Through Mystiko, we have demonstrated that private, federated gradient descent without loss of accuracy can be practical with reasonable synchronization overhead. Mystiko is the first federated learning platform to provide the same security and priv…

We have presented a new way to exploit pre-trained GANs for the task of large-scale super-resolution, up to 64× upscaling factor. We have shown that a pre-trained GAN can be used as a generative latent bank in an encoder-bank-decoder architecture. R…

In this section, we prove our main result and show that the GAN iteration converges and learns the one-layer Generator network (W∗,ϕ). Denote Σ∗=W∗WT∗. Without loss of generality, we can assume the underlying target network has the form (Σ1/2∗,ϕ) as…

Recently the majority of the research has been focused on proposing sophisticated newer models to train word embeddings. While newer models do provide better accuracies, they use classical models with default hyper-parameters as a baseline to showca…

We have introduced RAFT-3D, an end-to-end network for scene flow. RAFT-3D uses rigid-motion embeddings, which represent a soft grouping of pixels into rigidly moving objects. We demonstrate that these embeddings can be used to solve for dense and ac…

In this work, we studied the necessity of utilizing coverage as a metric for training of DRL tasks. We proposed a new metric to review the quality of the sampled states for high-dimensional state space. For the two standard tasks, we provided the ps…

We have presented the Panoptic FCN, a conceptually simple yet effective framework for panoptic segmentation. The key difference from prior works lies in that we represent and predict things and stuff in a fully convolutional manner. To this end, ker…

In this work, we investigated the application of fully convolutional deep neural networks for forecasting wave dynamics on fluid surfaces. In particular, we focused on a U-Net architecture with two skip connections, and we trained the network to pre…

A NN-based symbol detector for MIMO-OFDM systems is introduced in this paper. The NN incorporates structural knowledge of MIMO-OFDM signals that characterizes features in time, frequency and constellation where the training overhead is significantly…

We have described Sigma Worksheet, a new interface for visually constructing data warehouse queries using direct manipulation. Sigma Worksheet provides users with an expressive interface for constructing OLAP queries, as well as queries for data tra…

In this paper, we have shown that inference attacks against virtual keyboards can work with external microphones. While previous work which showed that the microphones in a device can be used to deduce PINs and passwords entered on its screen, we ha…

We have proposed a novel Retrieval-IVAE model, extension of the identifiable VAE for cross-modal retrieval, where it completely removes the ambient data decoding module via implicit encoder inversion that is achieved by Jacobian regularization of th…