Comment by mike210 | Hacker News Reader

Comment by mike210 | Hacker News Reader

Parent Full thread

Hi Andy,

We do have lots of unlabelled data, and we're also labeling a large portion of it. We do transfer learn for all of the models we're training, and the first backbone we use is partially self-supervised. Seems to help in overall performance, but it's not a huge effect in our experience.

Maybe once we get a lot of models we can release the backbone weights for nuclear segmentation or at least a competition set of some data we've labeled. Some IP issues here though.

What kind of alternative are you looking for? Specifically one for cells, or just for biologics in general? I'm guessing you're trying to have a better base of weights to transfer off so you can train your own model?

I would say we have medium diversity in terms of images - I think unless you have a similar application right now, you'd be better off transferring off of Imagenet just due to the amount of labeled data.

Thanks for the reply! I'm actually working in a different domain but it seems to have a lot in common with yours - lots of unlabelled data, images that have nothing in common with Imagenet, in that they are all essentially of the same thing and we are looking for variations or features. We found that self-supervised pre-training (with various contrastive models) underperformed vs. starting with weights trained on Imagenet.

So a model that has been pretrained on something else, with enough variability to work as a feature extractor, but closer to the problem framing I mention would be of interest.

For state of the art computer vision stuff, most of the benchmarks use imagenet or similar datasets. But unfortunately I'm coming around to the realisation that those datasets are not representative of most real world problems (except general purpose scene / object recognition). So it becomes very challenging to pick out a potential technique to apply, and hope it transfers.

Interesting - it would be great to chat and find out more. Maybe there are things we can learn about each other. Can you shoot me an email at michael at biodock dot ai?

Do I read it correctly that you're working with images with repeated pattern of instances of the same object on the image? I've been working with cell images, solving segmentation task - what biodock works on - and found interesting tricks to train models on vastly smaller number of labels than what you would think is possible with off-the-shelf models (e.g. Mask-RCNN or U-net + refinements).

Reply on news.ycombinator.com