• Publications
  • Influence
Universal Sound Separation
TLDR
A dataset of mixtures containing arbitrary sounds is developed, and the best methods produce an improvement in scale-invariant signal-to-distortion ratio of over 13 dB for speech/non-speech separation and close to 10 dB for universal sound separation.
cGANs with Multi-Hinge Loss
TLDR
The proposed new algorithm to incorporate class conditional information into the discriminator of GANs via a multi-class generalization of the commonly used Hinge loss shows that learning a single good classifier and a single state of the art generator simultaneously is possible in supervised and semi-supervised settings.
A Multi-Class Hinge Loss for Conditional GANs
TLDR
The trade-off between a generator-critic pair respecting class conditioning inputs and generating the highest quality images is shown, and a way to use the algorithm to measure the degree to which a generator and critic are class conditional is proposed.
Scattering transforms and classification of hyperspectral images
TLDR
A spatial-spectral scattering transform that combines Wavelet and Fourier representations is presented that obtains significantly higher classification accuracies and the results indicate that the Fourier scattering transform is effective at representing spectral data.
3-D Fourier Scattering Transform and Classification of Hyperspectral Images
TLDR
The 3-D Fourier scattering transform is introduced, an amalgamation of time–frequency representations with NN architectures that leverages the benefits provided by the short-time Fourier transform with the numerical efficiency of deep learning network structures.
Exploring the High Dimensional Geometry of HSI Features
We explore feature space geometries induced by the 3-D Fourier scattering transform and deep neural network with extended attribute profiles on four standard hyperspectral images. We examine the
Three-Dimensional Fourier Scattering Transform and Classification of Hyperspectral Images
TLDR
Results indicate that the Fourier scattering transform is highly effective at representing spectral data when compared with other state-of-the-art spectral-spatial classification methods.
Cortical Features for Defense Against Adversarial Audio Attacks
TLDR
This work applies several white-box iterative optimization-based adversarial attacks to an implementation of Amazon Alexa’s HW network and a modified version of this network with an integrated cortical representation, and shows that the cortical features help defend against universal adversarial examples.
Harmonicity Plays a Critical Role in DNN Based Versus in Biologically-Inspired Monaural Speech Segregation Systems
TLDR
The role of harmonicity on two state-of-the-art Deep Neural Networks (DNN)-based models Conv-TasNet and DPT-Net is analyzed, finding that performance deteriorates significantly if one source is even slightly harmonically jittered, making inharmonicity a powerful adversarial factor in DNN models.
...
...