Transfer Learning for Human & AI

In the past twelve months, our team has had to move rapidly from conducting most of our user experiments in a laboratory setting . This has challenged us to think about what types of experimentwe can run, and to improve our tools and methods to allow us to reliably capture the necessary data .…

Autopilot behavior can help to ensuresmoothhuman-vehicle collaboration during the initial exploration stagewith thevehicle . AutoPreview framework can provide a deeperunderstanding of autopilot behavior compared to direct interaction with the vehicle . We conducted acase study on humans-vehicles collaboration and built a prototype of our framework with theCARLA simulator .…

Training on datasets with long-tailed distributions has been challenging formajor recognition tasks such as classification and detection . We show that image-level andobject-level resamplings are both important, and thus unify them with a jointresampling strategy (RIO) Our method outperforms state-of-the-art long-taileddetection and segmentation methods on LVIS v0.5 across various backbones.…

This paper presents a Matlab toolbox to perform basic image processing and visualization tasks, particularly designed for medical image processing . The toolbox is entirely written in native Matlab code, but is fast and flexible . Main use cases for the toolbox are illustrated here, including image input/output, pre-processing, filtering, image registration and visualisation .…

Wepropose a 3D sphere representation-based center-points matching detection network (SCPM-Net) that is anchor-free and automatically predicts the position, radius, and offset of nodules without the manual design of nodule/anchorparameters . SCPM-net consists of two novel pillars: sphere representationand center points matching .…

Pseudo-labeling is a key component in semi-supervised learning . SemCo leverages label semantics and co-training to address this problem . We train two classifiers with twodifferent views of the class labels . We show that our method achieves state-of-the-art performance across various SSL tasks including 5.6% accuracy improvement on Mini-ImageNetdataset with 1000 labeled examples .…

We propose to use speech cues (on what, when and how the user talks) as an objective real-time measure of trust . This could be implemented in the robot to calibrate towards appropriate trust . However, we would like to open the discussion on how to deal with the ethical implications of this trust measure .…

The proposed BPR framework yields significant improvements over theMask R-CNN baseline on Cityscapes benchmark, especially on the boundary-awaremetrics . We reached 1st place on the CityScapes leaderboard by applying the BPR to the PolyTransform + SegFix baseline, we reached 1th place in the leaderboard .…

There are many potential measures that one might use when evaluating mixed reality experiences . In this position paper I will argue that there arevarious stances to take for evaluation, depending on the framing of the work . I will sketch out some directions for developing more robust measures that can help the field move forward.…

We present a comparison of several online machine learning techniques fortactical learning and proving in the Coq proof assistant . This work builds ontop of Tactician, a plugin for Coq that learns from proofs written by the userto synthesize new proofs .…

Text style transfer aims to controllably generate text with targetedstylistic changes while maintaining core meaning from the source sentenceconstant . Many existing style transfer benchmarks do not offer fine-grained control on sentence structure, emphasis, and content of the sentence . We introduce a large-scale benchmark, StylePTB, with pairedsentences undergoing 21 stylistic changes spanning atomic lexical,syntactic, semantic, and thematic transfers of text .…

Flash-based disk caches, for example Bcache and Flashcache, has gainedtremendous popularity in industry in the last decade . But these cache systemshave a worse write performance than the read performance because of theasymmetric I/O costs and the internal GC mechanism .…

In this work we address the task of observing the performance of a semanticsegmentation deep neural network (DNN) during online operation . Many high-level decisions rely on such DNNs, which are usually evaluated offline, while their performance in online operation remains unknown .…

The problem of lifting a preference order on a set of objects to a preferenceorder on a family of subsets of this set is a fundamental problem with a widevariety of applications in AI . The process is often guided by axiomspostulating properties the lifted order should have .…

Deep learning methods have reached state-of-the-art performance in cardiac image segmentation . Currently, the main bottleneck towards their effective translation into clinics requires assuring continuous high model performance . In this work, we present a novel learning framework to monitor the performance of heart segmentation models in the absence of ground truth .…

The open-domain sketch-to-photo translation is challenging due to the lack of training supervision and the large geometrydistortion between the freehand sketch and photo domains . We propose a framework that jointlylearns sketch to photo and photo-to sketch generation . We validate our method on the Scribble and SketchyCOCO datasets.…

