Computer Vision News Computer Vision News 46 This solution drew inspiration from the softmax layer used in multiclass classification tasks, which enforces probabilities that sum up to one. Paula slightly changed this to force it to sum up to the mass. By adapting and refining this, she created a tailored solution that addresses the unique complexities of the climate science problem. Something Paula says the team were insistent on from the very start was aligning the research with the practical needs of climate scientists. Collaborating closely with domain experts, including climate scientists at IBM, was crucial to understanding the specific requirements and desired outputs within the climate science community. “We had some ideas in mind, but as machine learning people, we didn’t know what the community exactly needs,” she recalls. “It was really important to have people from that domain with domain knowledge be like, okay, these are the kinds of outputs we want to have; these are the kinds of things we care about. We want the model to be something people can use for real-world climate forecasting. We just brought it out, but we’re talking with them, and people are very interested in starting to use it. We really hope that it can be built into operational climate and weather forecasts.” This work originally appeared in the Journal of Machine Learning Research earlier this year, which had a track for ICLR. Regarding the next steps, Paula envisions broader applications within and beyond The numerical models are just too computationally expensive, and that’s where deep learning methods from computer vision can help! ICLR Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=