SiOx ReRAMs Promise to Accelerate AI Self-Learning

Article By : Gary Hilson

Researchers at Politecnico Milan used Weebit ReRAM provide inference hardware with brain-like plasticity...

Recent research using Weebit Nano’s silicon oxide (SiOx) ReRAM technology outlines a brain-inspired artificial intelligence (AI) system which can perform unsupervised learning tasks with high accuracy results.

The work was done by researchers at Politecnico Milan (the Polytechnic University of Milan) and presented in a recent joint paper with the company that details a novel AI self-learning demonstration based on Weebit’s SiOx ReRAM. The memory technology is considered a prime candidate to succeed NAND flash memory because of its potential to be 1,000 times faster while using 1,000 times less energy than NAND, while at the same time lasting 100 times longer. Weebit’s SiOx ReRAM is also appealing because it can leverage existing manufacturing processes.

ReRAM has also been eyed for AI applications by several research organizations. The university developed a hardware design that uses Weebit’s ReRAM to combine the efficiency of convolutional neural networks (CNNs) with the plasticity of brain-inspired spiking neural networks (SNN) to enable the hardware to learn new things without forgetting trained tasks of previously acquired information. In addition, the system adapts its operative frequency for power saving, enabling feasible solutions for lifelong learning in autonomous AI systems.

ReRAM
Weebit’s ReRAM cell consists of two metal layers with a silicon oxide (SiOx) layer between them comprised of materials that can used in existing production lines. (Source: Weebit)

The research by professor Daniele Ielmini and his team looks at the inability of artificial neural networks (ANNs) to acquire new information without forgetting trained tasks, even though they outperform the human ability of object recognition. The team’s SiOx ReRAM-based inference hardware was able to merge the efficiency of convolutional ANNs and the plasticity of spiking networks. In an interview with EE Times, Ielmini said the research demonstrates that the circuit plastically adapts its operative frequency for power saving and enables continual learning of up to 50% non-trained classes. This optimizes the classification and enables the re-training of the filters, thus overcoming the catastrophic forgetting of standard ANNs, he said.

Ielmini said the biggest challenge for AI hardware to date has been limitations on what it can learn. For example, if the hardware is trained to recognize certain digits, it can only recognize the digits it was trained for, but not recognize any additional digits. Similarly, it won’t be able recognize letters on its own because it was exposed to digits. The purpose of the research was to develop a new hardware based on ReRAM that can continually learn, he said, and it showed that their inference could learn 50% more based on what it was already taught. For example, it could train the hardware on 100 figures, and it could recognize an additional 100 figures without being trained. “This is exactly what happens in the brain when we learn something.”

Essentially, when the brain sees something it recognizes, said Ielmini, there’s a neuron representing the target that spikes. Every time that neuron fires, it spends energy. Because the brain wants to be energy efficient, there is internal feedback which reduces the threshold for the neuron to spike, which ultimately allows for lifelong learning, he said, which the researchers we were able to mimic with their hardware using ReRAM, while achieving high energy efficiency within the system. “This is a big limit of AI hardware nowadays.”

For its part, Weebit Nano wasn’t surprised by the results achieved in Milan, said CEO Coby Hanoch, as the company has been working closely with Ielmini and his team for more than three years. “It is important for us to show that our SiOx ReRAM can serve not only as an advanced memory but also enable other advanced applications.” Weebit Nano has always strongly believed that ReRAM has big potential for neuromorphic applications and other advanced applications, he said.

The company engages with many researchers on potential applications for ReRAM. Hanoch said this research with Polimi research adds plasticity to the current AI systems. “The common approach to AI today is based on supervised learning where you have to spend significant efforts  to train the system, and once trained it can only perform the task that it was trained for.” The human brain, however, can classify objects without being trained massively, he since it has plasticity and is able to project from only few images.

ReRAM maker Crossbar is a founding member of the SCAiLE consortium formed to create AI platforms using the memory technology

Even though Weebit Nano continues to be bullish on potential of neuromorphic applications, said Hanoch, as a commercial company and pre-revenue startup, the company is keeping its focus on generating initial revenue by getting the embedded product out to market and developing discrete ReRAM offerings. “Nevertheless, we are preparing ourselves for the future and neuromorphic systems are poised to be a game-changer in the semiconductor industry,” he said. “Companies like Google, Facebook, Microsoft and Intel are spending significant efforts in this field since they believe in its great potential and need to improve the current ANN systems which require massive HW deployment in the cloud and have very high power consumption.”

Weebit has filed several patents focused on the manufacturing, optimization, and programming techniques of SiOx ReRAM for memory applications and design innovations. “The ReRAM devices will also be the basis of future ReRAM-based neuromorphic systems.”

Its most recent patent with French research institute and long-term partner Leti defines an efficient method to implement robust multi-level storage in ReRAM, which allows for the ability store more than one bit per cell. This in turn increases the memory storage capacity without increasing the number of memory cells or the memory array size, making the memory much more cost-efficient. While this method is based on Weebit Nano’s SiOx-based ReRAM, it can also be extended to any ReRAM technology. “We believe our patent will be required by many ReRAM companies who want to implement MLC,” said Hanoch. “Even those who can do MLC without it will most likely need to implement such a method if they want to implement more levels.”

Weebit is not the only company exploring ReRAM’s potential for AI. In early 2019, consortium dubbed SCAiLE (SCalable AI for Learning at the Edge that included ReRAM maker Crossbar was formed to create AI platforms using ReRAM.

Leave a comment