At SingularityNET, I researched biologically inspired learning approaches, including predictive coding and Neural Generative Coding, and developed a modular Predictive Coding Transformer with local inference and Hebbian weight updates. I reduced test perplexity from 56.65 to 7.12 on a character-level language model through architectural, initialization, inference, and hyperparameter improvements.
I also built benchmarking pipelines comparing predictive coding with backpropagation-based Transformers, and implemented multi-GPU training for predictive coding models. I contributed reusable PC-Transformer modules and language-model training pipelines to a JAX-based predictive coding library.
At iCog Labs, I developed a hybrid recommendation system and engineered a custom GPT architecture for Amharic NLP. I’m pursuing a Master of Artificial Intelligence and Machine Learning at Adelaide University.
