amelie's recent activity

  1. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    I don't envy you being a teacher in a time like this, I knew so many people that would have used them for everything. Wouldn't even know where to begin dealing with something like that. I think it...

    I don't envy you being a teacher in a time like this, I knew so many people that would have used them for everything. Wouldn't even know where to begin dealing with something like that.
    I think it would just hurt to see it happen, that they're so willing to take the easy route rather than actually learn. Though I guess that's been an issue for decades now.

    1 vote

  2. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    Sorry about the delayed response, I’ve been travelling the last week or so. It’s really worrying just how coherent it has become, I don’t think I would have believed such a thing even a couple of...

    Sorry about the delayed response, I’ve been travelling the last week or so.
    It’s really worrying just how coherent it has become, I don’t think I would have believed such a thing even a couple of years ago.
    It is such an amazing tool, that has been bastardised by high schoolers using it for homework and for people that think that it’s the new Google. I get that openai need to have a product and all that, but it’s just so dangerous.

  3. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    It’s the issue with and machine learning/ai they’re such useful tools but the people using them tend to be completely blind to what they actually are. LLM are a bit far from my personal expertise...

    It’s the issue with and machine learning/ai they’re such useful tools but the people using them tend to be completely blind to what they actually are. LLM are a bit far from my personal expertise but I see a similar trend which the misuse of the gpt models now that they’re ‘public’
    Sadly the paper was a prerelease, I’m not sure what actually happened with it, I’ll have a look though and see if I can find a copy.
    Honestly it’s the same for me, I think I made a grand total of one post and about 5 comments in just under ten years on Reddit. I’ve been amazed at just how conducive this platform is to meaningful conversation. Also just the chances of stumbling upon someone with experience in a somewhat similar and relatively niche subject matter is amazing.

    1 vote

  4. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    Honestly I completely agree with your takes on ML despite my experiences with it. It really rubs me the wrong way just kind of throwing my data into a black box, even if it’s a very big black box...

    Honestly I completely agree with your takes on ML despite my experiences with it. It really rubs me the wrong way just kind of throwing my data into a black box, even if it’s a very big black box that I designed.
    And sure it has its uses, I reviewed a paper that used a small CNN to solve an ill-posed inverse problem. And sure it appeared to work, and was incredibly efficient computationally, I just can’t bring myself to trust it, tools like that are so easily misused. I’m not going to trust the outcome of a paper if the analysis is entirely driven by ML.

  5. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    Honestly I just found any of the ‘biomedical engineering’ industry jobs super boring. Pretty much just turn into glorified tech support for existing products. There’s a weirdly large amount of...

    Honestly I just found any of the ‘biomedical engineering’ industry jobs super boring. Pretty much just turn into glorified tech support for existing products.
    There’s a weirdly large amount of overlap between radar & microscopy data, and it was a way out of being stuck in academia for the rest of my life so I went for it.
    Do you mind if I ask what your specialty is?

    1 vote

  6. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    You sound like you’ve got a really robust platform now I’m glad I’m of some use. I guess that would probably be the closest thing to describe what I did. I did the software and analysis for a lab...

    You sound like you’ve got a really robust platform now I’m glad I’m of some use.
    I guess that would probably be the closest thing to describe what I did. I did the software and analysis for a lab that focused mechanobiology. So most of my experience is with very specific analysis of cellular imaging.
    I’m actually a biomedical engineer by trade. Not that I’ve ever actually worked as one. I’m now working in the defence industry as a signal analysis engineer.

    1 vote

  7. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    You should really only need to overlap a few pixels, I never got around to it in my implementation it may also be worth blending the edges of the tiles together as well. It’ll just make the tiles...

    You should really only need to overlap a few pixels, I never got around to it in my implementation it may also be worth blending the edges of the tiles together as well. It’ll just make the tiles a bit more homogenous when put together.
    Also if you’re training on tiles I would add contrast variation to your augmentations, otherwise it can overtrain on the exposure of each pictures set of tiles.

    1 vote

  8. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    I’d be interested to see if the gravitational component would affect anything. I’m really not sure. If your images are rather large I’ve also had some success with tiling the original image before...

    I’d be interested to see if the gravitational component would affect anything. I’m really not sure.
    If your images are rather large I’ve also had some success with tiling the original image before feeding it through the Unet. Lowers GPU requirements massively. Though due to the convolution the edges of the tiles can get messed up. I had to write a script for tiling that overlapped each tile and reduced the convolutional artefacts. Also means that the dataset is much larger than it would be otherwise.

    1 vote

  9. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    Yeah I used keras and pretty much just recreated the model from the original paper. Though I’ve also heard good things about the Unet++ which basically just adds additional skip connections. For...

    Yeah I used keras and pretty much just recreated the model from the original paper. Though I’ve also heard good things about the Unet++ which basically just adds additional skip connections.
    For augmentation I only used reflection, rotation, and some noise addition. For my work I found that changing the scale to any meaningful degree was negatively impactful, though you may find otherwise.

    1 vote

  10. Comment on What programming/technical projects have you been working on? in ~comp

    amelie
    Link Parent
    If you’re willing to go through the effort of labelling your dataset I’ve had decent success training Unets with relatively small datasets + data augmentation for image segmentation of blood vessels.

    If you’re willing to go through the effort of labelling your dataset I’ve had decent success training Unets with relatively small datasets + data augmentation for image segmentation of blood vessels.

    1 vote