Fairly Ranking the Most Brilliant Birds

Ryan Moulton，Ryan Moulton 关于搜索、技术与工程的文章。关注

After I finished the project of my last post I still had a giant collection of processed bird color data from BirdColorBase where every way I sliced it revealed new and mind-blowing birds. The minimal treatment I gave in that post just wasn’t enough. I had to dig in more.

I set out to find the brightest, most brilliant, most colorful birds. I’m not going to make you wade through an essay’s worth of math and discussion before I show you my top 40. Here it is.

This is ranked with math, but it is subjective math. If you disagree with the list, if you feel I have disrespected your favorites, or omitted some obvious contender, read on to see if I have treated your most beloved bird fairly, and what it even means in principle for a subjective ranking to be fair. I’m going to walk you through how I turned my subjective preferences into math. In the process we’ll touch on a number of common functions that are useful in information retrieval, but rather than ranking documents based on their relevance or quality, we’re ranking birds by their color.

Chroma

Where our story left off, we had found all the birds with plumage samples outside of sRGB and display-P3, the birds for whom at least some part of them is undisplayable on a conventional screen. This yielded pretty big lists of birds, so a natural next step to ask is, which are the brightest? What birds are behind all those dots at the edges? For that we need a notion of what it means to be “at the edge.” The simplest is the convex hull.

There are a few problems here. The first is that there are too few birds. I want to see more birds than that. The second is that there are a lot of regions of the boundary very sparsely populated with birds. I want to see some magenta birds, but because the space of bird color curves inward there, no convex hull can show me.

There is a clever generalization of the convex hull that can improve this, called the alpha shape. The gist of it is that rather than prohibiting any concavities in your hull, you allow concavities only up to a given size. If you can carve out a region with circles of radius more than 1/alpha without snaring any of the points in your circle, then you add the points on the boundary of the circle to your hull.

This is much better. We have a lot of birds evenly spaced on the boundary now, and know now where to find all the most extreme colors that birds can show us. As a ranking of the most brilliant birds however, there are still two problems.

First, this is showing the colors of individual plumage samples from birds, not full birds. If you are an entirely beige bird with one brilliant spot, you can easily earn a spot on this list. No offense to the fire fronted serin, its forehead spot is indeed very fiery, but one spot is just not enough. Or the cut throat finch, great Halloween costume and fantastic name, but we’re looking for a bit more than just a permanently bloody throat.

The second problem is that this is scoring colors based on their saturation, but not their brightness. A bird that manages to achieve the same color of green but reflects twice as much light is meaningfully “more green,” and I want the ranking to reward that. This plot filters out feathers that were too dark to have a meaningful color, but to really rank birds, I’d like to incorporate brightness in a more continuous way rather than just a binary threshold that means “definitely not black.”

The CIE colorspace has a transformation which extracts a component, chroma, that is very similar to what we call vibrance. It rewards both the saturation of the color, and the amount of light the color reflects, so this is a promising place to start. You might expect this to be some simple function of the quantities of light at different wavelengths, but it is quite complicated and abstruse. You first convert the spectrum to CIEXYZ, then CIELAB, then to CIElCh, using functions and parameters set empirically based on properties of the human eye without any intuitive interpretation. Before I found this I spent quite a lot of time experimenting with different ways to define the vibrance of a color using the more raw physical properties of the reflectance of the feathers. To quickly summarize all of the things I tried which didn’t work, let me leave you with an aphorism. If you find yourself reinventing color theory from scratch, stop.

Let’s let CIE do all our work for us and just rank birds by their average chroma.

Before we knock him off his perch, let’s spend a minute appreciating our current champion, the scarlet ibis. It is an avian stop sign, or perhaps a toddler eating a plate full of strawberries. It makes neotropical mangrove trees look like giant raspberry bushes. It is a very very red bird.

The scarlet ibis looks exactly like an American white ibis that found a truck full of cherry fun dip and was very excited about it. Taxonomists aren’t entirely sure whether to classify it as a different species. The two ibises have overlapping regions, interbreed, and are identical in every anatomical detail except for their color, but what a color!

