Jonas Geiping

cramming. Cramming the training of a (BERT-type) language model into limited compute.

breaching. Breaching privacy in federated learning scenarios for vision and text

invertinggradients. Algorithms to recover input data from their gradient signal through a neural network

poisoning-gradient-matching. Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching

data-poisoning. Implementations of data poisoning attacks against neural networks and related defenses.

linear_cross_entropy_loss. A fusion of a linear layer and a cross entropy loss, written for pytorch in triton.

carving. Package to optimize Adversarial Attacks against (Large) Language Models with Varied Objectives

fullbatchtraining. Training vision models with full-batch gradient descent and regularization

dataaugs. Python

torchlmdb. LMDB wrapper for pytorch datasets.

poisoning-benchmark. A unified benchmark problem for data poisoning attacks

ParametricMajorization. New training strategies for energy minimization methods with learnable parameters (ICCV 2019)

meta-shampoo. For optimization algorithm research and development.

flexBox. FlexBox is a fexible MATLAB toolbox for finite dimensional convex variational problems in image processing and beyond.

flash-attention. Fast and memory-efficient exact attention