We study the computational complexity of Feedback Vertex Set on subclasses ofHamiltonian graphs . In particular, we consider Hamiltonian graphs that areregular or are planar and regular . We also study the less known class of $p$-Hamiltonian-ordered graphs, which admit for any $ p$-tupleof vertices a Hamiltonian cycle visiting them in the order given by the tuples of vertices .…

The goal of this research is to create a tree model to automaticallydetermine possible pruning points for stand-alone trees within meadows . The algorithm is capable of building a skeleton model based on a pre-segmented photogrammetric 3D point cloud . Good results were achieved in assigning the points to their leading branches and building a virtual treemodel, reaching an overall accuracy of 95.19% .…

UnifiedM2 is a general-purpose misinformation model that jointly models multiple domains of misinformation with a single, unified setup . The model is trained to handle four tasks: detecting news bias, clickbait, fake news, and verifying rumors . By grouping these tasks together, the model learns a richer representation of misinformation, which leads to comparable performance across all tasks .…

GarmentNets tackles the task of category-level pose estimation for garments . The key idea is to formulate the deformable object poseestimation problem as a shape completion task in the canonical space . The output representation describes the garment’s full configuration using a complete 3D mesh with the per-vertex canonical coordinate label .…

Stock price prediction is the focus of economics . Stock price forecasts can provide reference data . The Dempster combination rule is a classic method of fusing different information . This article will introduce the prediction method of evidence theory .…

The canonical polyadic (CP) decomposition of tensors thatcorresponds to matrix multiplications is studied . Finding the rank of thesetensors and computing the decompositions is a fundamental problem of algebraiccomplexity theory . In this paper, we present a novel decomposition . of the tensormultiplication of matrices of the size 3×3 with 3×6 with rank 40 .…

StereoPIFu can infer the human body’s spatiallocation in camera space and maintain the correct relative position of different parts of the body . Compared with previous works, our method significantlyimproves the robustness, completeness, and accuracy of the clothed human reconstruction, which is demonstrated by extensive experimental results .…

We propose a novel approach for the adaptivechoice of the learning rate in first-order gradient descent methods . Our approach modifies the objective function by adding a regularization term on thelearning rate . We theoreticallyprove that some well-designed regularizers can improve the convergenceperformance under the strong-convexity condition of the objectivefunction .…

The functorial structure of type constructors is the foundation for manydefinition and proof principles in higher-order logic (HOL) In this article, we tackle the preservation question forquotients, the last principle for introducing new types in HOL . Surprisingly, lifting the structure in theobvious manner fails for some quotients .…

In our framework, the features extracted by deep networks of each class arecharacterized by independent Gaussian distribution . The means of the Gaussian play a similar role as the centeranchor in existing methods . In addition, we introduce a margin to intra-class uncertainty to make each cluster more compact and reduce the imbalance of feature distribution from different categories .…

A multitude of classifiers can be trained on the same data to achieve similar performances during test time . This phenomenon is often associated with the blind selection of one model instead of another with similar performances . A model-agnostic algorithm, DIG, is proposed to capture and explaindiscrepancies locally .…

We extend the .neural basis expansion analysis (NBEATS) to incorporateexogenous factors . The resulting method, called NBEATSx, improves on a wellperforming deep learning model . We conduct a .comprehensive study of its application to electricity price forecasting (EPF)tasks across a broad range of years and markets .…

Pedestriancrossing intention should be recognized in real-time, especially for urbandriving . Recent works have shown the potential of using vision-based deepneural network models for this task . We fuse different phenomena such as sequences of RGBimagery, semantic segmentation masks, and ego-vehicle speed in an optimum way .…

In the Equal Maximum Flow Problem (EMFP), we aim for a maximum flow where werequire the same flow value on all edges in some given subsets of the edge set . We prove that the integer almost equal maximum flow problem (integer AEMFP) is in general$\mathcal{NP}$-complete .…

Auto-decoded deep latent embeddings(ADDLE) models tendencies of each rater using anauto-decoder framework . ADDLE can improve the partial areas under the curve (AUCs) fordiagnosing severe steatosis by 10.5% over standard classifiers . It can also improve the performance of other annotator-noise approaches, including those requiring 65 times the parameters.…

Beam search is the go-to method for decoding auto-regressive machinetranslation models . While it yields consistent improvements in terms of BLEU, it is only concerned with finding outputs with high model likelihood . Our aim is to establish whether beam search can be replaced by a more powerfulmetric-driven search technique .…

