[peppery]

Your work with ML and images sounds incredible!

Do you have a link (e.g. on arxiv or elsewhere) that describes your approach using ML for image reconstruction in greater detail? How would you recommend building up one's combined intuition in optical theory, the relevant ML techniques, and the biological substances themselves, to the level where you can innovate in this task as you have done?

Also, for those interested in the concept of building better images using higher fidelity simulations of the microscope itself, presumably Andrew meant studies along these lines: https://arxiv.org/abs/1702.07336

[andbberger]

That paper has the right general idea, you need a generative model of the scope and if you're clever you can use that to improve your reconstructions. But it's missing a couple key things that it workable in practice. One missing ingredient is a probabilistic model of the thing you're imaging. The others are secret ;)

Secret because I have a dream for a crazy startup based on this idea which I don't have the means to do now. Although I generally hate being secretive about, like, knowledge and for sure the value creation happens during execution - but just humor me this time, ok?

> How would you recommend building up one's combined intuition in optical theory, the relevant ML techniques, and the biological substances themselves, to the level where you can innovate in this task as you have done?

Well, I'm flattered but I haven't done any substantial innovation here... My recommendation would be to become a theorist, learn math and physics and work your way up the hierarchy. You need to understand the whole picture, how it works on each level, and how the levels fit together - then you can run thought experiments. A good generative model of the world.

[peppery]

Thanks for your sharing your thoughts!

> One missing ingredient is a probabilistic model of the thing you're imaging

Ooh, this (making the prior for the true image signal more informative by incorporating knowledge of the structure of the signal) is clever. Here, when you say model, you mean a description that is based on the (bio)physics of your sample? (E.g., knowing that objects being imaged obey diffusion equations informs your maximum likelihood estimation of the true signal?)

> Secret because I have a dream for a crazy startup based on this idea which I don't have the means to do now. Although I generally hate being secretive about, like, knowledge and for sure the value creation happens during execution - but just humor me this time, ok?

Of course; such is the right (and joy) of an innovator to define how one's own idea is disseminated/actualized! (~: It is intriguing that you feel your idea has the character of best being pursued via a startup (addressing some crucial unmet commercial need), rather than via the academic model (e.g. transformative Nature publication) more commonly used for improved microscopy techniques. Best of luck in this pursuit; I look forward to seeing your startup's innovations someday soon!

> You need to understand the whole picture, how it works on each level, and how the levels fit together - then you can run thought experiments. A good generative model of the world.

This idea of developing a "good generative model of the world" is a beautiful aspiration for all of us to have. Thanks for your insights!