UK project uses supercomputers, synthetic data to improve emotion recognition

UK supercomputing power will be used to test a new facial emotion recognition system that relies on synthetic image data.

A project from Goldsmiths University in London has been selected for the UK government’s AI Research Resource (AIRR) program, which provides free computing power to researchers, academics, and small and medium enterprises (SMEs). The computing power comes from two major supercomputers, the University of Cambridge’s Dawn and the University of Bristol’s Isambard-AI.

The three-month project will include intensive experiments aimed at addressing challenges in emotion recognition.

Although systems that recognize facial expressions such as anger, fear, happiness and sadness are already present in many technologies, emotional recognition systems have often been accused of misclassifying emotions from certain groups. The reason behind this is messy datasets, which are often mislabelled and lack diversity.

To solve this, Goldsmith University lecturer Georgios Mastorakis aims to train an emotion recognition system using synthetically generated images of people. The goal is to ensure a well-labeled image dataset, which systematically varies poses, lighting, occlusions and facial types and includes faces from different body types and demographics.

“We train on synthetic, then use this transition when we train and test on real human emotions and see if we can get high accuracy,” Mastorakis said in a statement published by the university.

The synthetic dataset was created with computer graphics software suite Maya and originates from work at the University of Washington.

Launched in July last year, AI Research Resource (AIRR) is a strategic UK initiative designed to support research into foundational AI, data sovereignty and innovation as part of the government’s AI Opportunities Action Plan. According to the plan, the country will invest £1 billion (US$1.36 billion) to expand computing capacity by twenty-fold by 2030.

The University of Bristol’s Isambard-AI is currently the UK’s most powerful public computing facility and the 6th-fastest supercomputer in Europe. It contains nearly 5,500 Nvidia GH200 Grace-Hopper superchips, supplied by HPE (Hewlett Packard Enterprise).

​The Dawn supercomputer at the University of Cambridge, on the other hand, is made up of more than 1,000 Intel Data Center GPU Max 1550 GPUs and operated in partnership with Intel and Dell.

The government has announced investments of over £350 million ($476.8 million) into the two clusters by 2030.

First-time users from academia or industry can use the capacity for algorithm testing and benchmarking, with AIRR providing up to 10,000 GPU hours within three months. UK-registered microenterprises and SMEs can get up to 20,000 GPU hours for early-stage AI product development within the same timeframe.

Finally, larger research partnerships can get between 50,000 and 150,000 GPU hours over six months for various tests, including developing new algorithms and AI-driven data synthesis.

Among the first projects that are testing the supercomputing power is one that analyzes recordings from wearable cameras and other devices to help people with dementia perform tasks more easily at home. Another project will train AI to analyze MRI scans and help identify cancer.

Article Topics

biometric dataset  |  biometrics research  |  emotion recognition  |  face biometrics  |  generative AI  |  synthetic data