TermAdventure (TA) is a suite of tools for creating interactive UNIX exercises . These resemble text adventure games, which immersethe user in a text environment and let them interact with it using textualcommands . The suite is released under an open source license, has minimal dependencies and can be used either on a UNIX-style server or a desktop computer running any major OS platform through Docker .…

The optimization of the Relay Transform Matrix (RTM) in a two-hop relaynetwork with an average relay power constraint and perfect channel stateinformation at the relay is addressed in this paper . The study considers themost general case in terms of number of transmit and receive antennas at thesource, relay, and destination, with arbitrary correlation of the receivednoise at all terminals .…

We outline the Great Misalignment Problem in natural language processing research . Only one paper was fully in line in terms of problem definition, method and evaluation . Only two papers presented a human evaluation that was in line with what was modeled in the method .…

GenerativeAdversarial Networks (GANs) involve training of overparameterized models where the number of parameters of the model exceeds a certain threshold . The role and benefits of modeloverparameterization in the convergence of GDA to a global saddle point in non-convex concave min-max problems is far less understood .…

We tackle the problem of action-conditioned generation of realistic and diverse human motion sequences . In contrast to methods that complete, orextend, motion sequences, this task does not require an initial pose orsequence . We learn an action-aware latent representation for human motionsby training a generative variational autoencoder (VAE) We evaluate our approach on NTU RGB+D, HumanAct12 and UESTC datasets and show improvements over the state of the art .…

This paper presents a biologically plausible method for converting real-valued input into spike trains for processing with spiking neuralnetworks . The proposed method mimics the adaptive behaviour of retinal ganglioncells and allows input neurons to adapt their response to changes in thestatistics of the input .…

In this work, we analyze the performance of a simple majority-rule protocolsolving a fundamental coordination problem in distributed systems . We prove that the Simple Majority Protocol (SMP) reaches consensus in only three communication rounds with probability approaching $1$ as $n$ grows to infinity .…

In this paper, an intelligent reflecting surface (IRS) assisted spectrumsharing underlay cognitive radio (CR) wiretap channel (WTC) is studied . We aim at enhancing the secrecy rate of secondary user in this channel subject tototal power constraint at secondary transmitter (ST), interference powerconstraint (IPC) at primary receiver (PR) and unit modulus constraint atIRS .…

The biggest challenge is the overhead incurred when the mobile terminal has to send the downlink channel state information or corresponding partial information to the base station . We propose a novel learning-based framework for denoising and compression of channel estimates .…

Mathematical information retrieval (MathIR) applications such as semanticformula search and question answering systems rely on knowledge-bases that linkmathematicical expressions to their natural language names . For databasepopulation, mathematical formulae need to be annotated and linked to semanticconcepts, which is very time-consuming .…

Deep Neural Networks (DNNs) have recorded great success in handling medical and other complex classification tasks . As the sizes of a DNN model and the available dataset increase, the training process becomes more computationally intensive . We have proposed a generic full end-to-end hybridparallelization approach combining both model and data parallelism forefficiently distributed and scalable training of DNN models .…

PPT-Multicore builds upon our previous work towards amulticore cache model . We extract LLVM basic block labeled memory trace using an architecture-independent LLVM-based instrumentation tool only once in anapplication’s lifetime . The model uses the memory trace and other parameters from an instrumented sequentially executed binary .…

Shuffler is an open source tool that makes it easy to manage large computervision datasets . It stores annotations in a relational, human-readabledatabase . It defines over 40 data handling operations with annotationsthat are commonly useful in supervised learning applied to computer vision .…

A novel deeplearning based cost model for automatic code optimization has been proposed in the Tiramisu compiler . The proposed model has only 16% of mean absolute percentage error in predicting speedups onfull programs . Unlike previous models, the proposed one does not rely on any heavy feature engineering .…

The Voynich MS is an illustrated 15th century manuscript, whose text is written in an unknown alphabet . In 2004 Gordon Rugg published a paper in which he proposed that this text islikely to be meaningless, and could have been composed by an alternative application of a so-called Cardan Grille .…

Cyber-physical systems and the Internet of things (IoT) are becoming anintegral part of the digital society . Malicious activities lead to confidential dataleakages and incorrect performance of devices are becoming critical . This study provides acritical analysis of the feasibility of using blockchain technology to protect IoT devices data .…