Abstract

This work studies the long-standing problems of model capacity and negative interference in multilingual neural machine translation (MNMT). We use network pruning techniques and observe that pruning 50-70% of the parameters from a trained MNMT model results only in a 0.29-1.98 drop in the BLEU score. Suggesting that there exist large redundancies even in MNMT models. These observations motivate us to use the redundant parameters and counter the interference problem efficiently. We propose a novel adaptation strategy, where we iteratively prune and retrain the redundant parameters of an MNMT to improve bilingual representations while retaining the multilinguality. Negative interference severely affects high resource languages, and our method alleviates it without any additional adapter modules. Hence, we call it parameterfree adaptation strategy, paving way for the efficient adaptation of MNMT. We demonstrate the effectiveness of our method on a 9 language MNMT trained on TED talks, and report an average improvement of +1.36 BLEU on high resource pairs. Code will be released

Abstract

rametric models that represent layout in terms of scene attributes are an attractive avenue for road scene understanding in autonomous navigation. Prior works that rely only on ground imagery are limited by the narrow field of view of the camera, occlusions and perspective foreshortening. In this paper, we demonstrate the effectiveness of using aerial imagery as an additional modality to overcome the above challenges. We propose a novel architecture, Unified, that combines features from both aerial and ground imagery to infer scene attributes. We quantitatively evaluate on the KITTI dataset and show that our Unified model outperforms prior works. Since this dataset is limited to road scenes close to the vehicle, we supplement the publicly available Argoverse dataset with scene attribute annotations and evaluate on far-away scenes. We show both quantitatively and qualitatively, the importance of aerial imagery in understanding road scenes, especially in regions farther away from the ego-vehicle. All code, models, and data, including scene attribute annotations on the Argoverse dataset along with collected and processed aerial imagery, are available

Abstract

Indian language machine translation performance is hampered due to the lack of large scale multi-lingual sentence aligned corpora and robust benchmarks. Through this paper, we provide and analyse an automated framework to obtain such a corpus for Indian language neural machine translation (NMT) systems. Our pipeline consists of a baseline NMT system, a retrieval module, and an alignment module that is used to work with publicly available websites such as press releases by the government. The main contribution towards this effort is to obtain an incremental method that uses the above pipeline to iteratively improve the size of the corpus as well as improve each of the components of our system. Through our work, we also evaluate the design choices such as the choice of pivoting language and the effect of iterative incremental increase in corpus size. Our work in addition to providing an automated

Abstract

In this paper, we present an audio-visual model to perform speech super-resolution at large scale-factors (8× and 16×). Previous works attempted to solve this problem using only the audio modality as input and thus were limited to low scale-factors of 2× and 4×. In contrast, we propose to incorporate both visual and auditory signals to superresolve speech of sampling rates as low as 1kHz. In such challenging situations, the visual features assist in learning the content and improves the quality of the generated speech. Further, we demonstrate the applicability of our approach to arbitrary speech signals where the visual stream is not accessible. Our “pseudo-visual network” precisely synthesizes the visual stream solely from the low-resolution speech input. Extensive experiments and the demo video illustrate our method’s remarkable results and benefits over state-of-the-art audio-only speech super-resolution approaches.

Abstract

Communication with the deaf community relies profoundly on the interpretation of sign languages performed by the signers. In light of the recent breakthroughs in sign language translations, we propose a pipeline that we term "Translating Sign Language Videos to Talking Faces". In this context, we improve the existing sign language translation systems by using POS tags to improve language modeling. We further extend the challenge to develop a system that can interpret a video from a signer to an avatar speaking in spoken languages. We focus on the translation systems that attempt to translate sign languages to text without glosses,

Abstract

The recent anomalies in b → slþl− transitions could originate from some new physics beyond the Standard Model. Either to confirm or to rule out this assumption, more tests of b → slþl− transition are needed. Polarized Λb decay to a Λ and a dilepton pair offers a plethora of observables that are suitable to discriminate new physics from the Standard Model. In this paper, we present a full angular analysis of a polarized Λb decay to a Λð→ pπÞlþl− final state. The study is performed in a set of the operator where the Standard Model operator basis is supplemented with its chirality flipped counterparts, and new scalar and pseudoscalar operators. The full angular distribution is calculated by retaining the mass of the final state leptons. At the low hadronic recoil, we use the heavy quark effective theory framework to relate the hadronic form factors which lead to simplified expression of the angular observables where short- and longdistance physics factorize. Using the factorized expressions of the observables, we construct a number of test of short- and long-distance physics including null tests of the Standard Model and its chirality flipped counterparts that can be carried out using experimental data.

