1

Gradvit: Gradient inversion of vision transformers

Images can be recovered from averaged gradients in Transformers, they are more vulnerable than CNNs.

Ali Hatamizadeh, Hongxu Yin, Holger Roth, Wenqi Li, Jan Kautz, Daguang Xu, Pavlo Molchanov

AViT: Adaptive Tokens for Efficient Vision Transformer

Transformer with adaptive inference where simpler images are classified faster. Tokens are automatically stopped at various depth once become irrelevant. Learned via differentiable loss inspired by ACT.

H. Yin, A. Vahdat, J. Alvarez, A. Mallya, J. Kautz, P. Molchanov

DexYCB: A benchmark for capturing hand grasping of objects

Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, others

Data-free knowledge distillation for object detection

Akshay Chawla, Hongxu Yin, Pavlo Molchanov, Jose Alvarez

Global Vision Transformer Pruning with Hessian-Aware Saliency

Global pruning of vision transformer networks. New parameter redistribution rule for ViT. 2x latency reduction with minor acucracy loss. More than 1.4% accuracy gain when pruned towards smaller model.

H. Yang, H. Yin, M. Shen, P. Molchanov, H. Li, J. Kautz

Kama: 3d keypoint aware body mesh articulation

Umar Iqbal, Kevin Xie, Yunrong Guo, Jan Kautz, Pavlo Molchanov

Optimal quantization using scaled codebook

Yerlan Idelbayev, Pavlo Molchanov, Maying Shen, Hongxu Yin, Miguel A Carreira-Perpinán, Jose M Alvarez

Optimizing Selective Protection for CNN Resilience

Abdulrahman Mahmoud, Siva Kumar Sastry Hari, Christopher Wardlaw Fletcher, Sarita V Adve, Charbel Sakr, Naresh Shanbhag, Pavlo Molchanov, Michael B Sullivan, Timothy Tsai, Stephen W Keckler

See through gradients: Image batch recovery via gradinversion

Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, Pavlo Molchanov

Dreaming to Distill: Data-free Knowledge Transfer via DeepInversion

Statistics stored in batch normalization layers contain information on training data. Via iterative optimization we recover images from train distribution and use for various applications.

H. Yin, P. Molchanov, J. M. Alvarez, Z. Li, A. Mallya, D. Hoiem, N..Jha, J. Kautz