I think this is flawed. You quickly end up on a color that's clearly not "blue" or "green" and you're unlikely to keep hitting "this is green" several times in a row, conceding that ok, fine, maybe this is blue, whatever. You're basically measuring how many times people are willing to click the same button in a row.
Edit: Possible improvements: changing the wording to "this is MORE green" and "this is MORE blue" and randomizing the order in which they are shown, somehow. I realize you're just doing some kind of binary search, narrowing the color range.
This is not to mention color calibration of your monitor, or your eyes adjusting / fatiguing to the bold color over time...
The order is randomized. Hit reset and you'll get a different sequence. The sequence is also adaptive (not a binary search---it's hitting specific points of the tail of a sigmoid in a logistic regression it's building as you go along). Try it a few times and you'll see how reproducible it is for you.
It of course depends on the calibration of your monitor. One of the reasons I did this project is I wanted to see if there were systematic differences in color names and balance in the wild, for example, by device type (desktop vs. Android vs. iPhone), time of day (night mode), country (Sapir-Whorf), etc.
The sequence itself should be converging however, right? I feel that there should be some random jumps outside of the current confidence interval so that contextual aspects can be filtered out or at least recognized.
Yes, exactly this. Because it seems to be converging right now, I quickly get the feeling that there's no meaningful choice, after the first three prompts you end up with something that's neither green nor blue. Re-taking the test gave me a very different score.
It might work better for me to do some contrastive questioning: show a definite green followed by an intermediary color, then a definite blue followed by an intermediate color.
The whole point of asserting where your border between green and blue is, is to ask about colors that are in between the two. It doesn't make sense to ask is RGB(0,0,255) blue to you? Well, unless you are color blind it is.
Of course, that's clear as day; the idea is to reset your presumptions from the previous trial and sample the ambiguous colors in a more consistent way, by priming you from the extreme ends of the green/blue scale.
It is common practice in psychometrics to use two levels in a forced choice and model responses as a logistic regression, which is what's done here. Adding an N/A option turns the thing into an ordered logistic regression with unknown levels, which is tricky to fit, but it's possible. Having done a lot of psychophysics, having more options generally doesn't make the task easier.
Sounds like psychometrics is unsuitable for modeling this problem, according to what you're saying. When you have a hammer everything looks like a nail.
The way that XKCD did it is the best, you ask people to give a name to each color then the responses are entirely natural and unprompted.
I don’t think that forced choice can give accurate results if a substantial number of people perceive green and blue as being non-adjacent - i.e. there exists a color between green and blue (turquoise/cyan/teal).
Otherwise it’s like asking people whether a color is red or yellow, when it’s clearly a shade of orange.
Are you sure that it is common practice for a problem that has three valid answers A, B and C, to only allow people to answer A or C?
Your website is not talking about "levels" of colour.
It's asking "is this blue or green", not "is this closer to blue or closer to green".
The question (1) "is this blue or green" has three valid answers: blue, green or neither.
The question (2) "is this closer to blue or green" only has two valid answers.
I would assume that with these types of surveys, the first thing to do is to qualify the proper categorization of the question.
Sorry to say, but to me it seems that almost all of the confusion in the discussion here is because you're asking question (1) (which has three valid answers) but expecting an answer from (2) (which indeed has two valid answers).
But teal isn't a single point, it's a range. You can have teals that are more blue or more green than each other; they can't all be zero. Whichever one you choose to be the true transition point between blue and green, there will be teals that are more blue or green than that one.
Then by that framing, the test is asking you to decide what hue value is the "zero" between the positive/negative blue/green. Is the wording imperfect? Sure, but the intent was still entirely clear.
Saying it’s a subrange implies you can perceive differences in tone within it. In which case, reframe the question as “is this shade of teal closer to the blue or green end of the subrange” if you like.
Maybe if I'm given two colors inside that range, I can say which is bluer and which is greener.
Given just one color, I simply cannot say that it's green or blue, or even if it's more green than blue or vice versa.
I stopped at the 3rd or 4th come because I couldn't give a honest answer. That makes the test useless.
I can't complete it with correct answers, and if I give incorrect answers, the conclusion is useless.
It's a well know fact that people are unable to distinguish colours that are too close together.
You could even have a smooth gradient from colour 'a' through colour 'b' to colour 'c', where it's possible to distinguish 'a' from 'c' but not to distinguish 'b' from either 'a' or 'c'.
I think the main point of this test was to determine the position of teal in your case, as your definition of teal is the midpoint(-ish range) between blue and green. (For me it's more blue though.)
