Transfer Learning for Human & AI

The nanoPU is a new networking-optimized CPU designed to minimize taillatency for RPCs . By bypassing the cache and memory hierarchy, the nanoPU places arriving messages into the CPU register file . The wire-to-wirelatency through the application is just 65ns, about 13x faster than the current state-of-the-art .…

Lung ultrasound (LUS) has emerged as a rapid non-invasive imaging tool for the diagnosis of COVID-19 infected patients . Concerns surrounding LUS include the disparity of infected patients and healthcare providers, relatively small number of physicians and sonographers . A 2-dimensional (2D) tele-operative robotic platform capable of performing LUS may be of significant benefit .…

Trap-Aware Model Predictive Control (TAMPC) is a two-level hierarchical control algorithm that reasons about traps and non-nominal dynamics to decide between goal-seeking and recovery policies . We evaluate our method on simulated planar pushing andpeg-in-hole problems against adaptive controland reinforcement learning baselines .…

For every partial combinatory algebra (pca), we define a hierarchy ofextensionality relations using ordinals . We investigate the closure ordinals ofpca’s, i.e.\ the smallest ordinals where these relations become equal . We calculate the exactcomplexities of the extensionality relations in Kleene’s first model, showingthat they exhaust the hyperarithmetical hierarchy .…

Millimeter wave technology enables unmanned aerial vehicles to offer broadband high-speed wireless connectivity in fifth generation(5G) and beyond (6G) networks . However, the limited footprint of a single UAV requires multiple aerial stations tooperate in swarms to provide ubiquitous network coverage .…

Stencil kernels dominate a range of scientific applications including seismicand medical imaging, image processing, and neural networks . Temporal blocking is a performance optimisation that aims to reduce the required memory bandwidth of stencil computations by re-using data from the cache for multiple timesteps .…

A regular expression is called vulnerable if there exist input strings onwhich the usual backtracking-based matching algorithm runs super linear time . Software containing vulnerable regular expressions are prone to algorithmic-complexity denial of service attack in which the malicious user provides input strings exhibiting the bad behavior .…

We present analysis techniques for large trajectory data sets that aim toprovide a semantic understanding of trajectories reaching beyond them being being beingpoint sequences in time and space . The presented techniques use a drivingpreference model w.r.t. road segment traversal costs, e.g.,…

This technical report describes the EML submission to the first VoxCelebspeaker diarization challenge . EMLonline algorithm decides about the speaker labels in runtime approximately every 1.2 sec. For the first phase of the challenge, onlyVoxCeleb2 dev dataset was used for training .…

This paper presents a novel algorithm to plan energy-efficient trajectoriesfor autonomous ornithopters . We propose an efficient approach to plan ornithopter trajectoriesthat minimize energy consumption by combining gliding and flapping maneuvers . The algorithm builds a tree of dynamically feasible trajectories and appliesheuristic search for efficient online planning, using reference curves to guidethe search and prune states .…

Onboard Self Localization (OSL) methods based on cameras orLidar have achieved many significant progresses . However, some issues such asestimation drift and feature-dependence still remain inherent limitations . Infrastructure-based methods can generally overcome these issues, but at the expense of some installation cost .…

In quest for a logic capturing PTime the next natural classes ofstructures to consider are those with bounded color class size . Structures with color classes ofbounded size can be reduced canonization preservingly to normal form in Choiceless Polynomial Time (CPT) This is the first result of this form for non-abeliancolor classes .…

Supervised deep learning methods for performing audio source separation can be very effective in domains where there is a large amount of training data . While some music domains have enough data suitable for training a separationsystem, many musical domains do not, such as classical music, choral music, and non-Western music traditions .…

Attention-based text-to-speech synthesis (TTS) is superior to conventional statistical methods in many ways . Transformer-based TTS is one of such successful implementations . While Transformer TTS models the speech framesequence well with a self-attention mechanism, it does not associate input text with output utterances from a syntactic point of view at sentence level .…

When training deeplearning separation models, a need for ground truth leads to training on synthetic mixtures . Training in noisy conditions requires either usingnoise synthetically added to clean speech, or training using mixtures of noisy speech . We demonstrate therelative inseparability of noise and that this noisy speech paradigm leads to significant degradation of system performance .…

A2O VC aims to convert any speaker, including those unseen during training, to a fixed target speaker . We utilizevq-wav2vec (VQW2V), a discretized self-supervised speech representation that was learned from massive unlabeled data . Our model can be generalized to only 5 min of data, even outperforming the same model trained with parallel data .…

IDLab top-scoring submissions for VoxCeleb Speaker Recognition Challenge 2020 (VoxSRC-20) in the supervisedand unsupervised speaker verification tracks . For the supervised verificationtracks we trained 6 state-of-the-art ECAPA-TDNN systems and 4 Resnet34 basedsystems with architectural variations . On all models we apply a large marginfine-tuning strategy, which enables the training procedure to use higher marginpenalties by using longer training utterances .…

