About
I'm an AI researcher and engineer with a PhD in Machine Learning and Meta-Learning from the University of Edinburgh. I've worked at Google, Amazon, and as principal scientist across AI startups in the UK and US. I specialize in large language models, multimodal learning, and self-supervised methods.
I believe intelligence emerges from structure, interaction, and information—not from sheer scale alone. The field has converged on one dominant recipe: make it bigger. But that's not how natural intelligence works. Brains compress experience, maintain state, and build hierarchical representations. They don't re-ingest the universe every time they need to think.
My work focuses on the first principles of learning—how structure, interaction, and information theory can replace massive parameter counts. I'm driven by that Cambrian Explosion spirit of the 2010s: new architectures, new paradigms, genuine exploration of the research tree.
Current Work
Cofounder & Research Lead
An open-science frontier AI company. We're exploring what's beyond the current scaling paradigm—better learning signals, novel architectures, and systems that actually understand.
Axiotic is also where I turn these ideas into deployed systems: LLM-assisted expert consensus platforms, research tooling, creative-companion workflows, and practical AI infrastructure for people who need more than a chatbot.
Open to a small number of advisory and technical due-diligence engagements for VCs and AI startups: eval design, data moats, model strategy, efficiency claims, and research-roadmap pressure-testing.
iam@antreas.io →Philosophy
My goal is to emulate the trajectory of human-like representation learning—starting from foundational representations akin to those seen in human infants, but without requiring eons of evolutionary fine-tuning. This serves as a springboard for my broader ambition: investigating how these infant-like representations can be fine-tuned in concert with higher-level abstract concepts to pave the way for general artificial intelligence. To achieve this, my research focuses on scalable, data-efficient, and generalizable self-supervised learning in a multimodal setting. I integrate insights from neuroscience and evolutionary computation to explore optimal learning sequences and curricula, paying close attention to architectural choices and their corresponding training recipes.
Leading my research interests is Multi-Modal Learning—the synergistic integration of text, images, audio, and video. This is followed by Self-Supervised Methods inspired by infant learning and evolutionary computation.
I operate within a pragmatic framework, aiming to identify high-leverage focal points conducive to in-depth investigation. This allows for efficient allocation of both computational and cognitive resources. In line with evolutionary tenets and the Pareto Principle, my methodology focuses on the "fittest" 20% of research avenues likely to contribute 80% of impactful results.
Research Agenda
I'm fundamentally inspired by information theory, evolution, and biology—how can we use what we observe about human learning and nature to build better artificial intelligence?
Beyond next-token prediction. Predicting masked past, present, and future states to force causal structure, long-horizon planning, and robust memory.
Inspired by the brain's multimodal neurons—"synesthetic" architectures processing text, images, and audio in a shared latent space.
Explicit memory and latent reasoning loops turning transformers into differentiable computers—infinite-context learning without expensive CoT tokens.
Fractal attention patterns operating at multiple resolutions simultaneously—seeing the forest and the trees without brute-force scaling.
New connectivity patterns, activation mechanisms, and dense routing schemes where deep layers attend directly to shallow features.
Data quality is the new frontier. Closed-loop engines where models generate, filter, and curriculum-sort their own training data.
Intelligence requires grounding. Using simulation as a primary data source, training agents on interaction physics before sim-to-real transfer.
SGD is greedy; evolution generalizes but is slow. Researching how to synergize them for robust, efficient learning.
Distillation through an information-theoretic lens—cracking the physics of compression to train small models to their maximum potential.
Efficiency in nature is achieved through local learning rules. Leveraging this for efficient learning without global backprop.
Rhythms across multiple temporal scales define natural systems. Exploring how neural architectures can internalize these for efficient learning.
LLMs as coordination and compression engines: mapping agreement, disagreement, uncertainty, and evidence across expert groups without replacing judgement.
Nature solved learning over billions of years. Drawing from biological neural networks, synaptic plasticity, and adaptive mechanisms.
Publications
A curated view: the three latest papers plus three highly cited foundations. See all on Google Scholar →
L. Ericsson, M. Espinosa, C. Yang, A. Antoniou, A. Storkey, S.B. Cohen, S. McDonagh, E.J. Crowley
A unified search space for neural architecture search based on einsum operations and fundamental building blocks.
A. Antoniou, E. Triantafillou, H. Larochelle, S. Montella, F. Rezk, K. Kim, L. Ericsson, P. Vougiouklis, J. Engelmann, E.J. Crowley, S. Humbarwadi, Y. Liu, G. Yang, J.Z. Pan, A. Storkey
Rethinking evaluation benchmarks for vision-language models to better measure true capabilities.
K. Kim, S.N. Gowda, P. Eustratiadis, A. Antoniou, R.B. Fisher
Using adversarial augmentation to improve robustness of action recognition models against natural distribution shifts.
T. Hospedales, A. Antoniou, P. Micaelli, A. Storkey
A comprehensive overview of the meta-learning landscape—how it works, why it matters, and where it's going.