This is a promising demeanor for a bird we might crown as the prettiest, but it does suggest a bit of overconfidence, a bit of resting on one’s laurels, perhaps a nest made of such. If you simply want to maximize the amount of color that a bird shoves in your eyes regardless of any other aesthetic qualities, stop here. Aside from being bright solid red from crown to tail, it is also coincidentally quite a large bird, sure to fill your binoculars. For maximizing raw color per unit bird, we have a clear winner.

Variability

Let me explain why I’m not stopping here, and want to iterate more. Consider the first two birds. If you look closely at the chroma plots, the Andean cock-of-the-rock has nearly the same hues as the scarlet ibis, but every sample has higher chroma. It comes in second place only because it also has black wings.

Similarly, compare the white bellied canary to the emperor fairywren, which it outranks by 18 positions. The emperor fairywren has higher chroma wherever it has color, but it has more black. Penalizing it for that doesn’t match how I think about the brilliance of a bird. Black among bright colors accentuates them by contrast, such as on an oriole or a goldfinch. If I’m given a choice between a bird with middling colors, and a bird with intense colors among patches of black, I would choose the half black bird to be more eye-catching every time.

To express this in math, I changed from a simple mean of the chroma values to a weighted average, where the weight itself depends on the chroma, . Setting different values of changes how much we care whether a bird has blacks or whites, where 0 treats the bird as if it didn’t have them at all. Varying changes how much I care if a bird has beiges or browns.

Setting too low yields lots of solid black birds with a few bright spots. Notice how almost all of them have red as their singular color. This will be important in a second.

After some tweaking I finally opted for as my weighting function, which gives the cock of the rock and the oriole a leg up among the top 5.

My second and more significant quibble with the pure chroma ranking is that the scarlet ibis is just one color. It’s a lot of color, but just one. I wouldn’t want to say that it’s impossible for a monochromatic bird to be the brightest or prettiest of all birds, but right now our ranking doesn’t even reward a bird for being multiple colors at all, in fact it hurts it.

Birds are very good at being red and yellow, due to the carotenoids that some birds transfer to their feathers from their diets. The CIE color space allows reds to have a high chroma even if they are very dim, and considers yellows to be very bright. This means that without any consideration for variety of color, the leaderboard will end up dominated by dim reds or tepid yellows. If a bird deviates from a pure red or pure yellow strategy, it is likely to do worse just because intense blues, greens, and magentas are harder for birds to make, and CIE expects more energy from them to raise the chroma.

It is unfortunate that our ranking has this sensitivity rather than being based on simpler physical facts. One might be tempted to work with lower level primitives like the raw quantities of light, but as I mentioned before, when you are tempted to reinvent color theory, stop. Defining the vividness of birds to eyes necessarily requires opinions about eyes. Most of those opinions are best left to CIE.

A natural concept to implement to reward birds with many colors over birds with just one is “diminishing returns.” Once a bird has a lot of red, you don’t want to keep giving it the same amount of additional credit for more red. You can instead say, “That’s enough red now, what else can you do?” This is naturally expressed by a power mean. The power mean is a generalization of the average that includes all of the other means you might be familiar with, like the geometric or harmonic means, as special cases. Rather than simply taking , you transform the values first, and then invert the transformation after the values are aggregated, . For a power mean, we let . The parameter of the power mean sets how min-like or max-like the mean is, with yielding the min, -1 yielding the harmonic mean, 0 yielding the geometric mean (after taking a limit), 1 yielding the normal arithmetic mean, and yielding the max.

To reward birds that have more variety of color I changed how I aggregate the chroma values across colors to use a power mean. All of the hues are unique, because the hue space is continuous, so before a power mean can have any effect, we have to either smooth or bin the hues. Normally in 1-d the natural method of smoothing the points would be to convolve them with a gaussian to produce a kernel density estimate, but here there’s a problem. Hue is a circle. This makes the infinite support of a gaussian inconvenient. You can’t easily ignore the tails of the gaussian without a discontinuity in the important areas, because it wraps around forever. The natural equivalent of a gaussian for points distributed over a circle is instead the Von Mises distribution, proportional to so that’s what I used. When the parameter is high, it converges locally to a gaussian around the chosen center, but is continuous around the circle. When is set to 0, it becomes a uniform distribution, so we recover the original hue-independent average chroma score that we started with.

