@bernhardsson I think if you take into account the infra credits and discounts available for small teams, it probably is not much of a problem. It also helps with hiring from the big firms if you're using the same environment.
If you speak Python, love Huggingface and SpaCy, and dig the idea of a corporate startup that applies the latest of NLP to a super exciting industry: @jorenvs is building a little dream team at @RepriseDigital and is hiring an NLP engineer! #Job https://t.co/tSFjo0bA0a
@ch402@rijulg Very interesting article! Do you think that superposition in neurons is a way of overfitting to a dataset that has no dogs and cars together for instance? What would happen if we force pure neurons? (Can we even do that?)
@vincent_spruyt Most super resolution approaches are heavily biased towards a specific downscale/degradation function, usually MATLAB's bicubic interpolation. "Blind" super resolution is only a very new field.