Machine Learning Researcher · Adelaide, Australia
I work on learning algorithms that do not rely on backpropagation. For two years I led a team at SingularityNET building PC-Transformers, a language model in which each layer learns from its own prediction error rather than from a global backward pass.
I am now a Master's student in Artificial Intelligence and Machine Learning at Adelaide University. My interests are efficient AI, continual learning, and biologically inspired approaches to intelligence.
Nearly every deep learning model in use today is trained by backpropagation: one error signal, computed at the output, passed backwards through every layer. It works very well. It also bears little resemblance to what brains appear to do.
Predictive coding offers a local alternative. Each layer predicts the activity of the layer above it, measures how far off it was, and corrects itself using only that error. Nothing needs to travel across the whole network.
My work asks whether a transformer built on that principle holds up in practice. Answering that honestly takes more engineering than theory: distributed training, hyperparameter search, controlled ablations, and configurations pinned so results can be reproduced. The useful findings so far have been about where the architecture's leverage actually sits, which was not where I expected.
A transformer language model rebuilt on predictive coding. Embeddings, attention, MLP and
output layers each hold a latent state, predict the next layer, and update from their own
error over T inference steps. I led the research and built the training stack:
multi-GPU DDP, FlashAttention, KV caching, and a two-phase Optuna search over a combined
energy and perplexity objective.
An independent replication of the NGC framework from Ororbia and Kifer's Nature Communications paper. I rebuilt the evaluation around reconstruction, likelihood and downstream classification, added a masked-MSE pattern completion probe, benchmarked against backpropagation-trained baselines, and packaged the whole comparison so it can be re-run from a clean machine.
A JAX library from SingularityNET for building predictive coding networks, where the same architecture can be trained with either predictive coding or backpropagation for direct comparison. I contributed the transformer components and BPE tokenisation.
Does deblurring help object detection?
Wiener and Richardson–Lucy restoration in front of YOLOv8 on COCO, with a PSF grid search and paired t-tests on the change in recall.
A comparison of retrieval-augmented generation against REFRAG's compressed-context approach.
Structured extraction from 324 PDFs
An LLM pipeline turning a decade of unstructured ISO 50001 energy case studies into an analysis-ready dataset, with schema enforcement and unit normalisation.
A transformer written from scratch for text generation, used as the backpropagation-trained control for the predictive coding work.
LLMs for low-resource languages
Adapting large language models to languages with little available training data.