Dark pictures, thrones, poems that take a thousand years to die: algorithms, butterflies, and enigmas of extinction

According to the most recent update to the Catalogue of Life, there are 1869 genera in the Geometridae. The larvae of these moths are known to most of us as the inchworms; “geometer” can be translated “earth measurer.” from Abaciscus alishanensis to Zythos turbata, more than 21,000 species have been named and described, transmitted from one natural historian to another, from Linnaeus onward.

What’s in a name? Each of these binomials originates in an act of capture, destruction, and description, the transmutation of living insects into data. For each of these names, somewhere there is a type specimen fixed by a pin, labeled, and secreted away in a drawer. Even before Darwin’s time, these activities were the rites and sacrements of natural history. That most famous of lepidopterists, Vladimir Nabokov, evokes in his poem “Discovery” the glamor of this work, for him an act of making more durable than literature:

Dark pictures, thrones, the stones that pilgrims kiss,
poems that take a thousand years to die
but ape the immortality of this
red label on a little butterfly.

In the buzzing, blooming world, however, many kinds of insects evade the killing jars of entomologists and poets alike. Shake a tree in the forests of the Amazon, and three-quarters of the insects that fall to the ground could be “new to science”—undocumented by the canonical forms of publication and museological preservation. Changes to earth’s planetary system, effected by humankind’s technological acceleration, are causing many of these species to disappear before they can be so named and described. This loss is not merely an aspect of some possible future, but an inescapable fact of our present moment.

“Nature,” William James said, “is but another name we give to excess.”

Recently, I’ve been trying to explore the sublime vastness of biodiversity—and the enormity of extinction—with the help of algorithms. Specifically, I’ve been running a recurrent neural network script trained on the full list of Geometrid genera from the Catalogue of Life. The output below, captured early in the run of the script, shows results that are still fragmentary; as the script runs, the output resolves into names like those of extant moth genera in the training data, but novel: names for unknown moths.

This activity seems biological, doesn’t it? At least it does so by dint of the language games we play with life qua biology. Neural networks generate output in a process meant to mimic the development of neural connections in the brain, but with superficial similarities to the ways we describe evolution by natural selection. Names arising from this process have the flavor of scientific genera — the Latinate syllables, the husks of European surnames and honorifics. We describe such algorithms as this neural network as proceeding without rules or rubrics, without plans or a priori conceptions—much as evolution has been called the “blind watchmaker.” How closely these descriptions match states of affairs in the world is arguable—while they seem empirical enough, they also mobilize certain kinds of ideological devotions. To be sure, though, the algorithm doesn’t know a moth from a butterfly; doesn’t know pollination or metamorphosis; has never seen the luminous commas of inchworms hanging by silk threads, or the apparition of Operopthera brumata, the winter moth, rising in fluttering clouds from November’s damp snow.

“Nature,” William James said, “is but another name we give to excess.” In the past, people thought biological life profligate, even wasteful, in its baroque abundance and astonishing diversity. Only now, with the advance of technology, has life on earth been revealed as fragile and ephemeral. Our computational systems, meanwhile, are entering what David Weinberger has described as a “post-scarcity” phase, in which machine-learning algorithms processing endless layers of data seem ever more like natural phenomena.

This is one of the stories we’re telling about the emergence of deep-learning networks. What other stories might we ask these systems to help us discover and tell? The algorithm, which has never seen a moth, gives strings of character that look to me like the names of insects; I’m using them to imagine all the moths we’ve never discovered, and all those we might never have the chance to see or to know.

And I’m asking the algorithm to challenge you, too — to prompt you to look for undiscovered insects, before the moth-making world is broken for good. I’ve set up a Twitter bot offering provocations to discovery and encounter with lepidopteran dark abundance.

Of course, most of the undiscovered moths inhabit biodiversity hotspots far from our common haunts. But maybe these challenges can tickle our attention. Between the loss of nature and the rise of inexhaustible machines, what work is there for us to do?

These notes accompany Earth Measurer, a mini-exhibition at Harvard’s Berkman Klein Center, up through July. Output generated using min-char-rnn.py, a minimal character-level Vanilla Recurrent Neural Network model written by Andrej Karpathy (@karpathy).