new breed of memory chip called a memristor could go a step further: helping the artificial intelligence systems of tomorrow actually understand the data and make more use of it.

Memristors could help computers connect the dots to identify diseases or help self-driving cars recognize objects based on probabilities and associations. Memristors are best used in machine-learning models to make predictions based on patterns and trends culled from large stacks of information, said Alex Nugent, CEO of Knowm.

Knowm is a New Mexico-based start-up and one of the companies working on memristor technology.

Knowm's memristors are designed around human brains, in which a synapse that connects two neurons gets stronger the more often a signal is passed. Similarly, the learning and retention of information on Knowm memristor circuits are determined by data flow characteristics and the current.

Knowm doesn't yet have a fully functional memristor chip. But it has introduced prototype test kits for researchers and academics on which its memristor design can be emulated. Knowm's test kit will include a chip with analog and digital circuits, software packages and algorithms.

Knowm's current memristor is a "learning processor" that works alongside CPUs, GPUs and other processors, Nugent said.

The company is going up against some big competitors, including HP.

HP plans to use them in a new type of computer called The Machine. It believes memristors could potentially replace both storage and memory in computers and is partnering with SanDisk to make the components. SanDisk says memristors could be 1,000 times faster and durable than flash storage.

Nugent believes memristors will lead to new computers that are better at learning and extracting intelligence from data patterns. Machine-learning is possible on today's computers, but it is not efficient and draws a lot of power, Nugent said.

It could be many years until the first chips based on Knowm's architecture appear in commercial products. The startup is being funded through equity investment and government grant programs, and will receive more equity funding from an undisclosed partner in the coming months.

Agam Shah