Source: siliconangle.com
Israeli deep learning startup DeepCube Ltd., which has built software to accelerate the process of running machine learning models in the data center and on edge devices, said today it has raised $7 million in a new round of funding.
The Series A round was led by Awz Ventures with participation from Koch Disruptive Technologies and Nima Capital, and brings DeepCube’s total amount raised to $12 million.
DeepCube, which is led by its co-founders Dr. Eli David (pictured) and Yaron Eitan, recently announced the launch of what it says is an industry-first software-based inference accelerator for deep learning, a subset of artificial intelligence that tries to mimic the way the human brain learns. The company reckons that its accelerator can dramatically improve deep learning performance on intelligent edge devices.
The problem that DeepCube is trying to solve is that deep learning deployments are still somewhat rare because of the size and speed of neural networks, plus their need for specialized hardware that’s not only expensive but also has serious compute and memory demands. Because of this, it says, deep learning remains very difficult and expensive to perform on edge devices, where resources are limited.
DeepCube’s software tries to fix all this by accelerating deep learning processes. It works by reducing the size of any deep learning model, including its training data, in a completely automated way and without any manual intervention.
The software can be deployed on central processing units, graphics processing units and application-specific integrated circuits, or ASICs. Those are computer chips that have been customized for a particular use, in both the data center and on edge devices.
With its software, DeepCube says it is able to deliver up to 10-times speed improvement and memory reduction, thereby enabling more advanced deep learning on any device.
DeepCube isn’t the only company in the deep learning acceleration business. For example, Nvidia Corp. has its Deep Learning Accelerator and Micron Technologies Inc. has its Deep Learning Accelerator platform. But DeepCube’s founder David told SiliconANGLE that DeepCube is unique in that it has created the only software framework that allows for both automatic optimization of any deep learning model, and over 10x inference speedup.
“The other so-called software accelerators do not algorithmically modify deep learning models, and so their speedup improvement is incremental, usually around 10%-30% at most,” David said. “DeepCube’s technology allows for aggressive automated restructuring of the deep learning model, so that it ends up being under 10% of its original size. Over 90% of the connections are removed.”
Awz Ventures founder and Managing Partner Yaron Ashkenazi said the inability to deploy deep learning at the edge, on small devices with minimal memory and processing power, has hindered adoption of the technology.
“DeepCube is the only company that has been able to demonstrate the necessary paradigm shift to change this,” Ashkenazi said. “DeepCube’s technology has the power to unlock truly autonomous decision-making in semiconductors, data centers and on edge devices, while improving speed and memory reductions. This is absolutely critical to the future of deep learning.”
Holger Mueller of Constellation Research Inc. told SiliconANGLE that one of the ultimate prizes of AI is to deliver deep learning capabilities that can change business outcomes.
“When software can learn by itself and change based on data, we will have achieved something critical with automation that can be done autonomously,” Mueller said. “Edge locations are of particular interest in deep learning automation because devices need to function without a human operator, and with power and bandwidth constraints.”
DeepCube said it will spend its new funds on research efforts and expanding the market for its software.
