Deep learning – not just for big business

Main visual : Deep learning – not just for big business

Deep learning has become the premier discipline in the world of Artificial Intelligence (AI). Thanks to deep learning methods, computers are now able to see better than humans and can even handle voice and text applications with ease. And that makes deep learning a decisive success factor in a competitive business environment.

However, using AI as a tool can be rather challenging. That’s why Fujitsu developed its Deep Learning Unit (DLU), which is an integral part of the FUJITSU AI Zinrai Deep Learning System (ZDLS). The idea is to provide a complete solution that introduces users to AI in a quick and easy way.

The added value of deep learning far outweighs the cost

With deep learning you can give computers the ability to learn by example. Human beings learn from pictures, text and acoustic signals. If they get the proper training input, deep learning computer models can learn this way as well, but much faster and with more precision than humans. That’s not magic – it’s strictly math.

Deep Neural Networks (DNN) serve as the basis for deep learning applications. These networks can have 150 or more hidden layers between their input and output layers. When compared with other machine learning processes, deep learning is extremely complex and involves much more extensive development in terms of setting up the deep learning models. However, the added value business enterprises can expect far exceeds the investment and effort – and that’s why everybody is so excited about this!

AI is expected to make sure that self-driving cars can recognize stop signs and distinguish between pedestrians and lamp posts for the scale of the fleet network we need to cover. Deep learning is helpful in the medical segment because it can analyze image data and identify unusual characteristics. Insurance companies rely on image categorization to uncover criminal activities like insurance fraud. In industrial environments, deep learning solutions combined with high-resolution cameras are used in automated optical inspection processes to find surface cracks, roughness or material contamination much faster and with more precision than the human eye.

Hurdles are often too high for SMEs

Many business enterprises recognize the potential of deep learning, but the technology is mostly deployed in large corporations. This is because rather complex models are integrated with the business applications – and this involves high levels of investment in AI expertise, time and resources. One factor is the huge volume of high-quality data involved. For example, millions of images and thousands of hours of video material are required for the development of self-driving cars. Extremely complex architectures and huge amounts of data mean that training deep learning models can take many days and weeks. And since those models can learn from their failures and successes, it is essential that they are continually fed with more and more data – over time they will become better and better.

The Deep Learning Unit saves a lot of time and energy

Deep learning in productive environments typically needs high-performance server and storage systems. To date servers have typically been equipped with General Purpose Graphics Processors (GPUs). However, GPUs have functions on board that have little to do with deep learning. In other words, they are not optimized for deep learning – and they consume a lot of energy/power. 

Fujitsu’s Deep Learning Unit (DLU) is different: It is designed and developed specifically for training neuronal networks – and is thus much faster and more energy-efficient than GPUs. Preliminary internal tests have shown that the DLU can boost performance per watt by a factor of ten. In a nutshell, the #DLU lets enterprises compute faster, while cutting energy consumption.

Researchers at the University of Massachusetts, Amherst, made a remarkable discovery related to AI models and energy. They trained various large-scale AI models and noticed that training one model can result in emissions that are the equivalent of 280 tons of carbon dioxide. That’s about five times the amount produced by an American car during its lifetime (including the manufacturing phase)!

The bottom-line: The DLU can be an important Green IT-component for enterprises.

End-to-end offering for an easy start

Fujitsu offers a complete framework within the scope of AI Zinrai. The Zinrai Deep Learning System (ZDLS) is an innovative end-to-end solution with AI services, software and infrastructure that ensure the quick setup of high-performance deep learning platforms. The DLU is the central hardware core of this AI solution.

It will be available from late 2019 in one integrated system which is ready for operation in a snap – and it also includes predefined learning models. The solution is affordably priced and keeps pace with growing customer requirements: ZDLS is equipped with two DLUs and can be expanded up to eight DLUs, and the systems can be switched in series.

Thanks to AI software components, plus AI services and integration services, Fujitsu makes life easier for its customers – we don’t have to deal with tedious configurations and productive AI solutions are up and running in no time at all!

The solution also has a remedy for the major headache that stems from having to identify applications and formulate the business problem in such a way that it can be transferred into an AI solution. The Fujitsu Co-Creation Program provides relief through in-depth AI expertise and design thinking methods that unleash creativity and result in digital transformation through innovative ideas.