Fujitsu’s labs have found the means to scale deep learning across multiple machines
Fujitsu is upping its development of artificial intelligence (AI) once again, championing an improved ability to scale deep learning neural networks across IT infrastructure and boosting the performance of machine learning algorithms.
Under its Zinrai AI initiative, the company’s European labs division has found a way to improve the performance of AI in carrying out tasks like image recognition and object classification, by allowing deep learning neural networks to run across multiple graphics processing units (GPUs) and systems.
AI’s and other machine learning based systems, such as autonomous vehicles, are trained using an deep learning artificial neural network (DNN), which in basic terms is a system of computational nodes that replicate the neurons and synapses of the human brain and essentially take apart large amounts of data and filter it through layers of nodes to recognise and learn patterns in them, for example learning to identify edges and objects in images.
Much like human learning, the training of AIs based on deep learning algorithms require large amounts of data. Processing this data requires a processor that can deal with a high throughput of data, rather than one that can carry out complex computational tasks, as such the parallel processing in graphics cards means, GPUs are favoured over the more serial processing capabilities of central processing units (CPUs).
Fujitsu postures that trying to train AIs to identify tiny and specific details in high resolution images, will push a deep learning neural network beyond the capacity of a GPUs on-board memory, there by limiting how complex a model a GPU can handled.
The soution to this problem is to develop high performance server systems and supercomputers with arrays of multiple and powerful GPUs.
However, this can be prohibitively expensive and thereby lock out the capabilities of advanced AIs from industries that could benefit from improved smart systems, for example using high resolution image renegotiation for use with diabetic retinopathy detection – the spotting of changes in the surface of the retina that can affect the vision of people with diabetes.
Fujitsu is touting the ability to scale deep learning across not the hardware infrastructure of an organisation, including GPUs, server-grade CPUs and application-specific integrated circuits, rather than just multiple identical GPUs.
To achieve this the company’s labs has figured out a means to distribute a large deep learning model beyond the memory capacity of a single system by creating sub layers of the models with in the neural network.
Model-parallelism is used to ensure these sub layers are functionally the equivalent of the layers normal found in a neural network yet are spread across multiple machines and are claimed by Fujitsu to be more computational efficient than running one large model on a powerful machine.
“Wider and deeper Neural Networks are needed, together with finer classification of categories, to address emerging AI challenges. Our solution addresses this directly, distributing DNN memory requirements onto multiple machines,” said Dr Tsuneo Nakata, chief executive at Fujitsu Laboratories of Europe.
“With our technology,it is possible to expand the size of neural networks that can be learned on multiple machines, enabling the development of more accurate and large-scale DNN models.“
With these more accurate deep learning models, Fujitsu see AI being put to more use in across multiple sectors where such smart technology will be of greatest use if it has powerful natural language processing and cognition and the ability to handle data at large scales without the need for prohibitively expensive hardware.
Fujitsu’s AI platform is already being put to use to solve complex problems, as seen with Japanese company Kawasaki Geological Engineering, which has taken Fujitsu’s Zirnai AI platform and used it to create a system aimed at detecting potholes before they emerge.