Category: Artificial Intelligence - HJS

Graph Convolutional Network (GCN)

9/8/2021

Image credit: https://tkipf.github.io/graph-convolutional-networks/

Graph Neural Networks (GNNs) have gained immense popularity as a new methodology of analysing graphs, similar to how we use deep learning to process images. GNNs can process many real‐world datasets, which inherently have graph structures. They also provide application scenarios in different domains, from recommendation systems to road traffic predictions and drug design. GNNs use the same principles as Convolutional Neural Networks (CNNs) with re‐definition for the graph domain. Like in images, the convolution framework of GNNs is capturing neighbourhood information of graph nodes to process graph structures represent in molecules, social‐, citation‐ and road networks. With the advancements of machine learning, graph data processing is nowadays possible in the most effective manner.

This repository contains the training and inference source code of a GCN for our technology platform. Please note that model training requires TensorFlow 1.15 and model inference our pyACL framework.

Please use the following link to visit our repository. http://shrnk.cc/sdrwl

A fireside conversation with the pioneer of computer science

16/4/2021

To young talents and fresh graduates! Please join our fireside conversations with John Hopcroft—a pioneer of computer science and Turing award winner. This event is the first appointment of a fantastic series to explore career progression themes within computer science and artificial intelligence.

We will have an exclusive live conversion to explore the industry's future, how it will help solve global issues and unearth career inspiration for everything—whether you are a novice or veteran.

You will hear first-hand experiences, personal stories and maybe a few predictions. Please do not miss it!

2 PM-3 PM CET on April 22, 2021
Sign up link: https://bit.ly/2PO0Cg1

Da Vinci — A scalable architecture for neural network computing (updated v6)

2/3/2021

The updated version of this presentation provides the detailed descriptions of two more projects:

Object detection
Body pose detection

Da Vinci — A scalable architecture for neural network computing (updated v5)

13/1/2021

The updated version of this presentation provides the following changes:

Additional application scenarios for artificial intelligence (upscaling and colourisation for video footage)
Information on our chip enablement layer and used computing language
Detailed instructions on how to prepare the development environment, the SD card image and the installation of third-party packages
Detailed description to create the first project: Colourful Image Colourisation

Da Vinci — A scalable architecture for neural network computing (updated v4)

4/11/2020

The updated version of this presentation provides the following changes:

Additional background information in the introduction section
Shows comparisons of different processors for AI
Describes how AI processor architectures will shift in the future
Describes the basic principles of Convolutional Neural Networks
Shows the advantages of our special compute units for AI
Shows Mind Studio in more detail
Shows an excerpt of listed models of our Model Zoo
Advises on how to start with our Atlas 200 DK developer board

Artificial intelligence that can write almost anything

30/9/2020

GPT-3 is an autoregressive language model that uses deep learning to produce human-like text. The following blog post gives an idea of the quality of this AI model. It is scary and fascinating at the same time, especially when you consider "fake news". Please note that the bold marked text is the input. The rest is the automatically generated text.

To read the full blog post, click here.

Book on AI applications using our Ascend chip

26/8/2020

Elsevier released the English version of the book Ascend AI Processor Architecture and Programming: Principles and Applications of CANN.

The book gives an in-depth description of artificial intelligence applications by using our Ascend chip and analyses the unique performance attributes of this processor. It also introduces theoretical aspects of artificial intelligence, describes the hardware and software architecture; and all related tools to program the technology. The book concludes with detailed case studies on data and algorithms for artificial intelligence. Enjoy!

Please use the following link to buy the book: http://shrnk.cc/ea922

The king is dead, long live the king! ... or key takeaways from the 55th TOP500 list

22/6/2020

The new TOP500 list of the fastest supercomputer was announced. I expected many things, but also have to admin that some things came as a surprise. Congratulation to Fujitsu, the entire Riken team and Japan to push an Arm-based system to the pole position of the new TOP500 list.

My key takeaways are:

The first supercomputer, which can already be called an exascale system, does not come from China or the USA. It is from Japan and used the Arm architecture!
The Japanese supercomputer tops the new 55th TOP500 list of supercomputers with 415 PFLOPS by using Arm-based processor.
The Japanese supercomputer is so powerful that it also scored number 1 on the HPCG and HPL-AI list.
The application requirements of OpenFoam, SPECFEM3D and WRF influence the design of the Japanese supercomputer strongly.
The company "Preferred Networks" has developed a highly efficient matrix accelerator (MAU), which reaches 21.1 GFLOPS/Watt by using PCIe adapters combines with low-power Intel Xeon CPUs. This system leads to the number 1 position on the Green 500 list, just ahead of NVIDIA.
China has with 226 by far the most systems in the new list, ahead of the USA with 114 and Japan with 29. France follows this with 19, Germany with 16 and the Netherlands with 15.
In terms of performance, the USA continues to lead with 639 PFLOPS, followed by China with 566 PFLOPS and Japan with 528 PFLOPS.
Among manufacturers, Chinese vendors are dominating. Lenovo has 180 systems, Sugon has 68 systems, and Inspur has 64 systems.
Overall, the total performance of the list has increased significantly by 35 per cent to 2.22 EFLOPS—previously it was only 5.5 per cent increase.
China continues to have a significant footprint in commercial systems on the new TOP500 list, which is 78% of commercial systems that represents 52% of the performance of all commercial systems.

Da Vinci — A scalable architecture for neural network computing (updated v3)

10/5/2020

The updated version of this presentation provides additional information on the applicability of artificial intelligence in modern medicine, shows more insights into the end-to-end life cycle of AI implementations in projects and gives more details of our software stack.

Da Vinci — A scalable architecture for neural network computing

27/3/2020

In this presentation, I give an introduction into microprocessor trends, describe two distinct eras of computing usage in training AI systems and show the wide variety of computing architectures in computer science. I also describe our advanced computing and artificial intelligence product portfolio, which focuses on innovation, continuous dedication and backward compatibility. The central part of this talk is insights into our Da Vinci architecture, descriptions of all building block, the core architecture and its micro-architectural configurations. Last, I show the process of how we execute artificial intelligence projects and the challenges which are still ahead of us.