Machine Learning Street Talk (MLST) — Episode #75

Emergence [Special Edition] with Dr. Daniele Grattarola
Date: April 29, 2022

Episode Overview

In this special edition, the MLST hosts are joined by Dr. Daniele Grattarola to deeply explore the concept of "emergence" as it appears in complex systems, AI, and in particular, through the lens of cellular automata and their generalizations to graphs. The episode weaves together foundational philosophical distinctions (weak vs. strong emergence), insights from leading scientists and philosophers, and cutting-edge research at the intersection of neural networks, biology, and computation. The conversation is rigorous yet exploratory, traversing cellular automata (CA), neural CA, morphogenesis, graph-based systems, and prospects for robust, self-organizing, self-healing AI systems.

Key Discussion Points & Insights

1. Defining Emergence

[00:58–08:00]

• Levels of Abstraction:
  Host 2 discusses how phenomena at higher abstraction levels (like trust, culture) become hard to quantify, and parallels this to emergent phenomena in machine learning and society.

• Specialization vs. Generalization:
  Dr. Grattarola notes population-driven algorithms highlight specialization, diverging from ML’s default emphasis on generalization. Populations tend to produce many specialized "exotic" behaviors not seen in generalist models.

  "The population implies... I want to see a lot of different things and like hyper specializations to all kinds of exotic things that... the generalist won't do." — Dr. Daniele Grattarola [02:20]

• Philosophical Foundations:
  Citing Melanie Mitchell, John Locke, and biologists, the hosts explain how simple agents, when aggregated, can show "superorganism" behavior — collective intelligence not reducible to individuals (e.g., army ants).

• Reductionism vs. Relationism:
  Tim summarizes connections to physics and philosophy, exploring how Western science’s reductionism contrasts with relationism—emphasizing system interactions and context.

  "[Reductionism] keeps chopping up things into smaller and smaller pieces... stark contrast with relationism." — Host 1 (Tim) [04:07]

2. Weak vs. Strong Emergence

[08:00–21:00]

• Weak Emergence:

  • Defined as new properties arising from interactions of simple entities, visible only at scale, and often only discoverable via simulation (computational irreducibility).
  • Marc Bedau: Weak emergence = macrostate P with microdynamics D is weakly emergent "if and only if P can be derived from D and S’s external conditions, but only by simulation."
  • Strong Emergence: Properties not deducible from lower-level facts, often connected to the "mystery" of consciousness.

• Chalmers and Hossenfelder:
  Professor David Chalmers and Dr. Sabine Hossenfelder debate the meaning and legitimacy of "strong emergence," especially in the context of consciousness and free will.

• Example from Chalmers’s work:

  • "He argues that consciousness isn't a logical necessity... he could imagine a universe... with the same physical laws where he would be a philosophical zombie." — Host 1 (Tim) [18:14]

3. Infinity and Computability in the Universe

[30:42–34:00]

• Infinity vs. Unboundedness:
  Keith and a guest debate whether actual infinity or just unboundedness exists in our universe, referencing Gödel’s results and the limits of language and computation.
• Implication for Emergence:
  Computationally irreducible systems, especially those involving real infinity, might always escape complete formalization — supporting "strong emergence" in principle.

4. Cellular Automata: Discrete, Continuous, and Universal Computation

[36:00–53:40]

• Basics and Historical Context:

  • Cellular Automata (CA)—simple rule-based models that, via local computation, can yield enormously complex behavior.
  • Conway’s Game of Life as a canonical example.

• Morphogenesis and Self-Organization:

  • Biological systems specify only simple developmental rules, which self-organize into complex, robust forms (the "genomic bottleneck").

5. Emergence in Neural Cellular Automata & Graphs

[47:01–57:23]

• Alexander Mordvintsev’s Work:
  Neural CAs can self-assemble (morphogenesis) and be trained to recover desired global patterns (like images) robustness and adaptability to perturbation.

  "They've turned a self healing image generation process into an emergent phenomena..." — Host 1 (Tim) [47:01]

• Decentralization vs. Centralization:
  Spontaneous centralization (leader election) often emerges in decentralized systems (brains, societies), but the process of that emergence gives adaptability, robustness.

6. Dr. Grattarola’s Research: Learning Graph Cellular Automata

[50:20–105:03]

• Graph Cellular Automata (GCAs):
  Dr. Grattarola generalizes CA by defining the local rule not over a regular grid but over an arbitrary graph, using graph neural networks (GNNs) to learn the rule.

