Definition
/ˈhaɪ.pəˌpraɪ.ə/ n. 超事前分布
[from Gk hyper over, and L prior earlier]
- 1. stats. A prior distribution on the parameters of another prior.the trouble began with a hyperprior that was far too sure of itself
- 2. A small private research group in London. It studies how learning machines and complex systems come to hold beliefs and act on them, and it builds the environments and graders that test them.
Origin From the Greek hyper, over, and the Latin prior, earlier. A prior on a prior, and so a way of writing down how sure one is about how sure one is.
See also prior, posterior, calibration, hyperparameter
The question
How sure are you about how sure you are?
A hyperprior is that second question, written down as a distribution.
Asked of a model, it is where its beliefs came from and how it would notice the world leaving them behind. Asked of an agent, it is what a grader has to measure. Asked of a human, it is the set of implicit or explicit beliefs that underpin our actions and feelings.
Research
Seven areas, the same question in each.
-
An agent is as honest as whatever scores it. We work on scoring rules that hold when an agent hunts for loopholes, and on simulated worlds where it is trained and scored before anything real is at stake.
-
Agents that learn from the consequences of their own actions. How should an agent explore, and what should it keep from experience that has stopped being true?
-
A model of the world lets an agent try a plan before it commits to one. We work on models that know where they are uncertain, and on planning with a model that is wrong somewhere, which is all of them.
-
Markets, ecologies, protocols, populations of learning machines, where behaviour belongs to the whole and to no part. Selection is the oldest grader, and what it breeds is seldom what was asked for. We study what emerges when learners adapt to one another, which of it was wanted, and how to tell the two apart early.
-
Hierarchical models, checked for calibration against what actually happened. Causal inference sits here too, for when a model’s actions change the data it learns from.
-
Synthetic worlds and agent-based simulation, built to test an idea before it meets real data, and experiments set up so that we can lose the argument.
-
Choosing what to work on is itself a decision under uncertainty, with priors of its own, and knowing when to stop is part of it.
Five working papers are in preparation. Read the abstracts
Engagements
We take on a few outside engagements when the problem sits close to the research. For each one we assemble a small team from the network, senior people who still write the code, quick to arrive, and we are usually the first people in and the last to leave. The brief is often something that has been called impossible, and part of the work is finding out whether that was true. There is no account team. You talk to the people doing the work.
-
An environment that runs at the scale a training run needs, a grader with its reasoning written down, and a record of how agents tried to game it and what we did about that. We build in domains we have worked in for years, quantitative ones among them, and the environment is yours to keep.
-
A test suite that is yours to run again, a report on where the agent breaks and why, and a measure of how far its confidence can be trusted. Miscalibration, brittle failure modes and reward hacking are the usual findings.
-
Models built for data that changes under them, the evidence that they hold on data they have never seen, and monitoring that says when they have gone stale. We have spent a long time on series that fight back.
-
The pipeline from data to a trained model to a served one, built so your own people can run it after we leave, and profiled until we can show you where the time goes. We run it on our own compute or on yours.
-
A written opinion on a team’s systems, protocols, data and claims, from people who have built the same kinds of thing. It says what is real and what is fragile, in time for the decision, and it covers AI teams and quantitative ones alike.
-
A plan with the hard parts named and costed, the first of them started, and the same people to call when it goes wrong. For firms deciding how much of this to do themselves.
Terms
- Start
- A conversation with the person who would do the work
- Form
- Fixed scope or retainer
- Confidentiality
- NDA by default
- Conflicts
- Checked before we say yes
- Compute
- Our own, on premises in London, with rented scale when a question earns it
- Location
- London or remote
- Capacity
- Limited
- Calibration
- Inspected. The seal is beside this line
Who we are
We learned our trade in global markets, where a wrong prior is priced at once and somebody else collects.
HyperPrior Research is a private research group in London, founded in 2022 and led by Patrick Steele. Before founding it he spent more than a decade in systematic and quantitative trading, most recently co-building a systematic desk at DRW (Cumberland), and before that trading APAC Delta One at UBS in Hong Kong. He read computer science at the University of Auckland and holds an MBA from Cambridge.
Quantitative finance is where we come from, and one of several domains we know well enough to build environments and graders for. The habit we kept from it is testing on the future.
We are small by design and privately funded, with no outside money. Around the group is a carefully drawn network of researchers and engineers who have worked at trading firms and AI labs, and we bring them in by project.
We work from Shoreditch, a short walk from Bunhill Fields, where Thomas Bayes is buried. From the Artillery Ground next door, in 1784, Vincenzo Lunardi went up in a balloon, the first flight in England with a person aboard. In 1897 the parish burned its own rubbish to make electricity and light its streets, among the first in the country to do so, under a Latin motto that reads light and power from dust. And for a century and more the small workshops off Curtain Road made much of London’s furniture, durable, useful things made by hand, which is the kind of making we have in mind.
Positions lists the three open roles.
Contact
Write with a problem or a question. If you want to work here, Positions says how.