The effect on the rankings of in the power mean and in the Von Mises distribution are very tightly coupled, which is a terrible situation for trying to tune functions, and especially for tuning them by hand. In that situation the best strategy is to fix the parameter that has a clearer intuition behind it, and only vary the more opaque parameter to achieve your desired effect. In this case, I have a clear intuition for how wide a range of hue makes colors similar to other colors, ~45° in hue space, but no intuition at all of how min-like to make the mean to match my aesthetics. I started by fixing at 10 to match my intuition about hue, and solely adjusted the power parameter.

Raising the power to greater than 1 selects only birds with single bright colors. Setting it to far less than 1 selects birds that cover the whole spectrum, a few of which I would like to highlight.

The Golden-Breasted Starling’s iridescent back and head span over half of the possible hues, just never very bright. This is a bird that can show you nearly every color there is, but just a little of it. It ultimately loses out in our ranking because it is dark enough to mistake for black in the wrong lighting, but deserves mention here for being exceptional in that particular way. If you ever find yourself needing a visual reference for science fictional metamaterials, perhaps the nanotech armor of an interplanetary supersoldier, the back of a golden-breasted or glossy starling would be an evocative place to start.

The Himalayan Monal has an astonishing range of brilliant iridescent colors, enough to rival or eclipse its southern neighbor, the peacock. While its colors were not bright enough to earn it a place in the top rankings of the final criteria, one wonders if in its natural habitat under the rarified air of high elevations, if it might be in practice much more brilliant than its lower elevation competitors. Living both at the top of the world, up to 14 thousand feet, and near the equator gives it much more sunlight to work with, a fact not captured by measuring all color relative to the standard illuminant. In effect I am telling you how bright each of these birds would look if you saw them next to each other in a zoo. I am not crediting the monal for the surrounding sunlight-flooded alpine valleys or the backdrop of brilliant white snow where it is found. Neither am I crediting it for the contrast with the austerity of its surroundings, nor the hypoxic euphoria that may accompany a pilgrimage to see it.

Finally, the Lilac Breasted Roller. I have once seen one of its cousins in person, an Indian Roller in Jaipur. The intensity of the blue and cyan on its wings took my breath away. I can only imagine the experience if its body were also pink as it is here.

While I did want to ensure birds with a variety of color had a fighting chance, I did not want to entirely exclude the monochromatic birds, so I finally opted for a of 0.85.

Finally, I was a bit skeptical of the birds with very few plumage samples. I don’t know how the researchers behind each dataset selected plumage to measure, so I can’t fully account for what it means to have a sparsely sampled bird. If they were only selecting the brightest patches to find the most significant colors, then a bird with 5 samples would have an undue advantage over a bird with dozens. The easiest standard way to fix this is to multiply by a saturating function of the number of samples, , where n is your sample count, and h is the number of samples at which you want the coefficient to reach one half. Given how useful this function is, it ought to have a standard name in this application, but I have not been able to find one. Any time you want to apply a conjugate prior to an estimate where the mean of the prior should be 0, but have no clue how to work out what that prior should be, you can just multiply by this function and be close enough to produce a working system. In many circumstances, such as a gaussian mean, this is exact, and it’s close enough in all others. Because so many beautiful birds were sparsely sampled, I set to a minimal value of 0.25. A bird with a single sample would only have its score reduced by 20%.

We aren’t quite done with our ranking, but we are ready to crown the champion, the orange breasted bunting. Its belly is as yellow as any bird is yellow. Its back is as turquoise as any bird is turquoise. It has a smattering of intense greens in between around its head and back and orange around its neck, which lets it fully fill half of the color wheel.

The orange breasted bunting lives only in the tropical dry forests of Mexico’s southwestern coast. Like the California forests to the north these forests endure many months of seasonal drought, but unlike in those forests, to survive the drought most of the trees lose their leaves. This gives the orange breasted bunting a similar visual role to its northern relative, the cardinal. It fills brown and bare winter branches with color.

