MediaTek’s AI research group has just published a paper at ICLR detailing its new algorithm using Fisher-Legendre (FishLeg) optimization to train AI models faster and more reliably than previously possible. Improving the efficiency of training is important because it can help to reduce the resource heavy and energy intensive process that is usually necessary when training large models. The research was conducted in collaboration with the neural dynamics and control group at Cambridge University, UK. The AI research group is presenting this work at the ICLR conference on May 1st.
- To learn more, the paper is available here >
- An implementation of the algorithm is publicly available to use here >