Transfer Learning for Human & AI

RETIoT ( Requirements Engineering Technology for the Internetof Things based software systems) aims to provide methodological, technical,and tooling support to produce IoT software system requirements document . It iscomposed of an IoT scenario description technique, a checklist to verify IoTscenarios, construction processes, and templates for IoT software systems .…

There exists many resource allocation problems in the field of wirelesscommunications which can be formulated as the generalized assignment problems . GAP is a generic form of linear sum assignment problem (LSAP) and is more challenging to solve owing to the presence of both equality and inequality constraints .…

The identification of black-box nonlinear state-space models requires aflexible representation of the state and output equation . Artificial neuralnetworks have proven to provide such a representation . A well-thoughtinitialization of these model parameters can often avoid that the nonlinearoptimization algorithm converges to a poorly performing local minimum of the considered cost function .…

Resource Constrained Training (RCT) only keeps a quantised modelthroughout the training, so that the memory requirements for model parameters are reduced . It adjusts per-layer bitwidth dynamically in order to save energy when a model can learn effectively with lower precision .…

Given a nonempty finite word $v$ let $PL(v)$ be the palindromic length of a word $x$ Let $x is an infinite non-ultimatelyperiodic word with $maxPL(x)=k<\infty . We show how to reduce the powers of $u$ and$u^R$ in $x' length in such a way that the length remains bounded . Less formally said, we show how we reduce the power of $U$ and $p(t) to a word with a finite or infinite word $y$ . We construct an infinitenon-ultinally periodic word such that $u …

Learning discriminative image representations plays a vital role in long-tailed image classification because it can ease the classifier learning inimbalanced cases . We explore effectivesupervised contrastive learning strategies and tailor them to learn better image representations from imbalanced data . We propose a novel hybrid network structure composed of a supervised contrastive loss to learn image representations and a cross-entropy loss .…

We are interested in unsupervised reconstruction of complex near-capillaryvasculature with thousands of bifurcations . Unsupervised methods can use many structural constraints, e.g.topology, geometry, physics . Common techniques use variants of MST on geodesictubular graphs minimizing symmetric pairwise costs, i.e. distances .…

Most state-of-the-art instance segmentation methods have to be trained on dense annotated images . We propose a proposal-free segmentation approach based on non-spatialembeddings . The method exploits the structure of the learned embedding space to extract individual instances in a differentiable way .…

We show that for a collection of $d$-dimensional quantumstates of cardinality $N$, the sample complexity is $O(\sqrt{N}d/\epsilon^2)$ The test is obtained by estimating the meansquared Hilbert-Schmidt distance between the states, thanks to a suitablegeneralization of the estimator of the distance between twounknown states by B\u{a}descu, O’Donnell, and Wright .…

High Energy Physics (HEP) experiments traditionally consume large amounts of CPU cycles for detectorsimulations and data analysis, but rarely use compute accelerators such asGPUs . As the LHC is upgraded to allow for higher luminosity, it may not provide enough computingpower to support simulation and data analyses needs .…

Few-Shot Learning (FSL) is a challenging task, i.e., how to recognize novel classes with few examples? Pre-training based methods effectively tackle theproblem by pre-training a feature extractor and then predicting novel classes via a nearest neighbor classifier . Nevertheless, due to the data scarcity, the mean-based prototypes are usually biased .…

For languages with insufficient resources to train speech recognitionsystems, query-by-example spoken term detection (QbE-STD) offers a way of accessing an untranscribed speech corpus . Yet retrieval performance can be poor when the queryand corpus are spoken by different speakers and produced in different recordingconditions .…

Recent QA with logical reasoning questions requires passage-level relations among sentences . We propose a discourse-aware graph network (DAGN) that reasons rely on discourse structure of texts . The model encodes discourseinformation as a graph with elementary discourse units (EDUs) and discourserelations, and learns the features via a graph network for downstream QA tasks .…

Training a multi-speaker Text-to-Speech model from scratch is expensive and adding new speakers to the dataset requires themodel to be re-trained . The naive solution of sequential fine-tuning of a model can cause the model to have poor performance on olderspeakers .…

The spatial visualization of higher-dimensional geometry has mostly remained a niche research topic for mathematicians and physicists . In this work, we attempt to revive the community’s interest in visualizing higher dimensional geometry by shifting the focus from the visualization of abstract shapes to the design of a broader hyper-universe concept .…

Tracking by detection, the dominant approach for online multi-objecttracking, alternates between localization and re-identification steps . In contrast, tracking in humans is highlighted by the notion of object permanence . We build on top of the recent CenterTrack architecture, which takes pairs of frames as input, and extend it to videos of arbitrary length .…

OnCOCO, a single FCOS-ResNetNet-50 detector equipped with Optimal TransportAssignment (OTA) can reach 40.7% mAP under 1X scheduler . Extensive experiments conducted on COCO and CrowdHuman further validate the effectiveness of our proposed OTA, especially in crowd scenarios . The code is available athttps://://://github.com/Megvii-BaseDetection/OTA…

