Cordyceps Septet

Electrical activity in a fungus rendered as light and sound, with the reading of a small model trained on the voltage data.

The recording

Andrew Adamatzky inserted seven pairs of electrodes into a culture of Cordyceps militaris. Each pair recorded the voltage between its two points once a second for twenty-two days. Everything in the work comes from these readings. At the default pace, one day of the recording passes in one minute.

The light

In the visualization, each pair of electrodes is a line of light with the seven lines forming a ring. The bright ring is the present. The past recedes behind it, six hours at a time, with a faint ring for each hour and a brighter ring at the turn of each day.

Each channel drifts slowly over hours. A spike is a sudden jump away from that drift, more than four times the channel's usual tremor, so each channel is measured in relation to its own tendencies. A spike flares as a bead and dims as it recedes.

When two channels spike within thirty seconds of each other, a gilt arc joins them.

The sound

No sound was recorded. Each pair of electrodes has its own note. The low tone follows the drift. Each spike strikes a channel's note, and each meeting sounds both notes together. Sound and light come from the same numbers at the same moment.

The model

The blue light belongs to a small transformer model, the design behind large language models. A language model learns by guessing the next word, checking that guess, and then adjusting. This model reads the last sixteen minutes of all seven channels and guesses whether each channel will spike in the next ten minutes. It was trained on the first four-fifths of the recording, which is about eighteen days.

The blue glow at the end of each line is its guess. The blue threads inside the ring show which channels it reads as it guesses. The gilt arcs outside the ring record what happened; the blue threads inside show what the model learned to attend to.

The model keeps none of the readings, but it remembers the relations. On the four days it never saw, it anticipates spikes better than chance on six of the seven channels, and it anticipates them less well when the other channels are scrambled. The quietest channel, which almost never spikes, is the one it reads most. The model cannot say whether the channels signal one another or answer a shared condition such as moisture or growth.

In numbers

Seven electrode pairs; 1.9 million readings per pair; 5,948 spikes; 527 meetings, of which about 425 would occur by chance. The model holds about 13,000 learned numbers in two layers.

Sources

Data: Andrew Adamatzky, “Language of fungi derived from their electrical spiking activity,” Royal Society Open Science 9 (2022): 211926. The recordings are published on Zenodo under a Creative Commons Attribution 4.0 licence.

Model design: Ashish Vaswani and colleagues, “Attention Is All You Need,” Advances in Neural Information Processing Systems 30 (2017).

The model’s code is open source under the MIT Licence.

Return to the work