This paper leverages a multi-lingualAutomatic Speech Recognition (ASR) system to perform Speech Activity Detection(SAD) In out-of-domain datasets, the proposed SAD model shows significantly better performance w.r.t. baselinemodels. In the Ester2 dataset, without using any in-domain data, this modeloutperforms the WebRTC, phoneme recognizer based VAD (Phn\_Rec), and Pyannote baselines (respectively 7.1, 1.7, and 2.7% absolutely) in Detection Error Rate(DetER) metrics.…

Singing voice correction (SVC) is an appealing application for amateursingers . Commercial products automate SVC by snapping pitch contours to equal-tempered scales, which could lead to deadpan modifications . SVC is achieved by temporally aligning thesource-target pair, followed by replacing pitch and rhythm of the source with those of the target .…

Robust voice activity detection (VAD) is a challenging task in lowsignal-to-noise (SNR) environments . Recent studies show that speech enhancementis helpful to VAD, but the performance improvement is limited . Here we propose a new joint optimizationobjective—VAD-masked scale-invariant source to noise ratio (mSI-SDR) The model has two decoders, one for speech enhancement and the other for VAD .…

Functional Reactive Programming (FRP) allows GUIs to be designed in a declarative fashion . Most FRP languages are synchronous and continually check for new data . This means that an FRP-style GUIs will “wake up” on each program cycle . This isproblematic for applications like text editors and browsers, where oftennothing happens for extended periods of time, and we want the implementation tosleep until new data arrives .…

Researchers have begun to study the differing access to information of individuals within a network . This previous work suggests that informationaccess is itself a potential aspect of privilege based on network position . Instead of using standard groupingmethods for graph clustering, we design and explore a clustering that incorporates models of how information flows on a network.…

This paper tackles the problem of verifying the authenticity of speechrecordings from world leaders . Previous work on detecting deep fake ortampered audio focus on scrutinizing an audio recording in isolation . We propose a subsequence alignment method based on theNeedleman-Wunsch algorithm .…

We investigate the impact of noise and topology on opinion diversity insocial networks . We find that opinion diversity decreases as communities and clusters are broken down . We test our predictions against data describingempirical influence networks between major news outlets and find that incorporating our measure in linear models for the sentiment expressed by suchsources on a variety of topics yields a notable improvement in terms of explanatory power .…

The open source community, plays a vital role in coping with the pandemic which is inherently a collaborative process . The study sheds light on the promising direction of adopting open sourcetechnologies and resources to rapidly tackle the worldwide public healthemergency in practice .…

This paper proposes a workflow to identify genes responding to a specifictreatment in an organism, such as abiotic stresses, a main cause of extensive agricultural production losses worldwide . The proposed approach is both ageneralization and an extension of WGCNA .…

The experimental landscape in natural language processing for social media is too fragmented . Each year, new shared tasks and datasets are proposed, ranging from sentiment analysis to irony detection or emoji prediction . In this paper, we propose a new evaluation framework(TweetEval) consisting of seven heterogeneous Twitter-specific classification tasks .…

Platforms such as Stack Overflow are available for software practitioners to solicit solutions to their challenges and knowledge needs . Academic research tends to provide validation for the practices and processes employed by Stack Overflows and other forums . However, previous work did not review the scale of scientific attention given to this cause .…

Automatic speech recognition (ASR) for under-represented named-entity (UR-NE) is challenging due to insufficient instances and poor contextual coverage in the training data to learn reliable estimates andrepresentations . We propose approaches to enriching UR-NEs to improve speech recognition performance . We first enrich the representations in a pre-trained recurrent neural network LM (RNNLM) by borrowing theembedding representations of the rich-represented NEs .…

There is an emerging interest in the application of deep learning models to source code processing tasks . We propose a simple yet effective method based on identifier anonymization to handle out-of-vocabulary(OOV) identifiers . We show that the proposed OOVanonymization method significantly improves the performance of the Transformer in two code processing task: code completion and bug fixing .…

This report contains an analysis of the queries defined in the SIGMOD 2014Programming Contest . We first describe the data set, then present the queries, providing graphical illustrations for them and pointing out their caveats . We also demonstrate the influence of this contest by listing followup works which used these queries as inspiration to design better algorithms or to define interesting graph queries .…

With vast amounts of high-quality information at our fingertips, how is it possible that many people believe that the Earth is flat and vaccinationharmful? We quantify the implications of an opinionformation mechanism whereby an uninformed observer gradually forms opinions about a world composed of subjects interrelated by a signed network of mutual trust and distrust .…

The annual number of publications at scientific venues, for example,conferences and journals, is growing quickly . Even for researchers it becomes harder and harder to keep track of research topics and their progress . Many such methods result in uninterpretable, purely numericalrepresentations .…