Image2Reverb is the first work that generates an IR from a single image . This IR is then applied to other signals usingconvolution, simulating the reverberant characteristics of the space shown inthe image . We use an end-to-end neural networkarchitecture to generate plausible audio impulse responses from single images .…

The article considers the quantitative assessment approach to theinnovativeness of different objects . The proposed assessment model is based on the object data retrieval from various databases including the Internet . Wepresent an object linguistic model, the processing technique for themeasurement results including the results retrieved from the different searchengines, and the evaluating technique of the source credibility .…

A fundamental problem in computer animation is that of realizing purposefuland realistic human movement given a sufficiently-rich set of motion capture clips . We learn data-driven generative models of human movement using autoregressive conditional variational autoencoders, or Motion VAEs . The latentvariables of the learned autoencoder define the action space for the movement and govern its evolution over time .…

The information inside these two visual domains is well complementary, e.g., 2Dimages have fine-grained texture while 3D point clouds contain plentifulgeometry information . However, most current visual recognition systems processthem individually . In this paper, we present a \emph{bidirectional projectionnetwork (BPNet) for joint 2D and 3D reasoning in an end-to-end manner .…

Monotonic Multihead Attention (MMA) shows comparable performance to offline methods in machine translation and automatic speech recognition (ASR) tasks . MMA is still a major issue in ASR and should be combined with atechnique that can reduce the test latency at inference time, such ashead-synchronous beam search decoding, which forces all non-activated heads toactivate after a small fixed delay from the first head activation .…

We introduce functorial language models: a principled way to computeprobability distributions over word sequences given a monoidal functor fromgrammar to meaning . This yields a method for training categorical compositional compositionaldistributional models on raw text data .…

We propose a vision-based architecture search algorithm for robotmanipulation learning . Our approach automaticallydesigns architectures while training on the task – discovering novel ways of combining and attending image feature representations with actions . The obtained new architectures demonstrate better task success rates, in some cases with a large margin, compared to arecent high performing baseline .…

MS-SVConv is a fast multi-scale deep neural network that outputs features from point clouds for 3D registration between two scenes . We computefeatures using a 3D sparse voxel convolutional network on a point cloud at different scales . We show significant improvements compared to state-of-the-art methods on the competitive and well-known 3DMatch benchmark .…

A permutation-based successive cancellation (PSC) decoding framework for polar codes attaches much attention . PSC framework isineffective for permutations falling into the lower-triangular affine (LTA)automorphism group . But BLTA equals the complete automorphisms of decreasing polar codes that can be formulated as affine trasformations .…

A single natural image is sufficient to train a feature extractor that outperforms state-of-the-art loss functions in single image super resolution, denoising, and JPEG artefact removal . Wepropose a novel Multi-Scale Discriminative Feature (MDF) loss comprising aseries of discriminators, trained to penalize errors introduced by a generator .…

Domain Adaptation with Virtual Domain Modeling (VDM-DA) achieves the state-of-the-art performances on all datasets . The virtual domain acts as a bridge between the source and target domains . In this way, we successfully achieve the goal of distributionalignment by training deep networks without accessing to the source domain data .…

Deep neural networks are vulnerable to small input perturbations known as adversarial attacks . We propose the anti-adversary layer aimed at countering this effect . In particular, our layer generates a perturbation in the opposite direction of the adversarial one, and feeds the classifier a perturbedversion of the input .…

DeFeat improves ResNet50 based Faster R-CNN from 37.4% to 40.9% mAP, and improves RetinaNet from 36.5% to 39.7% on COCO benchmark . It is a novel distillation algorithm via decoupled features(DeFeat) for learning a better student detector . The proposal is available at https://://://github.com/ggjy/deFeat.pytorch.…

The paper proposes novel sampling strategies to compute the optimal pathalteration of a surface vessel sailing in close quarters . Such strategy encodes the rules for safe navigation at sea, by exploiting the concept of minimal ship domain to determine the compliant region where the pathdeviation is to be generated .…

The paper studies planar surface reconstruction of indoor scenes from twoviews with unknown camera poses . Previous approaches have successfully created object-centric reconstructions of many scenes . They fail to exploit planes, such as planes, which are typically the dominant componentsof indoor scenes .…

At PODC 2014, A. Most\’efaoui, H. Moumen, and M. Raynal presented a new andsimple randomized signature-free binary consensus algorithm (denoted here MMR) that copes with the net effect of asynchrony Byzantine behaviors . MMR is optimal in several respects: it deals with up to t Byzantine processes where t < n/3and n is the number of processes, O(n\^2) messages and O(1) expected time . The proposed algorithm is the first loosely-self-stabilizing Byzantinefault-tolerant binary consensus algorithms suited to asynchronousmessage-passing systems . Furthermore, it only requires a bounded amount of memory. Furthermore, the obtained algorithm preserves its properties of optimal resilience andtermination, (i.e., t

