Hi @AIMLawg & @MicrobesAWG
Came across this really interesting piece on why AI keeps failing at Microbiome prediction (vs traditional algos/models)
https://blekhman.substack.com/p/ai-keeps-failing-at-microbiome-prediction
7 Likes
Great Substack post!
The reality is that foundation models are very complex, and using them for most downstream tasks can be tricky and very dependent on the data.
Sometimes the results in this field are real but still feel surprising. For example, this paper (https://ai.nejm.org/doi/full/10.1056/AIx2500261) shows that a task-specific foundation model (RETFound) is not significantly better than a general model pretrained on ImageNet for some eye-related tasks !
Another issue is that fine-tuning these models is complicated(billions of parameters!). Because of that, researchers started working on adaptation methods like LoRA ([2106.09685] LoRA: Low-Rank Adaptation of Large Language Models), which has now become its own research area !
Foundation models are very cool and powerful, but they are not always the best or simplest option for every problem 
2 Likes