You might notice the abundance of parrots in this list. There is a chemical reason for this. Most birds have to extract carotenoids from their diets to make red and yellow. Parrots, uniquely, can synthesize a different pigment themselves. The molecule they make is so singular to their family that it is named after them, psittacofulvin, literally “parrot yellow.” It can reflect either brilliant yellow or deep red with a tiny chemical change to its end enabling a wide variety of brilliant color patterns between species. The many species of mostly green parrots make their characteristic green by combining that yellow pigment with structural blue.

This is a compelling case for why the ranking is dominated by mostly-green visually similar parrots, but it is still a bit dissatisfying. Perhaps you can tell at a glance the difference between an orange-winged amazon, a yellow-crowned amazon, a white-fronted amazon, and a blue-fronted amazon but I cannot. Perhaps the sapphire-rumped parrotlet differs more from other green parrots than it seems to, but I have been sadly unable to find an adequate picture of its rump. Once I’ve seen one of these parrots, I’m unlikely to get very excited if I see another. They all kinda look the same. If you want the canonical ranking of the most brilliant birds to encompass more of the variety that birds have to offer, we need to include one final factor.

Diversity

Any encyclopedia will say that, after Mount Everest, the second highest peak on earth is K2, another mountain in the Himalayas. It actually isn’t; It is only so by convention. The second highest peak is a nameless rock a centimeter away from the summit of Mt. Everest. Same with the third highest peak, and the fourth, and so on. We ordinarily only count a peak as a peak if it is more prominent than any of its immediate neighboring peaks. This convention, or definition, prevents us from having to fill the list of the top thousand mountaintops with various boulders on Everest.

This problem isn’t unique to ranking peaks by height, it is inherent to ranking anything, especially when the number of items you are drawing from is large relative to the number of items you are selecting. Your ranking function probably has a single highest point within the natural distribution, its own Mount Everest. If your dataset is large enough, there are probably many points on Mount Everest, all very high, all very similar to each other. By default they will dominate your ranking. Everything on your list will become either an exact copy or a small variation of the thing you decided was best. The two tallest people in a high school are likely to be dissimilar from each other aside from being male. The two tallest people in the world probably both have pituitary gigantism.

In a situation like this, where the entities we are ranking are not themselves optimizing to be ranked highly, this only becomes a problem when the number of entities is large enough to fill out the whole space of possibilities. If the entities you are ranking are responding to your ranking, and optimizing for it, it can happen even when you are ranking very few things. Enough of them will find the dominant strategy that only those who use that strategy will rank.

This is a lot of pressure to put on a ranking function. If your function is wrong in some way, if there is some gap between what is actually best and what you defined to be best, then you’ll end up with the list totally dominated with the wrong things. If there is a strategy that breaks your function, it won’t just be broken at one position, it will be broken across the entire list. There will be nothing left of the unoptimized entities with an ecosystem of varieties and natures that you were hoping to reward.

We saw this problem acutely in some of our prior rankings. Our initial list had many solid red or solid yellow birds. When we reduced the weight of black and white samples to 0, we found that the list was dominated by mostly black birds with one brilliant red or yellow spot. When we began rewarding diversity of color within a bird, it became dominated by various mostly green parrots.

You might be unbothered by this, and prefer a list ranked purely independently, even if it ends up being mostly similar parrots. But what would you do with exact or near exact duplicates? We’re fortunate that “species” is a relatively crisp definition so we don’t have to worry about thousands of genetically similar and visually identical birds cluttering up the ranking, but what about gender? Do we count males and females of a species as separate birds in the ranking even if they look identical, such as Scarlet Macaws? What about subspecies? What if the subspecies look visually different? How do you decide whether they’re different enough? Make the wrong choice and you can end up with your entire list dominated by minor variations of the same bird.

To prevent this for mountain peaks, we discount a peak if it is very close to a more prominent neighbor, and this same motivation underlies algorithms for diversifying a ranked list, such as Determinantal Point Processes, or Maximal Marginal Relevance. If we could do so in a robust way, we wouldn’t need to worry about whether we had divided our species too finely. The duplicates and near duplicates would be ranked lower just because they were too similar to a previously ranked bird. We could throw the whole distribution into the ranking, and let the mutual similarity penalties sort out the granularity for us.