This paper reports on developing an integrated framework for safety-awareinformative motion planning suitable for legged robots . The information-gathering planner takes a dense stochastic map of the environment into account, while safety constraints are enforced via Control BarrierFunctions (CBFs) The planner is based on the Incrementally-exploringInformation Gathering (IIG) algorithm and allows closed-loop kinodynamic nodeexpansion using a Model Predictive Control (MPC) formalism .…

2D US has to perform scanning for each SP, which is time-consuming and operator-dependent . Automatically locating SP in 3D US is very challenging due to the huge search space and large fetal posture variations . Our approach achieves localization error of 2.52mm/10.26 degrees, 2.02mm/11.48 degrees, 3.61mm/9.71 degrees, 1.49mm/7.54 degrees for thetranscerebellar, transventricular, transthalamic planes in fetal brain,abdominal plane in fetal abdomen, and mid-sagittal, transverse and coronalplanes in uterus, respectively .…

Modern video question answering (VideoQA) systems often use additional human-made sources like plot synopses, scripts, descriptions or knowledge bases . In this work, we present a new approach to understand the whole story without external sources . We treat dialog as a noisy source to beconverted into text description via dialog summarization .…

Human-designed data augmentation strategies have been replaced byautomatically learned augmentation policy in the past two years . But two factors regarding the diversity of augmented data are still missing: the explicit definition of diversity and the quantifiable relationship between diversity and its regularization effects .…

This paper aims at answering the question of finding the right strategy to make the target model robust and accurate in the setting of unsupervised domain adaptation without source data . The proposed method of using non-robust pseudo-labels performs surprisingly well on both clean andadversarial samples, for the task of image classification .…

Key information extraction from document images of unseen templates is of paramount importance in office automation . We model document images asdual-modality graphs, nodes of which encode both the visual and textualfeatures of detected text regions, and edges of which represent the spatialrelations between neighboring text regions .…

Methodology for handling edge attributes has been developed for a number of network models, but comparable techniques for edge-related attributes remain largely unavailable . We address this gap in the literature by extending the latent position random graph model to the line graph of a random graph .…

HUGE is a system called HUGE to efficiently process subgraphenumeration at scale in the distributed context . HUGE features a novel two-stage execution mode with a lock-free and zero-copy cache design . Huge isgeneric such that all existing distributed subgraph enumeration algorithms can be plugged in to enjoy automatic speed up and bounded-memory execution .…

NL-EDIT can boost the accuracy of existing text-to-SQLparsers by up to 20% with only one turn of correction . The model is a model for interpreting natural language feedback in the interaction context to generate a sequence of edits that can be applied to the initial parse to correct its errors .…

Multi-label text classification (MLTC) is an attractive and challenging task . We propose a heterogeneous graph convolutional network model to solve the MLTC problem . We are able to take into account multiple relationships including token-level relationships . We evaluate our method on three real-world datasets and the experimental results show that it achievessignificant improvements and outperforms state-of-the-art comparison methods .…

This paper introduces Parallel Tacotron 2, a non-autoregressive neuraltext-to-speech model with a fully differentiable duration model . The duration model is based on a novelattention mechanism and an iterative reconstruction loss based on Soft DynamicTime Warping . This model can learn token-frame alignments as well as tokendurations automatically .…

Separating raypaths in a multipath shallow-water environment is a challenge . The proposed algorithm achieves a higher resolution and a stronger robustnesscomparing to the existing algorithms . Performance tests usingsimulation data in multipath environment, real data obtained in an ultrasonicwaveguide and ocean shallow water data, respectively, illustrate that the proposed algorithms achieve a higherresolution and a strong robustness .…

The generation of synthetic data is useful in multiple aspects, from testingapplications to benchmarking to privacy preservation . Generating links between relations subject to cardinality constraints (CCs) and integrityconstraints (ICs) is an important aspect of this problem . We provide a novel framework for the problem based on declarative CCs and ICs.…

We present a reinforcement learning algorithm for learning sparsenon-parametric controllers in a Reproducing Kernel Hilbert Space . We improve the sample complexity of this approach by imposing a structure of the state-action function through a normalized advantage function (NAF) Thisrepresentation of the policy enables efficiently composing multiple learnedmodels without additional training samples or interaction with the environment .…

We present a contrasting learning approach with data augmentation techniquesto learn document representations in an unsupervised manner . We hypothesize that high-quality document embedding should beinvariant to diverse paraphrases that preserve the semantics of the original document . Our method can decrease the classification error rate by upto 6.4% over the SOTA approaches on the document classification task .…

In this work, we present a database of multimodal communication featuresextracted from debate speeches in the 2019 North American Universities DebateChampionships (NAUDC) Feature sets were extracted from the visual (facialexpression, gaze, and head pose), audio (PRAAT), and textual (word sentimentand linguistic category) modalities of raw video recordings of competitive debaters .…