Learning Out Loud with AI-Generated Tutorials

Every researcher has a list of topics they have been meaning to learn properly “one day”. Mine has been growing for years, and the machine learning literature grows faster than I can read it.

So I built myself a helper.

A tutorial writer that works for one reader

It is an AI agent based on Claude. I give it a topic, and it goes off and does what I would do if I had a free month: it collects the relevant papers, reads them, and writes an in-depth tutorial that teaches the area from first principles. Not a survey that lists who did what, but a teaching document: intuitions first, then the mathematics, worked examples, code, cheat sheets, and an honest section on what does not work yet.

I use it in two ways. Sometimes I want to catch up on what is new in an area I already know. Other times I want to get completely into an area I have only watched from the outside.

I originally made these for myself. But some of them turned out good enough that it feels wasteful to keep them on my laptop, so I will be sharing the ones I think are worth reading.

A fair warning first: these tutorials are AI-generated. I read them, I learn from them, and they have been genuinely useful, but I have not checked every derivation and every citation line by line. Treat them as a very good starting map, and go to the original papers for anything you intend to build on. If you spot an error, I would be glad to hear about it.

Tutorial 1: Decision-Focused Learning

📄 Decision-Focused Learning tutorial

This one has been on my list for years.

The idea is simple to state and surprisingly deep. In the classic predict-then-optimize pipeline, we train a model to be accurate and then feed its predictions into an optimisation problem to make a decision. But accuracy is not what we actually care about; the quality of the decision is. A model can be slightly less accurate everywhere and still lead to much better decisions, if it is accurate where it matters. Decision-focused learning trains the predictive model on the downstream decision loss directly, which means differentiating through, or finding good surrogates for, the optimisation problem itself.

For someone who works on decision support systems and keeps arguing that prediction alone is not a decision, it was long overdue.

Tutorial 2: Hyperbolic Representation Learning

📄 Hyperbolic Representation Learning tutorial

This one started differently. Over the last couple of years I kept running into hyperbolic papers at NeurIPS, ICML, and ICLR, and at some point curiosity won.

The key insight is geometric. In Euclidean space, the volume of a ball grows polynomially with its radius; in hyperbolic space, it grows exponentially — just like the number of nodes in a tree grows with its depth. That makes hyperbolic space a natural home for hierarchical data: trees and taxonomies can be embedded in very few dimensions with low distortion, with general concepts sitting near the centre and specific ones pushed out towards the boundary.

Why these two, together

If you know my work, you can probably guess why these two caught my attention at the same time.

Much of my research revolves around DEX (Decision EXpert), a qualitative multi-criteria method in which a decision model is a hierarchy of attributes, and each aggregate attribute is computed from its children through an interpretable decision table of rules. DEX models are transparent, consistent, and easy to discuss with domain experts. They are also traditionally built by hand, which is slow.

Over the years, my colleagues and I have kept chipping away at that bottleneck:

• Integrating knowledge from DEX hierarchies into machine learning models for ski injury prediction;
• DIDEX, which extracts DEX models directly from data;
• Combining LLMs with DIDEX to predict internal migrations, and LLM + DEX for faster construction of hierarchical decision models;
• Linguistic-summary-based feature selection for qualitative decision models;
• and most recently DEX-LL, where LLMs propose candidate attribute groupings while rule induction stays data-driven, monotone, and auditable.

Two questions keep coming back in all of that work. Where does the hierarchy come from? And how should the decision rules be learned?

Those are, almost exactly, the two questions these tutorials speak to:

• Hyperbolic representation learning might let us discover hierarchical structure in data, rather than asking an expert or an LLM to propose it. Attributes that cluster together in hyperbolic space, and their position between centre and boundary, could suggest which aggregate concepts exist and how they nest.
• Decision-focused learning might let us learn or adapt the decision rules against the quality of the decisions they lead to, rather than against classification accuracy alone.

Put together, the picture I have in mind is a pipeline that finds the hierarchy with HRL and learns the rules with DFL, while keeping what makes DEX valuable in the first place: interpretability, monotonicity, and consistency.

I will be honest: I do not know yet how to make this work. DEX rules are discrete, ordinal, and monotone, which is not exactly the smooth, differentiable world either of these fields grew up in, and I suspect that tension is where the real research problem lives. I am still thinking about it. But the tutorials are helping me think about it much more clearly than I could a few months ago.

More to come

I plan to share more of these as they come out of the pipeline — on topics I have been wanting to learn, and on areas that I believe more people in decision support and operations research should know about.

If there is a topic you would like to see covered, or if you are thinking about similar ideas at the intersection of representation learning, decision-focused learning, and interpretable decision models, get in touch. I would love to talk.