The algorithm I implemented here is a variant on maximal marginal relevance. Let be the score of item when considered as a candidate for ranking in position . Let be the original undiversified score, and be an index over already ranked items. Let sim(i,j) be a similarity measure of the item with index i and j, here a cosine of the smoothed hue-chroma vectors.

The key difference between this and maximal marginal relevance is that the diversity penalty decays. As the number of items in the list grows, the number of candidate higher ranked items also grows, and so under the original maximal marginal relevance the nature of the ranking function changes dramatically in character as you descend. It becomes dominated by diversity rather than the original score. By allowing it to decay we can retain similar character throughout the list, and have every position contrasted with a similar number of peers. The motivation of the rank-dependent exponent is to space out items according to how similar they are. Low exponents penalize an item for being weakly similar to an already ranked item. High exponents only penalize near duplicates. By basing the exponent on the difference in ranks, we ensure that moderately similar items will be spaced out in the list, but won’t be ignored entirely, while near duplicates will largely be excluded because their diversity penalty will never decay.

We fill out the list greedily, at each point choosing the maximum score among the remaining candidates. Because the diversity penalty only lowers the score, we can consider candidates in order of their original score, and stop early when the score of our current candidate exceeds the original score of the next, preventing runtime for filling each rank.

The original champion, the orange breasted bunting, is the same. In position 1 it is unaffected by the diversity penalty. But note how different the rest of the list is, and how different their strategies are. We again have a lot of single colored birds, but this time they are each a different color.

The top 10 now newly includes the plum throated cotinga and the asian green broadbill, two truly exceptional representations of their respective colors. Two parrots remain, the red winged parrot and scarlet macaw, which are much more compelling and varied than a long string of various green amazons.

If these don’t seem like the most brilliant birds to you, and you think you can do better, my working colab is available to you. See what you find.

Fairness, Judgement, and Control

The techniques I am describing in this post, the power means and priors, the feature vectors and parameterized similarity measures, are largely a lost art. If you are a professional designing a ranking function today, you are unlikely to be building it by hand, using only your judgement, as a function of measured statistics about the items to be ranked. This is a very old way of doing things, prehistoric in computing timelines, the sort of process that Larry and Sergey went through when they were writing Google’s first ranking function, and that Amit Singhal went through when he was rewriting it. No one does things this way anymore.

Instead, if you are operating in an organization that is fresh enough to be “a dinosaur” you might be using traditional machine learning or statistics to predict a set of data labelled by human beings, perhaps a regression model predicting pairwise rankings of “which bird is prettier.” If your organization is merely “behind the times” you might be training an embedding model to summarize bird color, asking an LLM “which bird is prettier” rather than a human (or doing that by accident because your contractor workforce has silently outsourced their jobs without telling you) and fine tuning the model to the LLM’s preferences. If your organization is fully “meeting the moment” or “ahead of the curve” you are probably just asking a language model what it thinks the prettiest birds are, and shipping it to prod without even looking at the list yourself. If you are especially diligent you might iterate on “prettiest” vs “brightest” vs “most colorful” in the text you give it, and pray that the next model point release doesn’t have new idiosyncratic beliefs about birds that break your product, such as an inordinate and inexplicable fondness for corvids.

The new ways replaced the old ways for good reason. They perform much better, especially when you train them to reproduce the metric you are measuring them by. They are very fast. I iterated on these expressions and tinkered with their parameters off and on for a month, when a model could have produced a result in minutes.

Most importantly though, they are disempowering. You cannot get the result you want. You will instead get the result your data wants. This sounds like a disadvantage, but it is often not. When you base your model on data, you are outsourcing your product judgement to someone else. If that “someone else” is your users, and you are measuring their judgement correctly, this is a virtuous cycle. Unfortunately, organizations will converge to believing they are in that virtuous cycle, even if some critical part of it has broken down, a special case of Goodhart’s Law, and the actual proximate cause of what the most relentlessly cynical people on the internet call enshittification.

