Anyone who's worked with me knows the written word is not my strong point. Please enjoy the ideas presented here.
Eleven 7.4M-parameter protein encoders, identical except for the pretraining target, compared on secondary structure and fold retrieval. Regressing frozen ESM-IF1 structure embeddings with a linear head beat masked language modelling; predicting the same targets through a JEPA-style predictor did not.
It's more likely to snow in the UK at Easter than at Christmas, but I've never seen snow anywhere near my birthday at the start of April. I'd heard the statistics, but I don't think I had ever properly internalised what they actually meant. This interactive project allows you to understand what the climate should be like on any given day.
Satellites can see far more of the world than they can transmit home, but with how versatile modern image processing AI can be, it has become possible to deploy a general AI model and allow end users to reduce the data they are paying for by running experiments on the satellite and only transmitting the results.
Brain tumour classifiers are already accurate, but none can yet prove it is accurate enough to trust as a black box, so I compared models on how well they explain their answers.
Whatever your real reason for leaving an AI lab, saying you're worried about the models is a dominant strategy.
Machine learning has caught up with physics-based weather forecasting, but the weather services lack information on how to make this transition. For my undergraduate dissertation I conducted an empirical investigation on how best to develop future MLWP models.