Diffusion
There was an idea I wanted to explore in my PhD but never got round to. Explaining it to someone this week made me realize something.
Week 12 of life post-MAS. (Links to past weeks in my newsletter.)
A productive week. More doing than talking. Decks done. Proposals sent. Collaborations started. Cutting face-to-face meetings did wonders.
The familiar. Someone I have been working with at a global finance institute and his partner. Collaborators on an agentic AI risk management research paper. Planning a roundtable next week for executives from wealth management and family offices with folks I know.
New connections. An AI lead at a global communications firm, new to Singapore. A lawyer who wanted to explore collaborating on AI governance for the legal sector. A cybersecurity lead at a global network security firm. A senior figure at a financial association for fund managers. A course director at an international training institute thinking about AI supervision resources for regulators. Folks from an association working on AI and work. A workforce agency that wanted to feature my story. And collaborators on a forum on AI and work, later in the year.
I found a weird intersection between my life and my PhD research.
Graph Diffusion
This image in the header is from my PhD research, specifically from DynScan, a model I built to learn concepts from dynamic multimodal networks from the web. Each sub-graph represents a concept the model learned to recognise, some dense and highly connected, some sparse and peripheral. Strong signal or weak. Worth paying attention to or not.
In a conversation this week, I found myself explaining graph diffusion. Not an entirely foreign concept. My PhD research, such as DynScan, was built on graph neural networks, and graph diffusion is one method that GNNs use to propagate information across a network. What I never got to was a specific and interesting extension: using graph diffusion with learned functions for simulations.
The topic was on multi-agent systems for answering questions about the world. That led me to share how they might want to explore simulations with graph diffusion. And instead of learned functions, let agents drive simulations across complex networks.
The idea of graph diffusion is deceptively simple. It is how a signal travels across the network. It starts at a source node. Spreads along edges to neighbours. Then to their neighbours. Propagating outward, hop by hop, across the entire network. Repeat for every single node in the network. Kind of like ideas spreading across a social network. You usually do this until the whole network system converges on a fixed point.
What we focus on in such graphs are usually three things.
The embedding or meaning. If the embedding is weak or generic, what propagates is noise, not signal.
The edges or relationships. Not all connections are equal. Some edges carry signals well, others don’t.
Over-smoothing. Diffuse too much, too indiscriminately, and every node starts to look the same.
Explaining this made something click.
Three Months. One Hundred and Fifty Conversations.
3 months and close to a hundred and fifty conversations. I think I went way overboard trying to find that signal.
The rooms I walked into were diverse.
The people spanned veterans who had seen every hype cycle, researchers doing quiet foundational work, educators trying to close the gap between knowing and doing, practitioners living the messy middle, consultants orbiting the same problems, tech builders constructing the next thing, community builders that rarely talk to each other, and creatives who reminded me of my neglected watercolors.
The sectors were just as varied. Banking and finance, regulation and policy, academia and research, consulting, startups, the arts, sustainability, law, media, public service.
And the roles ranged from PhD students to chief risk officers, from corporate veterans to first-time founders, from bankers to anthropologists studying money itself.
I realised that I was doing my own graph diffusion.
Finding meaning. Building relationships. Making sure that I don’t wander too much and get lost.
Connecting to What Makes Sense
Graph diffusion accentuates the edges that matter. And I realized that not all connections made sense.
Some edges were strong. Understanding instead of hype. Practice instead of theory. Depth instead of awareness. Those went somewhere. Others were weak. Shallow and unclear. I walked away from several of those. And over-smoothing was the subtler risk. Saying yes too much dilutes you. A hundred and fifty conversations will do that to you.
The solution, in graph diffusion, is to learn the right meaning and find the right relationships to strengthen.
Still figuring out both. But a little clearer than last week. And I hope it gets clearer, week by week.
#GraphDiffusion #AIRiskManagement #Signal #Transitions #Reflections