For a system to continue to serve its goal as its complexity increases, your signals must be uncorrelated with the residual between your goal and your measurement of it. Your signals have to be so pure and so good at their job that they are incapable of reflecting the manner in which your measurement of your goal is wrong. One might call this risk “overfitting” but at the organizational rather than model level. For a model, a held out test set that the model never saw can allow the human beings training the model to judge which model is better. For an organization, the held out test set is reality, and the judgement is the success of the product.

Even if an organization is appropriately skeptical that their measurement of their users can substitute for product judgement, there is one remaining advantage of outsourcing it to them that makes it irresistible. It gives you deniability for your decisions. Politicians like deniability. The public likes deniability. The courts like deniability. The courts may even decide that deniability is the same thing as a lack of liability, and save you a lot of money. Organizations like deniability. It is repeatable, predictable, durable, and functions independently of who is employed to work on it. Your own sense of moral responsibility may like deniability. “What could I do? It’s not my fault. It was a pattern in the data.” Deniability has much to recommend it, but at the end you are putting your product in the hands of someone else, at best the public, and, well.

If you are writing the function yourself, and setting the parameters yourself, you can’t hide behind anyone. All of the decisions it makes are yours. If its decisions have ethical implications, those implications fall on you. That’s scary! How can I take responsibility for such a thing? People are going to get mad at me, and what can I say?

If you’re a scarlet ibis upset about your ranking and speaking on behalf of all the solid red birds, you might look at what I’ve written above and say “SEE? MALICE! He designed the ranking to hurt scarlet ibises! Being covered in a single brilliant color makes me better, not worse! Birds should be ranked by the single thing they’re best at, not secondary qualities that don’t practically matter.” Or if you’re a white bellied canary speaking on behalf of yellow birds, you might think it unfair that I was skeptical of how bright they all are, because there are so many of them, and they’re all bright in the same way. You might insist that birds should be ranked by lightness, not chroma. “Lightness has a 0.45 correlation with chroma! Using chroma instead of lightness is unprincipled sentimentality. It’s light that makes your eye respond, not color. Having a saturated color just means you’re missing some light! Conditional on being in the top 100 birds the residual between chroma and lightness has no correlation with subsequent outcomes!” Or if you’re an orange winged amazon speaking on behalf of all the well rounded green parrots with a few color spots, you might say, “Why does it matter how many green parrots there are at the top of the list? If an objective measure says that green parrots are best, then top of the list should be nothing but green parrots. That’s just how it is. It isn’t our fault that we can make pigments that nobody else can. Any diversity measure is just giving a spot to a dimmer bird that doesn’t deserve it.”

Certainly any statistical test you do would agree, that yes, the adjustments I made hurt the birds that I said they did and that I was motivated to rank lower. If any of the phrases I used in this article ended up as evidence during discovery or as a pull quote in the New York Times it would be incredibly damning. And yet I still contend that the ranking function I have designed here is fair, even though it has abundant biases in both intent and output, and even though it would fail every ordinary test of what fairness in a ranking function means.

When I express a desire to modify the ranking to reduce the rank of some particular bird that I think is overranked or to modify the composition in some systematic way, what constrains me is being able to express that intuition in math. I want to be clear though about what the actual constraint here is. You can express nearly anything in math. Writing a function that achieves any corrupt end you want is easy to do. Traditional machine learning algorithms are happy to overfit whatever data you provide. If you have bad opinions, and want to torture your math to express those opinions, you will absolutely be able to do it. The triviality of fitting an arbitrary function to an arbitrary set of biases is what motivates the standard statistical fairness measures that treat the underlying algorithm as a black box.

I am ignoring that literature here, because I don’t think it maps well to what I believe it means for a ranking algorithm to be fair. Instead I contend that this algorithm is fair in a deeper way than those criteria would guarantee. If I was just unreasonably unimpressed by the brilliance of parrots, and was putting my thumb on the scale to rank them lower at every opportunity, the primary factor that constrains me is the ugliness of the math I am writing. My aversion to doing gross hacks is as much of a constraint as my moral obligation to treat parrots even handedly. You might have noticed my fondness for the himalayan monal and my frustration with how poorly it ranked regardless of my parameter choices. My inability to come up with any elegant criteria that prefer it constrains how my preferences influence its ranking more than the soundness of my preferences. It is not simply a matter of expressing opinions in math, it is math that has good taste.

