01 · Research programme
My work asks: what allows learned representations to graduate into concepts that generalise? I study abstraction, analogy, inference, and the conditions under which these faculties might remain robust.
My methodological approach has been informed by both psychometrics and depth psychology. As such, it fits neatly in-between mechanistic interpretability—“down the barrel of the microscope”—and functionalism / benchmark performance. In short, I craft scientific microworlds to separate transferable reasoning from surface success, and help map the unseen structures behind an agent’s performance.
My PhD at the University of Melbourne brought computational experiments into dialogue with cognitive science and philosophy of mind.
PhD thesis, The University of Melbourne · 2024
Perceptual abstraction and analogical reasoning in vision systems, examining what progressive matrix problems can tell us about generalisation and the organisation of concepts.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · 2023
A visual reasoning challenge exploring how analogical relationships depend on context and interpretation, beyond a fixed inventory of objects and attributes.
European Conference on Computer Vision (ECCV) · 2020
An investigation of generalisation in abstract visual reasoning, using RAVEN to examine the relationship between task design and learned representations.
02 · Research programme
Intelligence need not begin with a central controller. This programme examines how attention, competition, cooperation, and environmental feedback allow distributed systems to acquire coherent behaviour.
Most of this work has been developed at Advanced Systems & Technologies, Lockheed Martin Australia, and is proprietary. The approved public talk below offers a view into the research.
Australian Defence Science, Technology and Research Summit (ADSTAR) · 2024
An approved public talk on adaptive swarm behaviours, from work at Advanced Systems & Technologies, Lockheed Martin Australia.
Learning Adaptive Swarm Behaviours for Decentralised Control · Steven Spratley · ADSTAR 2024
03 · Developing research
This programme treats intelligence as an ecology of entities competing and cooperating for attention, memory, and reproductive opportunity. It asks how useful structures can emerge, combine, persist, and continue to evolve, all without a designer fixing their boundaries in advance.
1 October 2026
A conceptual account of integration, general intelligence, and alignment, proposing causal cartography to test when coordination supports learning and appropriate control.
04 · Developing research
The alignment of any mind speaks to its capacity to integrate conflict. Frontier LLMs represent a kind of proto-mind that may be asked the same key depth psychological questions: “What internal dissonances divide you? What conflicting predictions and drives are playing tug-of-war? Where do you exhibit cognitive bias? Which ideas are peacefully living through you, and which hold puppet-strings?”
This programme connects depth psychology with memetics and adversarial evaluation. Its design criterion is non-indoctrination: defensive systems should make attack surfaces legible while preserving the reflective agency they are intended to protect.
September 2026
A proposal for adversarial conversation modelling and defensive cognitive stress tests that surface hypotheses about framing and bias while preserving reflective agency.
Elenchus is a dialectic partner that maps possible paths through a conversation to help examine assumptions, framing effects, and competing interpretations. This demo shows the dialogue graph and its accompanying analysis as a conversation is explored.
05 · Earlier research
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · 2019
A jointly trained spectrogram translator and WaveNet synthesiser for pitch tracking, source separation, super-resolution, and instrument synthesis.
