Gary Marcus

I haven't read your book (yet) but looking forward to some realism! In my experience deep learning is great for computer vision tasks where output is binary with high accuracy - BUT many medical tasks are not binary at all, and accuracy can be hindered by many confounders

The issue with any medical vision applications is that the number of possible medical diagnoses in a given situation is vast, yet not all are represented well in datasets, and rare entities can be critical. For example, if I get a biopsy from a chronic non-healing wound in a...

I've actually got what I think is a very relevant blog post coming out in the next few days. Preview of core idea: we don't need to reboot AI to make medical applications much safer than they are today, we need to use human smarts to much more effectively validate these models.

Deep learning is great in medicine for image segmentation. Clinical useful for accurately quantifying tumour/ventricular volumes, 3D models etc. CNN output can be easily checked/quickly edited if needed.

I’d have to take time to read the link—thanks for sharing. In medicine at least, safety without ability to explain and rationalize would be an oxymoron—if I can’t explain why a lab test would produce unreliable results and the circumstances where it does, then it’s not safe.

My current understanding on this: deep learning for critical medical decisions is inherently unsafe. Why? DL (appears to) interpolate, and this will always be from an insignificant subset of the space of all possible inputs => not a reliable way to fully specify the problem.