Day 2: Evolution of Neural Circuits, Eulogy for Yves Frégnac - Gilles Laurent and Florian Engert

Evolution's Playbook: Convergence, Inheritance, and the Making of Brains: Gilles Laurent and Florian Engert

Today's first session shifted our focus from the how of building neural circuits to the why they are the way they are. Led by Gilles Laurent, we explored the long timescales that have shaped nervous systems across the animal kingdom.

Setting the Stage: Deep Time and Shared Origins

Gilles began by laying out the timescales at which evolution operates. Life itself arose remarkably early on Earth (~4 billion years ago), with fundamental molecular building blocks like ion channels and neurotransmitter systems appearing long before complex nervous systems, often originating in single-celled organisms like bacteria. Neurons took a lot longer to appear on the stage, something like 600 million years ago. This coincides with the Precambrian/Cambrian boundary, where animal life diversified explosively, setting lineages like ours (vertebrates) and cephalopods or insects on independent evolutionary paths. However, this separating force of speciation has its counterpart in the concept of convergence:   

Convergence: Solving the same problem by different means

Convergence is the first of the 2 key ideas of the talk: the independent evolution of similar solutions to similar problems in different lineages. An illustrative example is given in STDP:

• STDP (Spike Timing Dependent Plasticity). This synaptic learning rule, dependent on the precise timing of pre- and post-synaptic spikes, was first described in vertebrates. Yet, Gilles's lab found a functionally identical STDP rule in insects. The molecular implementation differs here (glutamatergic in vertebrates, cholinergic in insects), pointing towards convergent evolution of the computational rule itself. However, keep in mind that the role might differ – perhaps learning in one context, but network homeostasis or synchrony preservation in another.   

The rundown: convergence hints that some solutions might represent particularly effective, or perhaps even necessary, ways to solve fundamental biological problems.

Inheritance: Working with the "Cards Dealt"

The second major force is inheritance, or historical contingency. Evolution doesn't design from a blank slate; it tinkers with what's already there, modifying existing structures and molecules. It's working with "a bag of tricks" accumulated over eons.

• Example 1: The Insect Eye. The compound eye of insects, made of many tiny lenses derived from cuticle, is optically limited compared to a single-lens eye like ours. This design arose early in arthropod evolution and constrains subsequent brain evolution, forcing solutions adapted to this specific input structure.
• Example 2: Molecular Toolkit. As mentioned, the basic ion channels and receptors used in our synapses were largely inherited from ancient bacterial ancestors, originally serving different purposes before being co-opted for neural communication.
• Loss of Function: Inheritance isn't just about keeping things; structures or abilities can be lost if they cease to be advantageous (or become disadvantageous) in a new context, like vision in cave-dwelling fish.

The rundown: Inheritance explains why biological solutions often seem odd or non-optimal from a pure engineering standpoint – they are products of a specific, constrained historical path.

Beyond Simple Optimization: Divergence and Biological Goals

The discussion also touched on divergence – solving the same problem in fundamentally different ways. Sound localization, for instance, is achieved via different neural and even physical mechanisms in barn owls, small mammals, and tiny insects facing vastly different physical constraints (like minuscule ear separation).

This complexity led to a key point debated in the session: evolution isn't a straightforward optimizer seeking a single perfect solution. It's driven by the messy realities of biology – the "Four Fs" (Feeding, Fleeing, Fighting and uhhhh.....Mating), sexual selection (like the peacock's seemingly impractical tail), and evolutionary arms races. Gilles argued that understanding these biological goals and constraints is essential, suggesting purely physics or information-theoretic views can miss the point if they ignore the adaptive context that gives rise to biological computation. Evolution finds what's "good enough" to survive and reproduce in a given niche, using the historical toolkit available.

Wrapping Up the First Session

The session provided a powerful reminder that the neural circuits we study and emulate are products of an incredibly long and complex history. Understanding the interplay between the functional pressures driving convergence and the historical constraints of inheritance offers a guiding perspective on why brains are structured the way they are, hinting at both potentially universal principles and the unique signature of evolutionary pathways. This is key for interpreting biological systems and perhaps for guiding our own neuromorphic designs.

Turtles, Drives, and the 'Easy' Art of Moving: More Evolutionary Insights

The morning's exploration of evolution continued, getting deeper into the origins of behaviour, the nature of emotions, and a provocative take on what constitutes the real hard problem for brains. But first, a moment of reflection.

