Strange Attractors Again
My life as an independent for the past 18 weeks has been quite chaotic.
New rooms. New people. Connections and rejections. Building, speaking, writing. The only thing I have not done in these 18 weeks is art. Which I should try to get back to.
Not random chaos. But chaos nonetheless.
At the end of the week, I read a research paper: “Machine Learning Predictions from Unpredictable Chaos” by Jiang et al. I’ve written about chaos theory and emergence before. But going back to this topic at week 18 of independent life sparked some thoughts. (Links to past weeks in my newsletter.)
The new. Folks from an agentic risk and compliance startup in UK. A veteran in the analytics space. The chief scientist of a technology company. A startup founder doing AI and corporate events. A yoga teacher & AI writer.
The familiar. Someone leading a AI bootcamp that I am also helping with. My cousin who did a long needed headshot for me. A chief AI officer who I have crossed paths with a number of times in the past year. A veteran in the financial crime space.
The Paper’s Insights
Chaos theory is not really about chaos. It’s about the hidden order that one can find in what seems random at first. Strange attractors, fractals, the butterfly effect. They all describe a random system that, at certain states, shifts into a fascinating pattern.
In this paper, the authors propose chaotic learning - a mashup of chaos theory and topology - to learn features that can be used for AI.
The idea is quite simple. Take any dataset - protein structures, brain waves, financial networks. Represent it as a network where nodes are data points and edges represent their relationships.
Attach a strange attractor or oscillator, such as the Lorenz (named after the person who described the butterfly effect) to every node. Then couple those oscillators together using the topology of the network. Let the system evolve. As connections strengthen, they start to sync. And we start getting features that help machine learning methods make better predictions.
The paper learnt that partial synchronization is better than both no synchronization and full synchronization.
Chaos is not noise. When shaped by the right connections to the right degree, order can emerge from chaos.
My PhD research was on something very similar. Learning features from dynamic multimodal networks - graph and time series modelling, multimodal learning. All of these involved encoding relational structure into learnable representations. How do you take a network, a set of nodes and edges with complex relationships, and turn it into something AI can use?
This paper on learning from chaos solves the same problem.
But instead of using learned weights to encode the network structure, it uses chaos + network dynamics. The topology shapes how the chaotic oscillators couple. The coupling patterns encode the structure. The encoding enables prediction.
But what’s interesting is my realization that this connects to life.
18 Weeks of Chaos
The past 18 weeks through the lens of the paper.
I’ve deliberately made it random. No plan. Nothing forced, from week 1 to now. There were nodes that synchronised. Then there were the ones that went nowhere.
Some days everything lines up - coffees lead to connections lead to collaborations. Some days nothing works - rooms that felt wrong, things that disappear into the ether, my discomfort at networking sessions.
The central finding of the paper on learning from chaos is that seemingly random and unpredictable chaotic dynamics counterintuitively offer accurate predictions, when shaped by the right topologies, i.e., the right connections.
In a nutshell. Topology is everything. Even if things seem chaotic. The relationships between the nodes. The strength of the coupling. The structure of the connections.
That seems to describe life when one is independent.
Chaos is never random. It’s shaped by the topology. Now I just need to figure out what synchronisation for me looks like. And maybe pick up the paintbrush again