• Morphogenesis on Arbitrary Graphs:

  • Demonstrated with point clouds (e.g., forming the shape of a bunny) via only local updates governed by a GNN.
  • Emergence happens as a distributed, decentralized process on arbitrary geometries.
  • Notable quote:

    "Does a rule exist that, starting from a random configuration of points, actually morphs these points into this coherent shape?... Yes, this can be expressed as a process that, iteratively and locally, kind of grows the image into what we want." — Dr. Daniele Grattarola [98:11–98:54]

7. Universality, Computation, and Iterative Dynamics

[91:29–96:23]

• Neural CAs and Computation:
  The neural CAs described (both grid and graph-based) use complicated neural rules (learned, not hand-coded) and often operate via many iterations.
  • "Without that iterative capability, without that kind of working space, without that temporal dynamics, you don’t get this kind of behavior." — Keith [95:13]

8. Practical Implications & Future Directions

[109:17–114:38]

• Robust, Self-Healing Systems:
  Potential engineering applications: systems that can self-heal (drawing from morphogenesis), robustify to changes, or coordinate in complex environments.

• Where to Learn More:
  Dr. Grattarola recommends resources:

  • For cutting-edge CA/neural CA: Twitter and GitHub communities.
  • For academic work on biological emergence: Michael Levin’s lab work (e.g. "xenobots").
  • For visually engaging explanations: YouTube channel "Emergent Garden".

• Theory Frontiers:
  Quantifying emergence via entropy, exploring the edge of chaos (Langton), and the notion that simple rules at multiple system levels yield recurring emergence.

Notable Quotes & Memorable Moments

• On the Surprising Power of Simplicity:

  "There is absolutely no reason why this should work, like at all. There is nothing that we can observe that says that these kinds of rules should exist at all. This model, in principle, it's like it's too simple for it to actually work. But in fact it turns out that these models are... universal models." — Dr. Daniele Grattarola [91:29]

• On Weak vs. Strong Emergence:

  "Any endorsement of strong emergence is a rejection of physicalism and reductionism, which is to say an appeal to magic and esoterica. Whereas weak emergence can be used to support the physicalist picture of the world..." — Host 1 (Tim) [20:41]

• On the Recurring Nature of Discrete and Continuous:

  "As you move along scales of emergence or reduction, you keep coming across the need to either view things as a continuum or as a discrete spectrum... this alternating series... never converges." — Guest 2 (Keith) [76:26]

• Neural Networks & Iteration:

  "A neural network as it's typically conceived... by itself is not Turing complete. What you need is the ability to do this iterative computation... and that's exactly what we have in this work." — Keith [95:13]

• On the Practical Side of Science:

  "What I came to realize is that at some point it becomes a matter of solving the problem. Solving the problem is more important than the way you solve it, in a sense." — Dr. Grattarola [105:42]

Timestamps for Key Sections

• [00:58] – Defining emergence, levels of abstraction, population vs. individual
• [03:08–05:00] – Melanie Mitchell on complexity, ant colonies, and reductionism
• [08:00–13:00] – Weak/strong emergence, Marc Bedau, David Chalmers, Hossenfelder
• [18:14] – Chalmers, consciousness, philosophical zombies
• [30:42] – Infinity, computability, fundamental limits
• [36:00–43:00] – Cellular automata, morphogenesis, nature’s robust design
• [47:01] – Neural CAs, self-organization, real-world examples
• [50:20–53:43] – Dr. Grattarola’s background and graph neural network journey
• [57:23–63:35] – CA as modeling tools: across scale, continuum vs. discrete
• [91:29] – Neural CA universality and surprise
• [98:11] – Point cloud "bunny" morphogenesis on a graph
• [109:56] – Further resources and recommendations
• [114:38] – Edge of chaos, entropy, the future of studying emergence

Additional Resources Mentioned

• Books:
  "Complexity: A Guided Tour" — Melanie Mitchell
  "A New Kind of Science" — Stephen Wolfram
  "Godel, Escher, Bach" — Douglas Hofstadter

• Papers/Authors:
  Sabine Hossenfelder — "The case for strong emergence"
  Sebastian Risi — "The Future of Artificial Intelligence is Self-organizing and Self-assembling"
  Michael Levin (biological emergence and morphogenesis)
  Alexander Mordvintsev — "Growing Neural Cellular Automata"
  Langton (Edge of Chaos)

• Online:
  Twitter & GitHub CA communities
  YouTube: Max Robinson’s "Emergent Garden"

Takeaways

• Emergence lies at the heart of both the natural world and artificial intelligence.
• Whether weak or strong, emergent phenomena highlight how systems’ complexity cannot always be deduced from simple parts—sometimes only accessible via simulation and iteration.
• Neural CA and graph CA research point to new ways of building robust, adaptable, self-organizing AI, but also raise profound theoretical and engineering questions.
• The field is rapidly evolving, bridging theory (physics, computation, philosophy) with practical applications (robust AI, neural modeling, biological insights).
• Staying at the bleeding edge means following both academic work and informal hacker/enthusiast communities online.