Transfer Learning for Human & AI

Nonlinear state-space identification for dynamical systems is most often performed by minimizing the simulation error to reduce the effect of modelerrors . This optimization problem becomes computationally expensive for largedatasets . This paper introduces a method that approximates the simulation loss by splitting the data set into multipleindependent sections similar to the multiple shooting method .…

We present a scalable block preconditioning strategy for the trace system coming from the high-order hybridized discontinuous Galerkin (HDG)discretization of incompressible resistive magnetohydrodynamics (MHD) We use a least squares commutator (BFBT)approximation for the inverse of the Schur complement . The remaining velocity, magnetic field,and Lagrange multiplier unknowns form a coupled nodal unknown block (the upper block), for which a system algebraic multigrid (AMG) is used for the approximate inverse .…

We present a new algorithm to automatically generate high-performance GPUimplementations of complex imaging and machine learning pipelines . It is fully automatic, requiring noschedule templates or hand-optimized kernels, and it targets a diverse range ofcomputations which is significantly broader than existing autoschedulers .…

Existing work on data-driven optimization focuses on problems in static environments, but little attention has been paid to problems in dynamic environments . This paper proposes a data streamensemble learning method to deal with the challenges presented by the dynamic environments.…

IBM’s Qiskit transpiler offers novel quantum compiler optimization for quantum computers . RPO leverages the single-qubit state information that can be determined statically by thecompiler . The circuits optimized with our optimization obtain up to 18.0% (11.7% on average) fewer CNOT gates and up to 8.2% (7.1%) lower transpilation time than that of the mostaggressive optimization level in the most aggressive optimization level .…

Pre-trained weights often boost a wide range of downstream tasks including classification, detection, and segmentation . Latest studies suggest that thepre-training benefits from gigantic model capacity . We examine the supervised and self-supervised pre-trained models through the lens of lottery ticket hypothesis (LTH) LTH identifies highly sparse matching subnetworks that can be trained in isolation from scratch, to reach the full models’ performance .…

Designing learning algorithms that are resistant to perturbations of the underlying data distribution is a problem of wide practical and theoreticalimportance . We present a general approach to this problem focusing onunsupervised learning . The key assumption is that the perturbing distribution is characterized by larger losses relative to a given class of admissiblemodels .…

A convex partition of a point set set P in the plane is a planar partition of theconvex hull of P with empty convex polygons or internal faces whose extremepoints belong to P . The problem has been shown to be NP-Hard and was recently used in the CG:SHOP Challenge 2020 .…

In this paper, we derive generalization bounds for the two primary classes ofgraph neural networks (GNNs) and message passing GNNs . We also show that our bound for GCNs is anatural generalization of the results developed in arXiv:1707.09564v2 [cs.LG] for fully-connected and convolutional neural networks .…

A coordination process model specifies coordination constraints betweenthe interdependent processes in terms of semantic relationships . At run-time, these coordination constraints must be enforced by a coordination process instance . Processes must be allowed to run asynchronously andconcurrently, taking their complex relations into account .…

The gold standard of a wireless network is that the throughput increaseslinearly with the density of access points (APs) In this paper, we propose to overcome the densification plateau of a mmWave network by employing spatial multiplexing at APs . We then demonstrate the necessity for deploying the multi-rate coding scheme in mmWave networks, especially when thespatialMultiplexing gain at mmWave APs is large, we say .…

In large-scale problems, standard reinforcement learning algorithms suffer from slow learning speed . In this paper, we follow the framework of usingsubspaces to tackle this problem . We propose a free-energy minimization framework for selecting the subspaces and integrate the policy of the state-space into the subspace .…

StyleGAN is able to produce photorealistic images almost indistinguishable from real ones . Embedding images into the StyleGAN latent space is not atrivial task due to the reconstruction quality and editing quality trade-off . In this paper, we introduce a new normalized space to analyze the diversity and the quality of the reconstructed latent codes .…

The Reseau de Transport d’Electricit\’e (RTE) is the French main electricitynetwork operational manager and dedicates large number of resources and effortstowards understanding climate time series data . The data is composed of temperature times series for 200 different possiblescenarios on a grid of geographical locations in France .…

Solving complex optimization problems in engineering and the physical sciences requires repetitive computation of multi-dimensional functionderivatives . In nonlinear inverse problems Gauss-Newton methods are used that require iterative updates to becomputed from the Jacobian . Computationally more efficient alternatives are Quasi-Newtons methods, where the repeated computation of theJacobian is replaced by an approximate update .…