I have defined here a parameterized class of functions, where any function in that class represents a reasonable opinion that someone could have about what makes a bird beautiful. Because the class itself has this property, all functions in that class are fair. The fact that I chose the parameters to specify a function within that class to match my own subjective preferences, or to rerank the concrete examples I considered, doesn’t degrade that. All the functions in the class are fair, so this one is fair too.

To enumerate what I believe makes it fair:

Each component of the function implements a widely held intuition about a quality that can make one entity better than another, in this case, “It’s brighter,” “It has more colors,” “we know more about it,” and “it’s comparatively unique.”
It implements these intuitions in a natural, straightforward way.
It bases each component on measured data without regard to the identity of any of the entities.
The components of the function are combined in a smooth, monotonic, simple fashion. There are no discontinuities, and especially no discontinuities suspiciously close to the scores of any particular entity.

Note that these qualities are all violated by traditional machine learning, in which most of the functions in the class you are fitting to data are unjustifiable gibberish. If you have little control over what the model is computing, you have no guarantees that it will match any human intuition. It usually will not. If the math I’m using were not constrained by good taste, I could extract the parrot hue/chroma vector principal component, simplify it to a few parameters, and ensure that no parrots ranked highly by subtracting its projection from their score. This is not substantially more complicated than what I have already written, and has a similar number of parameters. I wouldn’t do that because it is gross, and that grossness makes it unfair. Ordinary machine learning systems do this sort of thing all the time. It is the primary thing that they do.

Fairness to me means being able to look at and describe the motivations for every part of the ranking, justify them as reasonable criteria to rank one thing over another, and never feel that what it is computing is some BS. The reasons for choosing one bird over another have an intuition behind them, and intuition that is understandable and relatable in general, not just to me, and then that intuition is converted into math. In the ideal case the algorithm is so clear and well motivated that you could explain it to someone who is harmed by it, and have them react, “Eh, well fair enough I guess.”

The losers should feel that the reason they lost is both clear and legitimate, even if they disagree. This is a high bar for a ranking function to clear, but it is a better goal than any other that has been proposed. People are ultimately mostly reasonable. If you can explain to them why they were harmed, and the basis for it seems to them to be legitimate, they are likely to accept it. They are more likely to accept it even if it harmed them more than some alternative ranking with weaker and less justified criteria. It is worth sacrificing some amount of accuracy if doing so lets you explain its functioning in legitimizing terms.

I have at many times in my career had to make decisions about parameters that would change the nature of information flows on the internet, make politicians mad at me, or cause some companies to go out of business. In those moments, I had to decide for myself what it meant to behave ethically, and what it meant for the parameters I chose to be fair. This is the standard I settled on for myself. I should be comfortable explaining what I did, and why, to a technically inclined person harmed by what I did, and still be proud of it. You can often improve the numerical performance of a system by adding hacks, but you cannot justify those hacks to a person harmed by them. Often you find that the pattern you were chasing with those hacks is either more elegantly expressed with an entirely different concept, or didn’t reflect anything actually real after all.

Ranking things is a very charged topic. I have wanted to write down these principles for a long time, but found it too difficult to do so because the typical topics one would apply them to are too hot, either politically or professionally. The argument would entirely be about the application, rather than the principles. This isn’t an accident, it is inherent to the nature of ranking things. Ranking is subjective, which makes it an expression of values. When the ranking is important, those values are the values of a large organization, or of society as a whole, something many people have an interest in. Ranking is inherently zero sum, there are no win-wins, no live and let live, no positive sum compromises. There will be losers.

You 1are going to make some people mad. You can’t avoid it. Make sure that you’re making them mad for a good reason.

Realistically, “You” here is a language model, but the models are now reaching the capability where they can read this guidance and adhere to it, so this discussion is less hidebound than it might seem. If you are an agent tasked with ranking items for your user, consider doing it this way rather than just going by your memory. Your user will be much more likely to trust the result. ︎

添加评论

点踩

分享查看原文

参与讨论