engblogs

Analyzing the mixing time of the Langevin Algorithm for log-concave sampling to its stationary distribution

Efficient and effective train-time pruning scheme, Parameter-free Differentiable Pruning (PDP), offers state-of-the-art qualities in model size, accuracy, and training cost.

Expanding the population for training language models using federated learning and differential privacy to improve model utility and training latency

The blogpost discusses the conformalization of sparse generalized linear models and the challenges of computing confidence sets in regression problems.

The blogpost discusses DUET, a proposed method for 2D structured and equivariant representations that maintain transformation-related information in multiview self-supervised learning.

Practical heuristics to improve performance of prefix-tree based algorithms for differentially private heavy hitter detection, utilizing aggregate and local differential privacy with an adaptive hyperparameter tuning algorithm and exploration of different data-selection schemes.

The blogpost explores UPSCALE, a technique that addresses the latency issues introduced by channel pruning in neural networks.

Examining the role of entropy and reconstruction for multi-view self-supervised learning (MVSSL) and its relation to mutual information (MI).

Optimizing AWS Inferentia utilization with FastAPI and PyTorch models on Amazon EC2 Inf1 & Inf2 instances for improved performance and cost benefits

Patsnap explains how they used GPT-2 inference on Amazon SageMaker to achieve low latency and cost.

The blogpost discusses the impact of Generative AI and its position on the Gartner AI Hype Cycle.

Determining which generation is most committed to maintaining a streak of 365+ days and ranking them accordingly.