Scalable Visual Representation Learning

Model performance

At the Morgridge Institute for Research, I trained and benchmarked DINOv2, diffusion, and hierarchical VQ-VAE architectures on 1.5 million multi-channel Cell Painting images. DINOv2 representations increased gene-classification AUROC by 18% relative to an ImageNet-pretrained baseline.

Infrastructure performance

I implemented distributed training across 64 A100 GPUs using DeepSpeed ZeRO-3. The optimized training stack reduced epoch latency from 6.4 hours to 1.7 hours, delivering 3.8× higher throughput and 54% lower compute cost.

I also developed a discrete latent pipeline using VQ-VAE token spaces for conditional diffusion of unseen perturbations and temporal transitions. The project combined architecture benchmarking, multi-node training, performance profiling, and cost optimization.