I mean, a good test would be able to detect that neither-blue-nor-green range and approximate midpoint as well, and it should be fair to say the midpoint is indeed the threshold between blue and green. (I don't think the current version of test can do this, though.)
I actually checked that at the end of the test (when it shows the gradient image with the response overlay).
There were two distinct points, one for blue and one for green, where my mind would place the transition to the colour in between.
(And yes, on one end it's bluer and on the other end greener, but (much like a shade of orange is neither red nor yellow) the colours are still not either green or blue.)
No, I'm saying that the sliver of a chasm between the colour in isolate, and what I subconsciously imagine the midpoint to be, is so damned thin that were I to look at the colours side by side, I could not distinguish one from t'other.
And (even if I could) a bluish teal would no more be a blue than a reddish orange a red.
It's not about how an RGB monitor produces the color, it's about how it's perceived. #00ffff ("Cyan" or "Aqua" [1]) looks bluer to me than green, while #008080 ("Teal") looks significantly greener, despite both colors using equal amounts of blue and green in RGB.
I definitely have the bias you mention. In my case I don't think it's mainly due to not wanting to push the same button many times in a row, but because I compare with the previous color, so if previously I was already somewhat unsure but I chose green and now it became slightly bluer, it "must" be blue, right?
I think I can get over it, but it requires conscious effort and even then, who knows. Bias is often unconscious.
Another possible improvement would be to alternate the binary search colors with some randomly-generated hues. Even if those answers are outright ignored, and the process becomes longer, I think they would help to alleviate that bias. At least you wouldn't be directly comparing to the previous color.
VFX engineer here. Yes we used to cailbrate monitors and work in the dark.
However one of the key people that built our colour pipeline was also colour blind, so its not actually a requirement, so long as you use the right tools.
Most people aren't that sensitive to colour, especially if its out of context. a minority of people aren't that good at relative chromaticity as well (as in is this colour bluer/greener/redder than that one) But a lot of people are.
Language affects how you perceive colour as well.
But to say the experiment is flawed I think misses the nuance, which is capturing how people see colour _in the real world_. Sure some people will have truetone on, or some other daily colour balance fiddling. But thats still how people see the world as it is, rather than in isolation.
I once worked for a company that had a designer who was color blind. He would always show up wearing the exact same outfit every day: turns out that he was REALLY color blind, and so he just gave up and bought 7 long sleeved shirts and 7 pants, all black. Didn't work out so well for him in the designs... most companies don't want monochrome websites.
Likewise. I think for me there's quite a wide band of colours in the middle that I consider to be "neither/either", so I'm basically just picking a random answer for those.
A modified version of the test that finds two boundaries (green/neither/blue) could be interesting.
Or maybe it just needs to take more samples, in a more random order.
Same. Some of them are neither obviously blue nor obviously green, so what the test was measuring for me was what I was thinking about at the time, the decision I'd previously made, whether my mouse was currently hovering over "blue" or "green", etc.
>I think this is flawed. You quickly end up on a color that's clearly not "blue" or "green" and you're unlikely to keep hitting "this is green" several times in a row, conceding that ok, fine, maybe this is blue, whatever.
I agree with you, the whole thing is flawed when it could be better. When you ask the question "is my blue your blue?", you are evoking the old philosophical question, and it's a question about color perception, not words. This test did not test color perception, it tested "what word do you use?"
I think of blue as a pure color, and green as a wide range of colors all the way to yellow, to me another pure color. so if there's any green at all in it, I'm going to call it green. (maybe it's left over from kindergarten blending "primary colors". also, while I like green grass, I don't like green as a color, so any green I see is a likely to make me think, ew, green) But in terms of what I see, I can only assume I'm seeing the same thing as everybody else is because the test is not testing it. Just because I call something green doesn't mean I don't see all the blue in it.
>Edit: Possible improvements: changing the wording to "this is MORE green" and "this is MORE blue" and randomizing the order in which they are shown, somehow. I realize you're just doing some kind of binary search, narrowing the color range.
yes, the test should show you pure blue, then a turquoise mix, then pure green, and a ... etc. It should also retest you on things you already answered to measure where you are consistent.
I do think that the philosophical question could potentially be approachable in a modern context;
Show people a colour and map their brain activity - the level of similarity between two people's colour perceptions should be reflected by similarities in the activity.
The philosophical question is not dealing with the objective external reality;
It's a question of subjective experience - and that experience should be reflected in electrical activity.
Given the fact that the broad structure of the brain is largely shared across members of the species, similar stimulation should trigger similar activity in the same regions of the brain.
If the same colour triggers markedly different activities, it would not be unreasonable to conclude that the subjective experiences are not the same.