Remembering Yves Frégnac

Before the main discussion, Gilles offered a remembrance of Yves Frégnac, a distinguished neuroscientist and cherished regular at Capocaccia, who sadly passed away (see Yves and Irina photos here). Yves is remembered for his pioneering role in computational neuroscience stemming from his unique background combining engineering and neuroscience and his informed criticism HBP. Gilles touched upon his significant contributions as an experimentalist, his early adoption of computational modeling, his interest in neuromorphic systems, and his complex engagement with and reflections on the Human Brain Project. Yves was a passionate debater, a lover of French cinema, and a talented photographer, making use of Fuji cameras to capture the foggy hillsides of his native Burgundy. 

Following the Eulogy, Florian Engert picked up the evolutionary thread from the morning's session. He presented a baffling example of behaviour: the transatlantic migration of leatherback sea turtles.

"The Crossing of the Turtles"

• The Puzzle: Why do these turtles undertake an arduous East-West migration between Brazil and West Africa? It's not a typical seasonal shift, and feeding/predator conditions aren't drastically different.
• The Hypothesis (Continental Drift): Florian proposed an explanation rooted in deep time. 250 million years ago, Africa and South America were neighbours. Perhaps the turtles started a short crossing for some minor benefit. Then, as the continents drifted apart over millennia, the turtles, like "good conservatives," just... kept doing it.
• The Takeaway: Their current epic journey isn't necessarily optimal now; it's a legacy behaviour, an "evolutionary baggage" inherited from a vastly different past. As Florian imagined the turtles saying, "We've always done it this way, and we're doing fine, thank you very much!" This highlights how behaviours can persist if they're "good enough" not to cause extinction, even if costly or seemingly illogical today. We see the survivors of this historical path.

Emotions: Ancient Drivers, Ancient Goals

This turtle story led Florian Engert (Harvard) to his broader view on emotions and drives. He argued they are ancient, fundamental goal-setters shared across species (seeing fish emotions as homologous to ours, not convergent).

• Emotions First: In his view, evolution first establishes drives or urges ("I need to find food," "I need to find warmth," "I need to cross this water"). Cognition and complex nervous systems then evolve as mechanisms to achieve these goals, minimizing the "error" between the current state and the desired state. This contrasts sharply with engineering, where goals are typically imposed externally.
• Fish Happiness & Oxytocin: Florian illustrated this with his lab's work on larval zebrafish. They found oxytocin, often linked to love and bonding in mammals, seems to act more generally in fish as a hormone released during stress (loneliness, hunger, heat) that drives behaviour to solve the problem. Lonely, stressed fish explore more (likely seeking siblings) and eat less. The smell of other fish (the "stink of their brothers") can reduce their oxytocin, calm them, and trigger feeding – suggesting "happiness" can be measured and chemically modulated even in tiny fish. (He also half-jokingly proposed oxytocin's role in mammals might include helping females tolerate "obnoxious males")
• Seven Deadly Peptides?: Tongue-in-cheek, Florian floated the idea of mapping the Seven Deadly Sins (lust, gluttony, wrath etc.) onto distinct hypothalamic peptides, suggesting these ancient drives, recognized by the Church as powerful and inevitable human urges, might have specific neurochemical bases.

Innate Skills vs. Learning

Florian then offered a provocative perspective on learning, arguing that competence is largely innate.

• Evolution Equips: Evolution equips animals with the vast majority of knowledge and skills needed for survival and reproduction, encoded genetically.
• Learning as Calibration: What we often call learning, especially synaptic plasticity, is primarily about tuning and calibration – keeping the innate machinery functioning correctly in the face of bodily changes and environmental fluctuations.
• True Learning is Rare: Learning genuinely new knowledge or skills is rare, happens fast when needed, and isn't required for most basic animal success. Helpless newborns? Just born "prematurely" and need development to unlock innate potentials, not learn them from scratch (human language being a potential major exception).

What's Really Hard? Spoilers: It's Not Chess

Finally, Florian challenged conventional thinking about computational difficulty, invoking Moravec's Paradox.

• The Setup: Recounting a talk where a BMI researcher dismissed motor execution as simple compared to cognitive planning, Florian argued the opposite.
• The Paradox: Abstract reasoning (like chess strategy or solving a Rubik's cube algorithmically) is hard for humans but relatively achievable for computers. Conversely, dexterous, graceful motor control (physically moving chess pieces, manipulating a Rubik's cube, a toddler loading a dishwasher) is effortless for humans but extremely difficult for current robots.
• Why?: We perceive motor control as easy because hundreds of millions of years of evolution have perfected these innate abilities. Robots lack this evolutionary "pre-training" and the sophisticated, constantly-calibrating biological substrate (muscles, sensors, feedback loops). The true computational cost and complexity lie in interacting fluidly with the physical world.

That wraps up the morning talks! 

Afternoon brought a popular beach tutorial on modern neural architectures including diffusion networks from Alex Mordvintsev (sketched with toes in wet sand, quickly washed away by the waves) and lots of workgroup activity.