Space Warps Talk

Estimated percentage of lensed galaxy finds

  • Keith_C by Keith_C

    I wondered what sort of percentage of images are likely to contain a lensed object? So far I have 11.9% find rate (528 out of 4436 images) with a simulated rate of 10% (Sims seem to be counted in potential finds) so is 1.9% likely to be accurate or am I providing too many false positives?

    Will participants get information on confirmed finds and numbers of false positives in order to improve their spotting skills?

    Many thanks
    K

    Posted

  • Tom_Collett by Tom_Collett scientist

    Hi Keith,

    Great question. I would say 1.9% is a pretty reasonable rate to be tagging the 'not simulated' images at. It's probably a bit more than the rate of real lenses, but there are plenty of 'maybes' that you should be tagging, even though they might not all be lenses in the end. If you aren't tagging the images that have already been checked by the science team (and don't have a lens), then you should be confident that you're on the right track.

    Clearly I'm getting a bit too excited - I've got a ~5% rate for finding something in a non-simulated image. That probably is too high.

    Tom.

    Posted

  • Keith_C by Keith_C

    Thanks Tom. I think my rate when I started was a little over zealous and has reduced now as I am seeing more sims. They definitely help!

    Posted

  • keith_edkins by keith_edkins

    Can you let us in on how this project works - if one spotter marks a candidate (but doesn't chat about it) does it get looked at by an expert, or does it take 2 or 3 spotters to make it interesting? And are the sims really typical of what we're looking for - some of them seem over obvious (like I don't think to myself "That's a lens", I think "That's a sim").

    Posted

  • Tom_Collett by Tom_Collett scientist

    I'm not totally sure how the expert analysis is working at the moment. Personally I'm looking at anything people are talking about. The plan is for at least 10 people to look at each of the 400,000 images. I suspect that good lens candidates will be tagged by most of the people who see them. As for the images where only one person flags it, I suspect there will be too many for the experts to look at them all.

    The best way to flag up good candidates is to create your own collection. Mine is here: CSWS00060j

    Regarding the simulations looking too obvious, I had similar concerns about the brightness and color but I've been talking to @anupreeta about this and she says they are indeed realistic; some are bright and obvious, but there are some simulations that are faint - and easily missed. The simulations don't include the very faint sources that may occur (@anupreeta didn't want to put people off by making the sims ultra hard to spot), but lensed sources can also be as bright and obvious as the sims (like ASW0000x1l). Whilst there may be fainter (or different color) sources in the real dataset but they will definitely follow the same general shapes as the simulated systems.

    @drphilmarshall has said this: (DSW00000df)

    Regarding the "cartoonish" nature of some of the sims: we are not
    painting these things in with our fingers, you know! 😉 We draw real
    galaxy sizes, colours and magnitudes from the CFHTLS catalog, and
    realize the source behind a model mass distribution that we know fits
    lenses well. [...] believe me: lots of real lenses do actually look
    like the sims! 😃

    Posted

  • drphilmarshall by drphilmarshall scientist, admin

    I'll write a blog post for you about how the analysis of the markers works. Basically we combine results from everyone who sees a given image to compute the probability of that image containing a lens, and when that probability crosses some threshold we count it as either a detection, or a rejection. Rejected images we "retire" to make space for more images, which are then automatically displayed. We set the detection threshold quite high, to keep the false positive rate down. Because we make decisions based on calculated probabilities, we can account for uncertainty in the classifications. No need to worry! 😃

    I agree with Tom, @Keith_C : it sounds like you on about the right track! 😃

    Posted

  • Keith_C by Keith_C

    Thanks Tom and DrPhil. I suppose the find rate will be pretty low as there are only 400 known lensing galaxies. I have certainly found a few lenses that are very obvious and I thought that they were sims but were not flagged as training images, so that's reassuring.

    Is there any way that a 'next' button could be added when viewing images within collections? This would make the job of looking through candidates a lot easier.

    Posted

  • drphilmarshall by drphilmarshall scientist, admin

    That's a great suggestion - I'll issue it to the Talk dev team. Thanks! 😃

    Posted

  • drphilmarshall by drphilmarshall scientist, admin in response to Keith_C's comment.

    OK, done - I realized while writing to them that there already is a next/previous button on the collection page, that shows you ten more images! I guess the problem is that these images are too small to be good for lens spotting. So, I asked whether the collection page could show just 2 images at full size instead, and then keep the same next/prev button set. Would this do?

    Posted

  • Keith_C by Keith_C

    Great. Yes, that would be brilliant. Thanks.

    Posted