Welcome to the following installment of our Pigeon Artwork Undertaking.
At present, we’re going to get into the information, however first, let’s recap the challenge for many who are usually not acquainted. I direct the Comparative Cognition Lab at UCLA the place my college students and I examine animal cognition, primarily in folks, rats, and pigeons. Our pigeon analysis consists of coaching pigeons on “video games” introduced on a touchscreen. They be taught what to peck at for meals, after which we will take a look at all types of studying, reminiscence, and cognition.
In Might 2021, we found certainly one of our pigeons, Darwin (who can be a Netflix star!), had opened Microsoft Paint and made some paintings! We determined then and there to lean into the examine of pigeons making digital art. (I write concerning the origin of the challenge here. And yow will discover all my articles on the Pigeon Artwork Undertaking at my Medium page.)
Since then, dozens of pigeons have made a whole lot of digital artworks, and we’ve offered lots of them as NFTs. See our Genesis drop, an edition of 40 on the Solana blockchain (nonetheless minting), an open edition on the Polygon blockchain, and our present Stained Glass drop which is still minting.
Let’s get into the information!
It’s been a couple of months since my last post. However we now have not been idle! We’ve been creating our analytic instruments to reply some questions on pigeon artwork that I do know you’ve been dying to ask.
And, our pigeons have continued to make nice artwork! However how can we analyze such magnificant paintings?
As with every science challenge, the questions information the analytics. So, what are our burning questions?
Maybe the primary query to ask is: Are the pigeons simply randomly pecking on the display screen, or are their pecks intentional? If the pigeons are pecking randomly, then the information factors created by every peck must be randomly distributed throughout the show. If the pigeons are pecking deliberately, nevertheless, then the information factors (peck places on the display screen) ought to fall into clusters. These clusters must be organized spatially and temporally and present a development because the pigeon creates its paintings.
Our first step was to cluster the time course of pecking utilizing an iterative partitioning strategy utilizing Ok-means clustering, which is predicated on the evaluation of behavioral bouts produced by animals in normal foraging duties. We then in contrast the temporal clusters to their spatial places on the display screen to find out the diploma of randomness.
We utilized this evaluation to the paintings under.
Outcomes? Temporal clusters have been positively correlated with spatial clusters, giving an Adjusted Rand Index or ARI = 0.26. For comparability, and ARI = -1 is strongly random, whereas an ARI = 1 is strongly non-random. Thus, whereas an ARI of 0.26 is comparatively weak, it’s undoubtedly non-random!
Under is an video animation exhibiting the evolution of clusters of pecks because the paintings was created.
A pigeon’s acquired STYLE!
To our untrained eye, it appears like totally different pigeons create paintings with their very own particular person model. Listed below are some examples.
However can we decide creative model?
To reply this query, we developed an SVM (Assist Vector Machine) mannequin utilizing machine studying. An SVM is only a fancy-sounding instrument that makes use of an algorithm to study a coaching set of knowledge in order that it may well classify them. For our functions, we educated the mannequin on a set of various artworks from 5 totally different pigeons. The mannequin was educated on the peck information for a set of artworks from every pigeon artist. Coaching was primarily based on 8 behavioral attributes of the pecking conduct, reminiscent of spatial variance, inter-response time, line size, and many others. Thus, the mannequin realized to categorise the artworks to the 5 classes, every class being a selected pigeon artist.
We then examined the power of this mannequin to appropriately determine new artworks from these similar birds as belonging to the right artist. After coaching, our mannequin appropriately categorized novel paintings at 60% accuracy on the primary guess! And accuracy improved for the highest 2 (85% accuracy) or high 3 (92% accuracy) guesses.
Sure, these birds have gotten model!
Once we look extra intently on the information desk exhibiting how strongly every issue decided prediction accuracy, we see that spatial variance (that’s the variation within the x, y display screen places of every peck in relation to one another peck for a given paintings for a given artist) was the strongest inter-bird artist distinction. We’re not precisely positive what this implies, however we’re giving it a whole lot of thought!
That is only a glimpse of some preliminary analyses. We’re excited to delve deeper into the information and pull out some extra nuggets of knowledge.
Thanks for studying and please let me know what you assume?
Additionally, the pigeon artwork is still available for minting, so please help our pigeon artists! All proceeds from gross sales go on to the pigeon artwork challenge!
Yours in gratitude,
Aaron
