Emojis have changed the way humans communicate today. They are the most convenient non-linguistic social cues available to us in this era of social media. But there is no methodology wherein users can creatively interact with their systems to generate personalised emojis or edit existing ones. While there have been some experiments that enable networks to create images, there is no comprehensive solution that gives users the control to create personalised emoticons. In this work, we propose an end-to-end architecture to create a realistic emoji from a roughly drawn sketch. Our generated emojis show a PSNR value of 20.30dB and a SSIM of 0.914. Additionally, we look at a multi-modal architecture which generates an emoji when given an incomplete sketch along with a handwritten word describing the associated emotion.

Auto-completion is a major feature in all keyboard based mobile devices. However, not a lot of work has been done to extend this feature into handwritten text based input. Generally, auto-complete options are ranked on the basis of previous context and word probabilities. However, this leads to missing out on the user intended word simply because there are more frequently used words with the same prefix, in the same context. In this paper, we propose a gesture-based solution to recognize and complete partially-written words, where the user indicates to the system the missing text length by a stroke gesture, thus ruling out a huge subset of possibilities. We also propose a length dependent hybrid score ranking system, to improve the prediction accuracy and speed. Our results show that using a gesture stroke length based searching method not only reduced the processing time by 27% but also showed an accuracy increase of 10%, when the top 3 candidates are considered.

This paper demonstrates two novel methods to estimate the global SNR of speech signals. In both methods, Deep Neural Network-Hidden Markov Model (DNN-HMM) acoustic model used in speech recognition systems is leveraged for the additional task of SNR estimation. In the first method, the entropy of the DNN-HMM output is computed. Recent work on bayesian deep learning has shown that a DNN-HMM trained with dropout can be used to estimate model uncertainty by approximating it as a deep Gaussian process. In the second method, this approximation is used to obtain model uncertainty estimates. Noise specific regressors are used to predict the SNR from the entropy and model uncertainty. The DNN-HMM is trained on GRID corpus and tested on different noise profiles from the DEMAND noise database at SNR levels ranging from -10 dB to 30 dB.

We propose a Low-Dimensional Deep Feature based Face Alignment (LDFFA) method to address the problem of face alignment “in-the-wild”. Recently, Deep Bottleneck Features (DBF) has been proposed as an effective channel to represent input with compact, low-dimensional descriptors. The locations of fiducial landmarks of human faces could be effectively represented using low dimensional features due to the large correlation between them. In this paper, we propose a novel deep CNN with a bottleneck layer which learns to extract a low-dimensional representation (DBF) of the fiducial landmarks from images of human faces. We pre-train the CNN with a large dataset of synthetically annotated data so that the extracted DBFs are robust across variations in pose, occlusions, and illumination. Our experiments show that the proposed approach demonstrates near real-time performance and higher accuracy when compared with state-of-the-art results on numerous benchmarks.

On-line handwriting recognition has seen major improvements in the last few years with the successful application of deep learning techniques. While the training data is massive and training times are large, the current accuracies are very good even with reasonable small model sizes (when compared to models in image recognition tasks). However, the important task of personalizing such deep models (after a new user has downloaded the model onto his mobile device) to the writing style of the new user, resulting in an improved model with significantly higher accuracy, has not been investigated so far. It is clearly not feasible to download the massive training data for each new user, and re-train the model on the device itself (after mixing original data with the limited quantity of training samples a new user can supply). To our knowledge, it is not known so far with what adaptation technique, with what old training data (in addition to the user's new samples), how flexibly the retraining can be done on the device, how much accuracy gain can be expected, and whether the model can be simplified as a result of the retraining. In this paper we address all these important questions regarding deep network models on device: we demonstrate simple but robust continued-training techniques for adapting a pre-trained model to a specific user's writing style. The pre-training is done using Long Short Term Memory (LSTM) network and a Connectionist Temporal Classification (CTC) loss function. It is seen that only a small set of the old training data can be used along with the user's samples to successfully retrain the model. The retrained (adapted) model achieves around 2.5% more accuracy than the roughly 92% accuracy of the non-personalized model for on-line handwritten Hindi text recognition, also surpassing all existing techniques. We also derive some additional insights into the questions posed above.

The maximum cut problem is known to be an important NP-complete problem with many applications. The authors investigate this problem (which they call the normal maximum cut problem) and a variant of it (which is referred to as the connected maximum cut problem). They show that any n-vertex e-edge graph admits a cut with at least the fraction 1/2+1/2n of its edges, thus improving the ratio 1/2+2/e known before. It is shown that it is NP-complete to decide if a given graph has a normal maximum cut with at least a fraction (1/2+ epsilon ) of its edges, where the positive constant epsilon can be taken smaller than any value chosen. The authors present an approximation algorithm for the normal maximum cut problem on any graph that runs in O((e log e+n log n)/p+log p*log n) parallel time using p(1>

We give a new characterization of the planar separator theorem in terms of mutually non-containing closed Jordan Curves. using this, we develop an O(n √n logn) maximum flow algorithm for directed planar networks (hence for any planar network).

Algorithms are given that compute maximum flows in planar directed networks either in O((logn)3) parallel time using O(n4) processors or O((logn)2) parallel time using O(n6) processors. The resource consumption of these algorithms is dominated by the cost of finding the value of a maximum flow. When such a value is given, or when the computation is on an undirected network, the bound is O((logn)2) time using O(n4) processors. No efficient parallel algorithm is known for the maximum flow problem in general networks.