Source code processing heavily relies on the methods widely used in naturallanguage processing (NLP) but involves specifics that need to be taken into account to achieve higher quality . In thiswork, we develop a recurrent architecture that processes code with all variablenames anonymized, i.…

Multi-channel speech separation sometimes doesnot necessarily need such a heavy structure for all time frames . To deal with this problem, we propose an early exit mechanism, which enables the Transformer model to handledifferent cases with adaptive depth . Early exit mechanism accelerates the inference, but also improves the accuracy of the separation task, says the authors .…

In this paper, we aim to improve the robustness of dynamic quadrupedallocomotion through three aspects: 1) fast model predictive foothold planning, 2) LQR control for robust motion tracking and 3) adaptive feet for terrainadaptation . The specially designed foot with adaptive sole aims at improving thetraversability of rough terrains with rocks, loose gravel and rubble byenlarging the contact surfaces with ground .…

During the steady gait, humans stabilize their head around the vertical orientation . While there are sensori-cognitive explanations for thisphenomenon, its mechanical e fect on the body dynamics remains un-explored . We introduce a simplifiedanthropometric D model to reproduce a broad walking dynamics .…

The method takes features for both edges and faces of a 3D mesh as input and dynamicallyaggregates them using an attention mechanism . At the same time, we introduce apooling operation with a precise geometric interpretation, that allows handlingvariations in the mesh connectivity by clustering mesh faces in a task-driven fashion .…

Dense synthesizer attention (DSA) dispenses with dotproducts and pairwise interactions . Local DSA restricts attention scope of DSA to a local range around the central frame . Combination method combines DSA with SA to extract the local and global information simultaneously .…

DBLog is currently used in production by tens of microservices at Netflix . It is a commonly observed pattern for applications to utilize multipleheterogeneous databases where each is used to serve a specific need such as storage the canonical form of data or providing advanced search capabilities .…

Spatial approximations have been traditionally used in spatial databases to accelerate the processing of complex geometric operations . However, many emerging applications (e.g., visualization tools) requireinteractive responses, while only needing approximate results . Besides, real-world geospatial data is inherently imprecise, which makes exact dataprocessing unnecessary .…

Transformer encoder-decoder model leverages multilingual data well in IPA transcriptions of languages presented during training . However, the representations it learned were not successful in zero-shot transfer to unseen languages . We show that the gain from modelingcrosslingual phonotactics is limited, and imposing a too strong model can hurt too strong models .…

We study the geometric separability problem whenthe separator is a rectangular annulus of fixed orientation . We give polynomial time algorithms to construct separators of each of the above type that also optimizes a givenparameter . We also give an optimal algorithm that computes a separating orthogonal convex polygon withminimum number of edges, that runs in $O(n\log n) time.…

PEAGNN trains multilayer GNNs to perform metapath-aware information aggregation on collaborative subgraphs . The model is compatible with the mainstream GNNstructures such as GCN, GAT and GraphSage . The empirical analysis on threepublic datasets demonstrate that our model outperforms or is at least on par with other competitive baselines .…

We study query containment in three closely related formalisms: monadicdisjunctive Datalog, MMSNP and OMQs . Containment inMMSNP was known to be decidable due to a result by Feder and Vardi, but itsexact complexity has remained open . We prove 2NEXPTIME-completeness and extend this result to monadic disjunctiveDatalog and to OMQ .…

Wepropose SapBERT is a pre-training scheme based on BERT . It self-aligns therepresentation space of biomedical entities with a metric learning objective function leveraging UMLS, a collection of biomedical ontologies with 4Mconcepts . SapberT outperforms many domain-specific BERT-basedvariants such as BioBERT, BlueBERT and PubMedBERT .…

We consider ontology-mediated queries (OMQs) based on expressive descriptionlogics of the ALC family and (unions) of conjunctive queries . Our results include exact characterizations of when such a rewriting is possible and tightcomplexity bounds for deciding rewritability . We also give a tight complexitybound for the related problem of deciding whether a given MMSNP sentence is equivalent to a CSP .…

Billions of text analysis requests containing private emails, personal textmessages, and sensitive online reviews, are processed by recurrent neuralnetworks (RNNs) deployed on public clouds every day . Prior securenetworks combine homomorphic encryption (HE) and garbled circuit (GC) topreserve users’ privacy .…

We present two methods to reduce the complexity of Bayesian network (BN)classifiers . We introduce quantization-aware training using thestraight-through gradient estimator to quantize the parameters of BNs to fewbits . We extend a recently proposed differentiable tree-augmented naiveBayes (TAN) structure learning approach by also considering the model size .…

Multi-agent active perception is a task where a team of agents cooperatively gather observations to compute a joint estimate of a hidden variable . The task is decentralized and the joint estimate can only be computed after the task ends by fusing observations of all agents .…