Learning Robust Representations (LR2)
Description
Sponsored by DARPA’s GARD program, LR2 develops deep learning models that are robust to adversarial attacks. Our work covers models for numerous tasks spanning different modalities, such as image classification, object detection, action recognition, object tracking, speaker identification, and automatic speech recognition. Our general approach is layered defenses, where inner layers of the defense deploy modality-agnostic techniques based on robust representation learning, and the outer layers deploy modality specific techniques, such as attack detection and removal.
Team
Collaborating Institutions
Publications
September, 2021
arXiv preprint
In this paper, we introduce a novel non-linear activation function that spontaneously induces class-compactness and regularization in the embedding space of neural networks. The function is dubbed DOME for Difference Of Mirrored Exponential terms. The basic form of the function can replace the sigmoid or the hyperbolic tangent functions as an output activation function for binary classification problems. The function can also be extended to the case of multi-class classification and used as an alternative to the standard softmax function.
July, 2021
Computer Speech & Language
In this paper, we introduce a novel non-linear activation function that spontaneously induces class-compactness and regularization in the embedding space of neural networks. The function is dubbed DOME for Difference Of Mirrored Exponential terms. The basic form of the function can replace the sigmoid or the hyperbolic tangent functions as an output activation function for binary classification problems. The function can also be extended to the case of multi-class classification and used as an alternative to the standard softmax function.
June, 2021
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
In this paper, we introduce a novel non-linear activation function that spontaneously induces class-compactness and regularization in the embedding space of neural networks. The function is dubbed DOME for Difference Of Mirrored Exponential terms. The basic form of the function can replace the sigmoid or the hyperbolic tangent functions as an output activation function for binary classification problems. The function can also be extended to the case of multi-class classification and used as an alternative to the standard softmax function.
Sponsors
Defense Advanced Research Project Agency
SPONSOR PROGRAM: DARPA – GARD