Some of the hardest challenges in time-series modelling come down to two questions: how do signals relate to one another and how to learn efficiently across times.
Federico has been tackling the first of these at Imperial College London, studying one of the most complex sources of time-series data: the human brain. His research uses graph neural networks to uncover the structure of interacting neural signals, revealing patterns in brain activity associated with motor learning.
Yi Heng has been tackling the second. He’s the lead author of the ICLR 2024 paper Parallelizing non-linear sequential models over the sequence length, which demonstrated how nonlinear recurrent neural networks can be parallelised across time. His work laid the foundations for a new generation of parallelised recurrent models, including Apple’s ParaRNN, which scaled nonlinear recurrent language models to 7 billion parameters.
Both bring expertise directly relevant to our mission of building efficient foundation models that learn from irregular, asynchronous and multiscale time-series data.
Welcome to the team, Federico Nardi and Yi Heng Lim!