The Sparse Multi-Family Deep Scattering Network(SMF-DSN) is a novel architecture exploiting the interpretability of the DeepScattering Network (DSN). The DSN extractssalient and interpretable features in signals by cascading wavelet transforms, complex modulus and extract the representation of the data via atranslation-invariant operator .…

Public resources and services have been widely adopted to ease the development of DeepLearning-based applications . However, if the third-party providers areuntrusted, they can inject poisoned samples into the datasets or embedbackdoors in those models . We propose a systematic approach to discover the optimal policies for defending against differentbackdoor attacks by comprehensively evaluating 71 state-of-the-art dataaugmentation functions .…

Syntactic structure of a sentence text is correlated with prosodicstructure of the speech that is crucial for improving the prosody and naturalness of a text-to-speech (TTS) system . In this paper, we propose a syntactic representationlearning method based on syntactic parse tree traversal to automaticallyutilize the syntactic structure information .…

Organizations collect vast quantities of sensitive information in ERP systems . Leakage of such information poses a severe threat for companies as the number of incidents and the reputational damage to those experiencing them continue to increase . Understanding the nature of such attacks becomes increasinglyimportant for the (internal) audit and fraud examination practice .…

We give a new approach to the elliptic curve discrete logarithm problem . It is based on a transfer: First an$\mathbb{F}_q$-rational . isogeny from the Weil restriction of the curve under consideration with respect to . the Jacobian variety of a genus three curve .…

Meticulous ObjectSegmentation (MOS) is focused on segmenting well-defined foreground objects with elaborate shapes in high resolution images (e.g. 2k – 4k) Compared with common image segmentation tasks targeted at low-resolution images, higher resolution detailed image segmentations receives much less attention .…

Relational stores are able to store large-scale knowledge graphs and convenient in updating knowledge . Native graph stores are efficient in processing complex knowledge graph queries due to its index-freeadjacent property . But they are inapplicable to manage a large-sized knowledgegraph due to limited storage budgets or inflexible updating process .…

Computational notebooks have emerged as the platform of choice for datascience and analytical workflows, enabling rapid iteration and exploration . NBSafety detects and prevents errors that users make during unaided notebook interactions, all while preserving the flexibility of existing notebook semantics .…

There have been growing uses of semantic networks in the past decade, such asleveraging large-scale pre-trained graph knowledge databases for various natural language processing (NLP) tasks in engineering design research . The survey reveals that engineering design researchers have primarily relied on WordNet,ConceptNet, and other common-sense semantic network databases trained on non-engineering data sources .…

We present a novel framework for boosting neural semantic parsers via iterative utterancesegmentation . Given an input utterance, our framework iterates between twoneural modules: a segmenter for segmenting a span from the utterance and aparser for mapping the span into a partial meaning representation .…

We present the barycentric Lagrange Lagrange dual tree traversal (BLDTT) fastsummation method for particle interactions . The scheme replaces well-separatedparticle-particle interactions by adaptively chosen particle-cluster, cluster-clusters approximations . The performance of the BLDTT is demonstrated for calculations with different problem sizes, particledistributions, geometric domains, and interaction kernels, as well as forunequal target and source particles .…

We study fair allocation of undesirable indivisible items (or chores) We show that determining the existence of anenvy-free allocation is NP-complete even when agents have binary additivevaluations . We provide a polynomial-time algorithm for computing anallocation that satisfies envy-freeness up to one chore (EF1) under monotonevaluations, correcting a existing proof of the same claim in the literature .…

Primal-dual technique, which has been particularly successful indealing with outliers and penalties, has not been very successful in dealing with capacities . In this paper, we present a framework to design approximation algorithms forcapacitated facility location problems with penalties/outliers using LP-rounding .…

Random devices that exploit ambient thermal noise have recently attracted attention as hardware primitives for accelerating probabilistic computing applications . Such devices can be used as hardware accelerators in energy-efficient computing schemes that require alarge throughput of tunably random bits .…

Terahertz (THz) communications have been envisioned as a promising enabler toprovide ultra-high data transmission for sixth generation (6G) wirelessnetworks . An intelligent reflectingsurface (IRS) is put forward to smartly control the incident THz waves by adjusting the phase shifts . An adaptive gradient descent (A-GD) algorithm is developed by dynamically updating the step size during theiterative process, which is determined by the second-order Taylor expansionformulation .…

Homogeneous, autonomous mobile robots place themselves equidistant apart apart on the grid . The robots operate by executing cycles of the states”look-compute-move” The robots aresemi-synchronous, anonymous and have unlimited visibility . Eventually, the robots uniformly distribute themselves on alternate nodes of a grid, leaving adjacent nodes of the grid vacant.…