A. Antoniou, H. Edwards, A. Storkey
MAML is elegant but tricky to train. We figured out what actually matters and what doesn't.
A. Antoniou, A. Storkey, H. Edwards
Using GANs to create training data. Especially useful when you only have a few examples to work with.
Teaching
A single home for lecture materials, invited talks, workshops, tutorials, and teaching recognition.
A teaching package on using generative AI as a collaborator for writing, music, visual storytelling, and agentic creative workflows.
Alternative slide treatments for the same teaching material: terminal-native, cinematic, and interactive versions.
A standalone teaching session on using Cursor, Claude, Codex, and agentic workflows for practical software development.
The latest LLM-101 session on model compression, efficient language models, and the practical tradeoffs behind smaller systems. Earlier recordings remain archived below under Quest Lectures.
University of Edinburgh and invited sessions · 2025–2026
Machine Learning Systems Course · 2026
Teach-A-Thon presentation
Workshop on using LLMs in education
What I learned from extensive LLM usage
Practical guide to coding with AI assistants
Introduction to UoE's compute infrastructure
Lead TA at University of Edinburgh. Created comprehensive tutorials, coursework materials, and supervised research projects. Course Website →
University of Edinburgh — Recognition for exceptional teaching assistance
View Award →Background
Axiotic AI
Building an open-science frontier AI company. Exploring efficient learning, novel architectures, and what's beyond scaling.
Pieces for Developers
Led ML research agenda. Built nano-models and foundation models for efficient deployment across CPUs, GPUs, NPUs, and LPUs.
Malted AI
Efficient LLM training, distillation, synthetic data generation, and LLM-as-a-Judge systems.
Amazon
Applied machine learning research and development.
Machine learning research internship during PhD studies.
University of Edinburgh
Supervised by Prof. Amos Storkey. Part of BayesWatch and the Adaptive and Neural Computation (ANC) institute.
University of Edinburgh
Thesis: "Meta Learning for Supervised and Unsupervised Few-Shot Learning"
Community
Building communities and democratizing ML research
Lead Organizer & Competitor · 2025
Originally conceived and initiated bringing the global LeRobot Hackathon to the University of Edinburgh. Built the organizing team and managed 60% of organizational work. Sourced and provided all specialized hardware and high-performance GPU infrastructure. Served as expert mentor for all teams throughout the 30-hour competition. Also competed and led my team to 7th place out of 1000+ teams worldwide.
Early Adopter & Community Lead · 2023
One of the early adopters and key community support members for Edinburgh's EIDF A100 GPU cluster. Created the Slack server, answered hundreds of user questions, scheduled community meetings for key issues, and served as a bridge between users and developers. Authored early documentation and developed kubejobs, a Python package simplifying Kubernetes job specifications.
Co-organizer · 2024-2025
Co-organized the International Workshop on Efficient Generative AI 2024, a GAIL-funded event bringing together leading researchers from academia and industry to discuss efficient approaches to generative AI.
A journey across modern neural network inductive biases
Discovering learning priors across domains, tasks, and modalities
Temporally and semantically Aligned Audio, Language and Images
...and other mythical creatures
NeurIPS Workshop paper presentation
NeurIPS 2019 paper presentation
Farewell presentation at University of Edinburgh
Cat ownership, biology, evolution & psychology
Pieces for Developers Podcast · 2025
A practical collection of academic writing habits and go-to paper-writing resources
Pieces Blog · August 2025 — A vision for local-first AI built on biology-inspired architectures
Pieces Blog · July 2025 — The "Great Amnesia" of AI's scaling monoculture
My conviction in the democratization of ML research stems from the irreplaceable value of individual expertise and the power of collective collaboration. Open source is not merely a development model—it's a fundamental necessity for driving innovation and maintaining ethical standards in the field.
User-friendly framework simplifying MAML meta-learning. Top-3 finalist in UK Open Source Awards.
Scalable ML template designed to minimize setup overhead for researchers.
Python package simplifying Kubernetes job management for ML workloads.
Multi-domain, multi-task benchmark suite for diversifying empirical evaluations.
University of Edinburgh · 2021-2024
Playground
Interactive explorations of my research. Dive in!
TALI is my response to the growing demand for multimodal understanding in deep learning. It's a large-scale, tetramodal dataset that aligns text, video, images, and audio—a playground for innovative self-supervised learning tasks and multimodal research. With TALI, we're exploring how different modalities and data/model scaling affect downstream performance. I'm excited about the diverse research ideas it will inspire!
Symphonia is a platform for synthesising structured expert judgement. Experts submit positions, reasoning, evidence, and confidence; an LLM drafts a transparent map of agreement, disagreement, and uncertainty; participants then refine it through iterative feedback cycles. The point is not to automate judgement — it is to make collective judgement cheaper, clearer, and more scalable.
This is the same research story in applied form: LLMs are most powerful when used as coordination and compression engines over human expertise, preserving disagreement rather than flattening it.
More demos coming soon...
Contact
Interested in research collaboration, consulting, or just want to chat about AI?
