Requests for name changes in the electronic proceedings will be accepted with no questions asked. However name changes may cause bibliographic tracking issues. Authors are asked to consider this carefully and discuss it with their co-authors prior to requesting a name change in the electronic proceedings. Use the Report an Issue link to request a name change.

Outlier Features (OFs) are neurons whose activation magnitudes significantly exceed the average over a neural network's (NN) width. They are well known to emerge during standard transformer training and have the undesirable effect of hindering quantisation in afflicted models. Despite their practical importance, little is known behind why OFs emerge during training, nor how one can minimise them.Our work focuses on the above questions, first identifying several quantitative metrics, such as the kurtosis over neuron activation norms, to measure OFs. With these metrics, we study how architectural and optimisation choices influence OFs, and provide practical insights to minimise OFs during training. As highlights, we introduce a novel unnormalised transformer block, the Outlier Protected block, and present a previously unknown benefit of non-diagonal preconditioning optimisers, finding both approaches to significantly reduce OFs and improve quantisation without compromising convergence speed, at scales of up to 7B parameters.

Vision transformers have revolutionized computer vision, but their computational demands present challenges for training and deployment. This paper introduces LOTUS (LOttery Transformers with Ultra Sparsity), a novel method that leverages data lottery ticket selection and sparsity pruning to accelerate vision transformer training while maintaining accuracy. Our approach focuses on identifying and utilizing the most informative data subsets and eliminating redundant model parameters to optimize the training process. Through extensive experiments, we demonstrate the effectiveness of LOTUS in achieving rapid convergence and high accuracy with significantly reduced computational requirements. This work highlights the potential of combining data selection and sparsity techniques for efficient vision transformer training, opening doors for further research and development in this area.