Vivek
Ramanujan

Studies when and how better image reconstruction leads to better generation. CRT achieves state-of-the-art ImageNet generation (2.18 FID) with 2-3x improved compute efficiency even with worse reconstruction.

360-1M dataset and Odin model for novel view synthesis from the largest real-world multi-view dataset to date.

Synthetic data can be beneficial, but is consistently matched or outperformed by real images from a simple retrieval baseline.

Data quantity is the primary factor influencing downstream robustness, while other factors have limited impact.

Enabling large pretrained models to adapt to distribution shifts with minimal labeled data.

A benchmark for multimodal dataset creation with 12.8B image-text pairs. DataComp-1B achieves 79.2% zero-shot ImageNet accuracy.

Learning object-centric representations through novel view reconstruction using a dictionary of object codes.

Flexible representations that adapt to multiple downstream tasks with varying computational resources.

Preparing for future model versions by saving cheap auxiliary information about present training.

As parameters grow in magnitude, networks approximate discretized networks with saturated activations—a new characterization of inductive bias in GD.

Learning extremely low-dimensional binary codes (~20 bits for ImageNet-1K) while maintaining near-optimal accuracy.

Hidden networks for continual learning—learning thousands of tasks without catastrophic forgetting.

Finding untrained subnetworks at initialization that match trained network performance.

A pruning strategy based on soft threshold reparametrization, enabling very sparse but highly performant trained models.

Improving on CycleGAN by allowing better shape deformation between more disparate domains.