It’s interesting that the cross-species UCE model is the most performant. InstaDeep recently published their ChatNT genomic model and showed that training on multiple species genomes simultaneously, and selectively switching between objectives, produced a model that generally outperforms specialized models. Makes me wonder whether providing several “escape hatches” in terms of objectives makes the training process more robust to getting stuck in local minima.
I really enjoyed reading this post. I'm particularly excited about it because I'm preparing for an upcoming STEM outreach program for high school students on the applications of AI in single-cell biology. With your permission, I'd love to incorporate some of the points you discussed in my presentation.
It’s interesting that the cross-species UCE model is the most performant. InstaDeep recently published their ChatNT genomic model and showed that training on multiple species genomes simultaneously, and selectively switching between objectives, produced a model that generally outperforms specialized models. Makes me wonder whether providing several “escape hatches” in terms of objectives makes the training process more robust to getting stuck in local minima.
Hello Abhishaike,
I really enjoyed reading this post. I'm particularly excited about it because I'm preparing for an upcoming STEM outreach program for high school students on the applications of AI in single-cell biology. With your permission, I'd love to incorporate some of the points you discussed in my presentation.