Bezawit Abebaw — Machine Learning Researcher

I am passionate about understanding how artificial intelligence systems learn and what happens inside them. For the past two years, I have been exploring learning algorithms inspired by biological intelligence. Rather than relying solely on backpropagation, I have experimented with predictive coding, where models learn through local prediction errors at each layer instead of relying on a single global backward pass.

I am currently pursuing a Master's in Artificial Intelligence and Machine Learning at Adelaide University. My research interests include efficient and ethical AI, continual learning, and biologically inspired approaches to intelligence.

Research

Nearly every deep learning model in use today is trained with backpropagation: one error signal is computed at the output and propagated backwards through every layer. It works remarkably well. It also bears little resemblance to how brains appear to learn.

Predictive coding offers a local alternative. Each layer predicts the activity of the layer above it, measures its prediction error, and updates itself using that local error rather than relying on a global backward pass.

My work asks whether a transformer built on this principle can hold up in practice. Answering that honestly takes work on two fronts. The research side focuses on the architecture itself: how layers should connect, what energy functions to use, which weight and latent update rules to apply and when, whether to add lateral connections, and how to set scaling factors. The engineering side is about making those findings trustworthy: distributed training, hyperparameter search, controlled ablations, reproducible configurations, and careful evaluation of what the model learns, including its representations and internal behavior.

Selected work

PC-Transformers

2025–2026

A transformer language model rebuilt around predictive coding. The embeddings, attention, MLP, and output layers each maintain a latent state, predict the activity of the next layer, and update from their own prediction error over T inference steps. I led the research and built the training stack, including the model architecture, multi-GPU training, KV caching, a two-phase hyperparameter search, and experiment visualizations.

Case study·Code

Neural Generative Coding replication

2025

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.

Case study·Code·Paper

FabricPC

2026

A JAX library from SingularityNET for building predictive coding networks, designed to train the same architecture with either predictive coding or backpropagation for direct comparison. I contributed transformer components, regularization techniques, and BPE tokenization, along with a demo for training a simple model using both predictive coding and backpropagation.

Case study·Code

Other projects

Does deblurring help object detection?

A computer vision pipeline investigating whether image deblurring improves object detection. I compared Wiener filtering and Richardson-Lucy deconvolution with YOLOv8, evaluating both image quality and detection performance.

Can AI recognize emotions in abstract art?

A computer vision project exploring emotion recognition from abstract art using the ArtEmis dataset. I implemented a Vision Transformer from scratch and compared it with ResNet-50, SVM, and Naive Bayes baselines, evaluating their performance under the same experimental setup.

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.

LLMs for low-resource languages

An NLP project exploring text generation for Amharic, a low-resource language. I built a character-level bigram model and a GPT-style transformer from scratch, and experimented with adapting pretrained GPT-2 models at different scales for Amharic text generation.

Experience

Machine Learning Research Engineer · Technical Team Lead, SingularityNET

2024–2026

Machine Learning Engineer, iCog Labs (internship)

2024

Software Engineer, Vintage Technologies (internship)

2022

Education

MSc Artificial Intelligence and Machine Learning, Adelaide University

2026–present

BSc Computer Science, University of Gondar

2019–2023