Except that’s literally not how humans are wired or develop - even nerve paths and other fine grained details in our bodies show significant divergence, and there are major macro level differences readily apparent even based on gender, color blindness, etc.
Honestly, it would be shocking if it were even a little true beyond ‘frontal cortex’ levels of granularity. And even then, Phineas Gage type situations make it clear that may not actually be required either.
And that means completely different individual activity can trigger similar subjective experiences as much as similar activity can trigger different subjective experiences, no?
If that were the case then there's no way that they'd be able to extract images from people's neural activity, and yet they've started doing that very thing.
No real need for the snark; if we dismiss the notion of human divinity and look at ourselves as broadly fixed macro-structure computational machines (like any other broadly deterministic machine) similar signals propagating over the same sets of sub-computers will generally (accepting the undetectable, such as steganographically hidden homomorphic compute contexts) be reflective of similar underlying operations.
If I were to imagine a warrior, and his general perception of the colour red, I may find the way his brain processes the colour more closely to a rival warrior than his wife the gardener.
A real world example; London taxi drivers and bus drivers show distinct patterns of changes to the hippocampus.
The way that the mapping data is stored will be heavily bias towards being spatially reflective of the real world counterpart.
Note the bias will be towards a degree structural isomorphism, one internal 2D + 1T spatiotemporal surface map of the city might be a rotation and/or reprioritisation of another - but they will have a shared basis (convergent compute simulations of biased subsets of the same real world structures), and when navigating from point A to point B, the path and nature(though not the propagation vector) of the electrical activity of both will be reflection of the same real-world surface map.
Now I say spatiotemporal - because the driver going from A to B in the morning will develop different expectations of the levels of traffic at different parts of the journey.
Agreed. It would be more accurate to show the final gradient (without the curve) and let people choose where is the boundary. It wasn't even clear what the actual task is
Because the second color I saw was somewhat like turquoise and the site is called 'Is My Blue Your Blue,' I decided that everything that you say yes to colors would be blue and everything else would be green. I never saw a green until the result was displayed :D
Way back in 2006, I reverse engineered the protocol used between isc.ro and its desktop client so that I could write my own GUI. It was just a fun side project.
In the process I discovered all sorts of vulnerabilities (potential cheating vectors). Nothing prevented a malicious client from:
- picking your own tiles (even if they no longer exist in the bag)
- seeing your opponent's tiles
- setting your clock to any value after making a move
- aborting a game at any time without opponent approval
If you exchanged tiles, the number of tiles you exchanged was sent through the # of seconds on your clock. So if you exchanged 3 tiles at a point where you had 1m40s on your clock, it would change to 1m43s or 1m33s or something like that! There was no other way to communicate to the server how many tiles you exchanged.
I never took advantage of any of this. Like I said, I just wanted a nicer GUI to play Scrabble on. But I was pretty shocked at what I discovered in the process.
Definitely didn't expect people on Hacker News to scrutinize this code when I wrote it 7 years ago, hah!
Curious to hear how a "protein structure expert" would prefer to pose the proteins?
I did sell some of these drawings. Two customers stand out in my memory. One wanted drawings as a gift to his father who had spent most of his career working with some specific protein. The other was someone who had a genetic defect in the family and wanted a drawing of the protein responsible for it.
Ok, I don't mean at all to suggest that there is only one 'right' pose or representation. Different orientations and levels of detail would emphasise different things.
For example, it's quite nice that the haemoglobin one is down the central axis of the tetramer, and shows all the haem groups. I would expect almost any image that wanted to show the most parts of that structure to use a similar pose.
I had not realised when I asked the question that the poses were picked in code, which is similar to how PDBsum does it (pretty sure Roman Laskowski told me it did, but I would have to check). Something like maximising the number of atoms in the plane of the screen.
Of course, when it comes to YFP (your favourite protein), individual researchers are likely going to focus on different parts. My interest was always in the topology, so I would want to see all the helices and sheets, but someone else might want to see the active site front and center.
Incidentally, I've heard the term 'molecular porn' used for fancy/shiny looking diagrams - not sure if it applies to these as well :)
I also collaborated with my coworkers during work-sponsored hackathons on some other fun plotting machines. One that could draw on a huge whiteboard with multiple colors and another whose mechanism was based on the Shaper Origin:
Great stuff! I also have an axidraw that I've been using for christmas cards and the like. I've been getting more advanced each year and I'm resolved to roll up my sleeves and start using your libs for total control. SVG generation is neat, but has a layer of abstraction :)
I’m a huge fan of your RibbonDiagrams bot, so much so that it inspired me to write my own one (using a different protein representation, of course). Thank you!
