Day 11 - Flying and Floating Minds (Iris Adam, Richard Hanhloser, Gilles Laurent)

 Day 11: Flying and Floating Minds (Iris Adam, Richard Hahnloser, Gilles Laurent)

Welcome back to Day 11 of the workshop. Today we are entering back into the minds of animals, from treetops to ocean floors. We look at how brains control messy muscles, how bird brains remember, and how squids can perform real-time, high-dimensional image rendering on its own skin.

The Biological Control Problem

If you want to understand complex motor control, look at a songbird vocalisation. Iris Adam kicked off the day by framing bird vocalization not just as a beautiful behavior, but as a somewhat mysterious engineering feat.

In machines, actuators (motors) are generally predictable (at least we hope so). In biology, actuators are muscles, which fatigue, overextend (just ask the workshop members playing football in the sport break), and are "plastic“ (meaning they change over time). Furthermore, the muscles a bird uses to sing are multi-taskers, simultaneously responsible for all kinds of functions including breathing, flying, and feeding.

How does a bird achieve reliable, nigh-operatic vocal control when its own hardware is constantly fatiguing or changing? Unlike human limbs, birds have no proprioceptive feedback (the physical sense of where body parts are) for their vocal muscles. They also don’t have a spinal cord, instead the signal path goes from the HVC (the top-level timing circuit) down through the RA, straight through the brainstem, and out to the muscles. So, their only feedback loop is acoustic. They just have to listen to their own song.

Is this an open-loop or closed-loop system? According to Iris, birds don't generally compensate "online" within the exact second they are singing. Muscle fatigue sets in fast (within 1 to 2 seconds) which might set a biological hard limit on the length of a song motif. Instead of fixing things mid-note, they listen, practice, and correct offline.

But even this is still not complete. Specifically, if a bird's song sounds off, how does its brain know which of the many overlapping, multi-functional muscles caused the error (very similar to the credit assignment problem in machine learning)?

An audience member asked whether Reinforcement Learning (RL) in the basal ganglia could support this credit assignment, but Iris pointed out a flaw in the timeline: adaptive responses to muscle fatigue happen on the scale of a single day, which is far too fast for how we think about biological RL. Experienced roboticist Guide de Croon suggested that a tunable forward model (combined with feedback) might be the key to predicting fatigue and adapting to it "on the fly." Unfortunately, proving this in birds would difficult because almost all vocal muscles participate in every vocal motif, making their individual functions hard to disentangle.

Pitch Shift

Next up, Richard Hahnloser opened with a classic machine learning problem: Catastrophic Forgetting. When an artificially intelligent system learns something new, it often completely overwrites and destroys the old information. How does biology deal with this?

Richard walked us through how songbirds deal with new auditory information, revealing that birds struggle less with learning new syllables and far more with mastering the transitions or sequencing between them.

The connection to continual learning and catastrophic forgetting came with his explanation of a "Pitch Shift" experiment. Researchers took a song the birds knew (e.g., ABABABAB) and shifted the pitch of one syllable (B to B’ or B’’) during their critical period (the time of the life they can still learn new songs). The bird's ability to learn depended on the severity of the shift.

Shift Size	Human Analogy	Bird's Response	High-level Interpretation
Small	A German speaker getting by with a little bit of bad Dutch.	Learns an "in-between" mix of the old and new song.	Generalization over similar tasks.
Medium	A German speaker forced to learn English from scratch.	Learns the new pitch completely, discarding the old one.	Forgetting (Overwriting).
Large	A German speaker not even attempting Hungarian.	Fails to learn it, or uses a totally different call.	Task Separation / Ignoring extreme variance.

How does a bird learn these shifts? Well it’s still conjecture but probably through intrinsic variability during its song. Signals in the RA are sometimes referred to as "frozen noise" because they are highly deterministic but still look very random in multiple dimensions. During undirected singing (practicing alone), a brain area called LMAN sends bursts that increase the variability of the song. When the bird sings to a female, this variability drops. The variability may be useful as an exploratory measure, but need to be traded off with actually exploiting a well-honed song.

Richard also floated an interesting evolutionary hypothesis: Why are birds and whales so good at vocal learning? Because they live in 3D. When you can move up, down, left, right, forward, and backward in a dense forest or open ocean, it is easy to get lost. Strong, complex affiliative vocalizations become an evolutionary crutch to recognise the flock or pod and maintain social cohesion.

Cuttlefish Camouflage

Finally, Gilles Laurent discussed the cephalopod, making a point to emphasize three principles: Brains evolve to control adaptive behavior (to do stuff!), natural selection leads to functional convergence, and we can learn general/common principles about biological intelligence by investigating the convergence across creatures with different evolutionary history.

Enter the cuttlefish. It has about 500 million neurons (same as a squirrel) but operates in an entirely different niche. To hide from predators, cuttlefish have to replicate the statistics of real-world textures, which mathematically requires hundreds of parameters to fit. They achieve this via chromatophores, little pigment sacs in cells controlled by radial muscles that come in yellow (young cells), red (middle-aged), and black (old). This chromatophore camouflage is tightly linked to vision. The cuttlefish’s photoreceptors interface with an optic lobe, passing through a tight neural bottleneck of just tens of thousands of neurons, before expanding back out into a high-dimensional output on the skin.

For years, biologists assumed cuttlefish possessed a handful of pre-programmed camouflage "outfits" they could switch between to break up their outline. According to Laurent the opposite is true: camouflage is a continuous spectrum. Every time a cuttlefish hides, it generates a different solution based on the statistics of the visual environment it finds itself in. And even at rest, the skin is not static; at a microscopic level, the chromatophores are constantly "twinkling“.

When moving from one camouflage pattern to another, the cuttlefish navigates a complex state space in terms of the states of its chromatophores. Essentially, the high-dimensional display (that is the „skin“) dynamically "hop" between successive states when moving from one camouflage state to another. During the acceleration phase of a hop, the patterns align directly toward the target camouflage pattern. But during the deceleration phase, the changes become random, resulting in a convoluted, unique path from one camouflage pattern to another.

This directed hopping seems to imply a closed-loop feedback system. But, like we already asked ourselves with Iris, where would this feedback come from? They don't have proprioception, their vision can only see a small fraction of their own body, and efference copies (internal copies of motor commands) would require constant calibration. The mechanism behind this real-time, high-definition skin rendering remains an open, beautiful biological mystery.

Final Thoughts

With the end of day 11, we come to the last day of discussions in the workshop. Learning about the incredible computational techniques across the animal kingdom hopefully inspires the engineers among us to build better control systems and learning algorithms that will be more robust and adaptive. Still, whether it is a bird navigating a noisy, fatiguing vocal muscle system using only auditory feedback, or a cuttlefish projecting its surroundings onto its skin without fully knowing how it's doing it, biology proves that the best control systems in the world can be wet, messy, and mysterious.