MEng, Artificial Intelligence · Carnegie Mellon University
Building machine learning that reads the signals around us, from cardiac sounds to medical images, with a focus on healthcare and social good.
I am a research associate, teaching assistant, and Masters graduate in Artificial Intelligence from Carnegie Mellon University's College of Engineering. My research focuses on using machine learning to process, understand, and make inferences from signals in the world around us, with a particular interest in visual and multimodal representations.
Prior to CMU, I earned a Bachelor's degree with First Class Honors in Software Engineering from The ICT University in Cameroon, graduating first in my department. I currently work with the AI Healthcare Research Lab at CMU, where I work on machine learning methods for understanding cardiac health through the joint analysis of electrocardiogram (ECG) and phonocardiogram (PCG) signals.
Beyond work, I enjoy traveling, reading, and watching documentaries.
Email: bfonya [at] alumni [dot] cmu [dot] edu
Efficient deep learning for computer vision and healthcare, with a focus on representation learning, medical imaging, and AI that benefits humanity.
A zero-shot EEG decoding framework that learns subject and task-invariant neural priors from large-scale HBN data, enabling robust cross-subject generalization without calibration.
Bidirectional self-supervised representation learning with joint-embedding predictive architectures for learning robust representations from unlabeled data.
An empirical study showing how signal preprocessing can act as an implicit prior in self-supervised learning for cardiac sound classification.
Analyzed health facility distribution in Rwanda using geospatial mapping and ML to propose equitable resource allocation based on disease prevalence.
Introduced VARET, a valve-aware transformer that separately evaluates robustness to missing auscultation sites and shifts across patient cohorts for multi-site phonocardiogram classification. The framework demonstrates improved patient-level classification performance and reveals limitations in zero-shot cross-cohort generalization.
Evaluated the impact of PCG preprocessing filter choices and deep learning architectures (LSTM and CNNs) on murmur severity classification, with the aim of understanding their influence on classification performance and severity calibration. Found that filter choice had minimal impact on overall accuracy but introduced consistent severity prediction bias, while model architecture had a significantly stronger effect, highlighting important considerations for reliable cardiac screening and triage in low-resource settings.
A controlled benchmark of five SSL objectives (SimCLR, BYOL, MAE, DINO, JEPA) for heart sound screening. We showed that SSL helps only in the low label regime, when labeled data is scare: JEPA and MAE gain 9.44 and 8.89 balanced-accuracy points at 10 percent of labels, but the advantage decays to nothing at full supervision, and no method survives transfer to an external pediatric cohort.
Localizes where clinically relevant heart sound information is lost when phonocardiograms are compressed into discrete tokens. Two distinct bottlenecks emerge: tokenizer training determines what survives compression, while interface geometry determines what a bounded model can access. A factorized categorical interface holds murmur recoverability near constant across a 3.9-fold rate reduction, while encoding the same states as flat token IDs drops it sharply at every rate.
Outside of research, I mentor students in research and across the STEM fields, helping early-career scientists and engineers find their footing and build work they are proud of.
I am also happiest with a boarding pass in hand. I love traveling the world, meeting new people, and collecting stories from every place I land.
A quick ML quiz. Beat the clock, build a streak.
8 rounds, 8 seconds a question, faster answers score more.