Noninvasive brain-computer interface (BCI) decodes brain signals tounderstand user intention . decoding of sensation imagery based on brain signals could provide various industries such as developing advanced displays and more immersive virtual reality and augmented reality . Thispaper introduces a preliminary study to develop a neurohaptic-based BCI system using actual touch and touch imagery paradigms.…

Cold-start problem is a fundamental challenge for recommendation tasks . Pre-training GNN simulates the cold-startscenarios from the users/items with sufficient interactions . Self-attention-based meta aggregator to enhance the aggregationability of each graph convolution step, and an adaptive neighbor sampler to select the effective neighbors according to the feedbacks from the pre-trainingGNN model .…

In this paper, we study the circle formation problem by multiple autonomousand homogeneous disc-shaped robots (also known as fat robots) The goal of the robots is to place themselves on the periphery of a circle . Circle formationhas many real-world applications, such as boundary surveillance .…

We characterize the lattice decoding problem from a neuralnetwork perspective . The notion of Voronoi-reduced basis is introduced to restrict the space of solutions to a binary set . This problem is shown to be equivalent to computing a continuous piecewise linear (CPWL)function restricted to the fundamental parallelotope .…

Collaborative filtering tries to predict the ratings of a user over someitems based on opinions of other users with similar taste . The ratings areusually given in the form of a sparse matrix, the goal being to find themissing entries (i.e.…

We propose novel randomized optimization methods for high-dimensional convexproblems . We consideroblivious and data-adaptive subspaces and study their approximation propertiesvia convex duality and Fenchel conjugates . A suitable adaptive subspace can begenerated by sampling a correlated random matrix whose second order statistics mirror the input data .…

In this paper, we propose a theoretically founded sequential strategy for training large-scale Recommender Systems (RS) over implicit feedback, mainly inthe form of clicks . The proposed approach consists in minimizing pairwiseranking loss over blocks of consecutive items constituted by a sequence ofnon-clicked items followed by a clicked one for each user .…

In many practical few-shot learning problems, there are abundant auxiliary data sets that potentially contain useful information . We propose amasking module that adjusts the features of auxiliary data to be more similarto those of the target classes . We show that this masking module can improve accuracy by up to 18 accuracy points, particularly when the auxiliary data issemantically distant from the target task .…

Designing efficient and scalable sparse linear algebra kernels on modern multi-GPU based HPC systems is a daunting task due to significant irregular memory references and workload imbalance across the GPUs . This is particularly the case for Sparse Triangular Solver which introduces additional two-dimensional computation dependencies among subsequent computation steps .…

Let L be a Desarguesian 2-spread in the Grassmann graph $J_q(n,2)$ We prove that the collection of the 4-subspaces, which do not contain subspaces from Lis, are a completely regular code . Similarly, we construct acompletely regular code in the Johnson graph from the Steinerquadruple system of the extended Hamming code .…

A large number of tasks for facial image processing in thermal infrared images that are currently solved using specialized rule-based methods or notsolved at all can be addressed with modern learning-based approaches . We have used USTC-NVIE database for training of a number of machine learning algorithms for facial landmark localization .…

We study stochastic games with energy-parity objectives . Themaximizer aims to avoid running out of energy while simultaneously satisfying aparity condition . We show that the corresponding almost-sure problem is decidable and in $NP \cap coNP$. The same holds for checking if such a $k$ exists and if a given $k# is minimal .…

This paper investigates the construction and the decoding of a remarkable set of lattices and codes viewed as group codes . We treat in a unified framework the Leech lattice and the Golay code in dimension 24, the Nebe lattice indimension 72, the Barnes-Wall lattices, and the Reed-Muller codes .…

Multimodal learning mimics the reasoning process of the human multi-sensory system, which is used to perceive the surrounding world . In this work, we propose a novel multimodal fusion module thatlearns to emphasize more contributive features across all modalities . The MSAF module is designed to be compatible withfeatures of various spatial dimensions and sequence lengths, suitable for bothCNNs and RNNs .…

This paper investigates the performance of reconfigurable intelligent surface(RIS)-aided massive multiple multiple multiple-input multiple- input multiple-output (MIMO) systems with directlinks . We first derive the closed-form expression of the uplink ergodic data rate . Then, based on the derived expression, we use thegenetic algorithm (GA) to solve the sum data rate maximization problem .…

Targeting the right group of workers for crowdsourcing often achieves better quality results . We discuss the benefits andtrade-offs of recruiting paid vs. unpaid community-situated workers . We consider the context of Alcoholics Anonymous (AA), the largest peer support group for recovering alcoholics, and the task of identifying and validating AA meeting information .…