Thu Mar 16 2023
Tue Mar 14 2023

Eliciting Latent Predictions from Transformers with the Tuned Lens

Deep Learning
Natural Language Processing
Cybersecurity
improving language models
detecting malicious inputs

Analyzes transformers from the perspective of iterative inference, seeking to understand how model predictions are refined layer by layer.

Can improve the accuracy, reliability, and bias of autoregressive language models through the use of the tuned lens method. Can also detect malicious inputs with high accuracy.

MeshDiffusion: Score-based Generative 3D Mesh Modeling

Generative Modeling
Computer Graphics
Simulation
automatic scene generation
physical simulation

Considers the task of generating realistic 3D shapes using a generative modeling method that takes advantage of the graph structure of meshes.

Can generate more realistic 3D shapes with fine-grained geometric details, which is useful for automatic scene generation and physical simulation.

Mon Mar 13 2023
Sun Mar 12 2023
Thu Mar 09 2023
Wed Mar 08 2023