IBM uses analog computing for artificial intelligence to reshape AI computing

IBM has been researching ways to reshape the way artificial intelligence computes. IBM researchers have published a paper describing a breakthrough in using analog computing for artificial intelligence (AI).

When building an artificial intelligence system, the data model needs to be trained. It is to assign different weights to different subsets of training data, such as image data describing different characteristics of cats.

When training an artificial intelligence system on a traditional (digital) computer, the artificial intelligence model is stored dispersedly in memory. Computing tasks require constant transfer of data between memory and processing units. IBM says this process slows down computing and limits the energy efficiency that can be achieved.

Using analog computing for artificial intelligence may provide a more efficient way to achieve the same results as artificial intelligence running on digital computers. IBM defines simulated in-memory computing, or simulated artificial intelligence, as a technology that borrows key features of how neural networks in biological brains operate. Researchers say that in the brains of humans and many other animals, the strength of synapses, called weights, determines the communication between neurons.

In simulated artificial intelligence systems, these synaptic weights are stored in-situ in the conductance values ​​of nanoscale resistive memory devices such as phase change memory (PCM), IBM said. They are then used in deep neural networks to perform cumulative multiplication operations.

IBM said this technology can reduce the need to constantly send data between memory and processors.

In a paper published in Nature Electronics, IBM Research introduced a mixed-signal analog artificial intelligence chip that can run various deep neural network (DNN) inferences Task. According to IBM, this is the first analog chip that performs computer vision AI tasks as well as digital chips in tests, and is more energy efficient than the latter.

The chip is manufactured at IBM's Albany Nanotechnology Center. It consists of 64 analog memory computing cores (or chips), each containing a 256 x 256 crossbar array of synaptic cells. IBM said a time-based analog-to-digital converter is integrated into each chip to convert between analog and digital data. Each chip also integrates lightweight digital processing units that IBM says can perform nonlinear neuron activation functions and scaling operations.

IBM stated that each chip can perform calculations related to a layer of DNN model. The authors of the paper said: "Using this chip, we conducted the most comprehensive study on the computational accuracy of analog memory computing and achieved 92.81% accuracy on the CIFAR-10 image data set."

The above is the detailed content of IBM uses analog computing for artificial intelligence to reshape AI computing. For more information, please follow other related articles on the PHP Chinese website!