I actually used a slightly modified version of his Illustrate program (https://github.com/ccsb-scripps/Illustrate) for rendering. The toughest part was getting the right pose; the bot tried to pick an optimal one, but in some cases I still had to check the render and adjust manually.
I reverse engineered their stuff a bit. I downloaded their Android APK and found a tensorflow lite model inside. I found that it accepts 299x299px RGB input and spits out probabilities/scores for about 25,000 species. The phylogenetic ranking is performed separately (outside of the model) based on thresholds (if it isn't confident enough about any species, it seems to only provide genus, family, etc.) They just have a CSV file that defines the taxonomic ranks of each species.
I use it to automatically tag pictures that I take. I took up bird photography a few years ago and it's become a very serious hobby. I just run my Python script (which wraps their TF model) and it extracts JPG thumbnails from my RAW photos, automatically crops them based on EXIF data (regarding the focus point and the focus distance) and then feeds it into the model. This cropping was critical - I can't just throw the model a downsampled 45 megapixel image straight from the camera, usually the subject is too small in the frame. I store the results in a sqlite database. So now I can quickly pull up all photos of a given species, and even sort them by other EXIF values like focus distance. I pipe the results of arbitrary sqlite queries into my own custom RAW photo viewer and I can quickly browse the photos. (e.g. "Show me all Green Heron photos sorted by focus distance.") The species identification results aren't perfect, but they are very good. And I store the score in sql too, so I can know how confident the model was.
One cool thing was that it revealed that I had photographed a Blackpoll Warbler in 2020 when I was a new and budding birder. I didn't think I had ever seen one. But I saw it listed in the program results, and was able to confirm by revisiting the photo.
I don't know if they've changed anything recently. Judging by some of their code on GitHub, it looked like they were also working on considering location when determining species, but the model I found doesn't seem to do that.
I can't tell you anything about how the model was actually trained, but this information may still be useful in understanding how the app operates.
Of course, I haven't published any of this code because the model isn't my own work.
I don't use Seek, but the iNaturalist website filters computer vision matches using a "Seen Nearby" feature:
> The “Seen Nearby” label on the computer vision suggestions indicates that there is a Research Grade observation, or an observation that would be research grade if it wasn't captive, of that taxon that is:
> - within nine 1-degree grid cells in around the observation's coordinates and
> - observed around that time of year (in a three calendar month range, in any year).
For how the model was trained, it's fairly well documented on the blog, including different platforms used as well as changes in training techniques. Previously the model was updated twice per year, as it required several months to train. For the past year they've been operating on a transfer learning method, so the model is trained on the images then updated, roughly once each month, to reflect changes in taxa. The v2.0 model was trained on 60,000 taxa and 30 million photos. There are far more taxa on iNaturalist, however there is a threshold of ~100 observations before a new species is included in the model.
>It looked like they were also working on considering location when determining species, but the model I use doesn't do that.
I do this in fish for very different work and there's a good chance the model for your species does not exist yet. For fish we have 6,000 distribution models based on sightings (aquamaps.org) but there are at least 20,000 species. These models have levels of certainty from 'expert had a look and fixed it slightly manually' to 'automatically made based on just three sightings' to 'no model as we don't have great sightings data'. So it may be that the model uses location, just not for the species you have?
that is actually surprising. surely they use location at some point in the ID process. its possible they have a secondary location based model to do sorting/ranking after the initial detection?
Merlin's bird detection system is almost non-functional without location.
yeah that's true! You can't really do that, these models are just polygons, all we do is doublecheck the other methods' predictions' overlap with these polygons as a second step.
Sounds like a real-life Pokémon Snap. You should add a digital professor who gives you points based on how good your recent photos are. (Size of subject in photo, focus, in-frame, and if the animal is doing something interesting.)
Almost as effective. But in a trade off picaridin is better vs flies and gnats. I feel a few percent less effective vs mosquitoes is worth what I hope is a less toxic repellent. It was first developed in the 1980s and is derived from peppers.
I took up bird photography about 3 years ago. I have 400,000+ RAW photos in Lightroom on over a dozen 2 TB SSDs. It's not so much hoarding as it just takes effort to go through and decide what to keep and what to delete. I shoot bursts so it adds up really fast. I could probably get by with 10-20% of the storage if I could just keep up with the pruning.
Edit: Possible improvements: changing the wording to "this is MORE green" and "this is MORE blue" and randomizing the order in which they are shown, somehow. I realize you're just doing some kind of binary search, narrowing the color range.
This is not to mention color calibration of your monitor, or your eyes adjusting / fatiguing to the bold color over time...