Abstract

We explore the possibility of CP violation in baryonic Λb → ðΛþ c ; pþÞπþμ−μ− decays which are mediated by two Majorana sterile neutrinos and are jΔLj ¼ 2 lepton-number-violating processes. Appreciable CP asymmetry can be obtained if there are two on-shell Majorana neutrinos that are quasidegenerate in mass with the mass difference of the order of average decay widths. We find that, given the present constraints on the heavy to light mixing element jVμNj, the Λb → pþπþμ−μ− and Λb → Λþ c πþμ−μ− decay rates are suppressed but could be within the experimental reach at the LHC. If searches of the modes are performed, then experimental limits on the rates can be translated to constraints on the Majorana neutrino mass mN and heavy to light mixing element squared jVμNj 2. We show that the constraints on the ðmN; jVμNj 2Þ parameter space coming from the jΔLj ¼ 2 baryonic decays are complementary to the bounds coming from other processes.

Abstract

In the left-right symmetric models, a heavy charged gauge boson can decay to a lepton and a right-handed neutrino (RHN). If the neutrino masses are generated through the standard type-I seesaw mechanism, the Yukawa couplings controlling two-body decays of the RHN become very small. As a result, the RHN decays to another lepton and a pair of jets via an off-shell . This is the basis of the Keung-Senjanovi'{c} (KS) process, which was originally proposed as a probe of lepton number violation at the LHC. However, if a different mechanism like the inverse seesaw generates the neutrino masses, a TeV-scale RHN can have large Yukawa couplings and hence, dominantly decay to a lepton and a boson, leading to a kinematically different process from the KS one. We investigate the prospect of this unexplored process as a probe of the inverse seesaw mechanism in the left-right symmetric models at the HL-LHC. Our signal arises from the Drell-Yan production of a and leads to two high- same-flavour-opposite-sign leptons and a boosted -like fatjet in the final state. We find that a sequential with mass up to ~TeV along with a TeV-scale RHN can be discovered at the HL-LHC.

Abstract

The quest for stabilizing planar forms of tetracoordinate carbon started five decades ago and intends to achieve interconversion between [R]- and [S]-stereoisomers without breaking covalent bonds. Several strategies are successful in making the planar tetracoordinate form a minimum on its potential energy surface. However, the first examples of systems where stereomutation is possible were reported only recently. In this study, the possibility of neutral and dications of simple hydrocarbons (cyclopentane, cyclopentene, spiropentane, and spiropentadiene) and their counterparts with the central carbon atom replaced by elements from groups 13, 14, and 15 are explored using ab initio MP2 calculations. The energy difference between the tetrahedral and planar forms decreases from row II to row III or IV substituents. Additionally, aromaticity involving the delocalization of the lone pair on the central atom appears to help in further stabilizing the planar form compared to the tetrahedral form, especially for the row II substituents. We identified 11 systems where the tetrahedral state is a minimum on the potential energy surface, and the planar form is a transition state corresponding to stereomutation. Interestingly, the planar structures of three systems were found to be minimum, and the corresponding tetrahedral states were transition states. The energy profiles corresponding to such transitions involving both planar and tetrahedral states without the breaking of covalent bonds were examined. The systems showcased in this study and research in this direction are expected to realize molecules that experimentally exhibit stereomutation.

Abstract

The step-by-step workflow adopted for the analysis of water molecules in the binding sites of small molecules of our interest bound to protein crystal structures. (Note: We examined the local water network around O and S atoms of carbonyl (>C=O) and thiocarbonyl (>C=S) in small molecule ligands that are non-covalently bound to protein crystal structures (≤2.5Å) deposited in the Protein Data Bank (PDB). To minimize the influence of heterogeneous chemical micro-environment on the hydration profile of O and S atoms, we chose to confine our analyses to ligands that possess both carbonyl and thiocarbonyl moieties (in thioamides, thioureas, thiocarbamates, and dithiocarbamates). A subset of protein-ligand pairs where the ligand has at least one water molecule in its First Solvation Shell (FSS is defined as a sphere containing any point that lies within 4.5 Å of a ligand atom) was selected for further analysis.