Transfer Learning for Human & AI

This paper explores how different ideas of racial equity in machine learning can present trade-offs that are difficult to solve computationally . Machine learning is often used in justice settings to create risk assessments, which are used to determine interventions, resources,and punitive actions .…

Koopman operators provide tractable means of learning linear approximationsof non-linear dynamics . Many approaches have been proposed to find theseoperators, typically based upon approximations using an a-priori fixed class of models . However, choosing appropriate models and bounding the approximationerror is far from trivial .…

The robust principles of treating interference as noise (TIN) and avoiding it when it is not, form the background of this work . Combining TIN with the topological interference management (TIM)framework that identifies optimal interference avoidance schemes, we formulate a TIM-TIN problem .…

The problem of describing the way in which an incoming acoustic wave isscattered by an elastic body immersed in a fluid is one of central importance in detecting and identifying submerged objects . The problem can be treated by the boundary-field equation method, provided an appropriatescaling factor is employed .…

Simultaneous Localization and Mapping (SLAM) has found an increasingutilization lately, such as self-driving cars, robot navigation, 3D mapping, virtual reality (VR) and augmented reality (AR) The employment of datasets is essentially a kind ofsimulation but profits many aspects – capacity of drilling algorithm hourly,exemption of costly hardware and ground truth system, and equitable benchmarkfor evaluation.…

The backpropagation algorithm for neural networks is widely felt hard to understand . This paper provides a new derivation of this algorithm based on the concept of derivative amplification coefficients . The concept is found to well carry over to conventional feedforward neural networks and it paves the way for the use of mathematical induction in establishing a key result that enablesbackpropagating for derivative amplification .…

We introduce a novel approach to automated termination analysis of computer programs . We train neural networks to act as ranking functions . The existence of a valid ranking function provesthat the program terminates . We present a custom loss function for learning lexicographic ranking functions and uses satisfactioniability modulo theories for verification .…

We study high-dimensional nonlinear approximation of functions inH\”older-Nikol’skii spaces $H^\alpha_\infty(\mathbb{I}^d) on the unit cube$\mathbb {I}$d:=[0,1]^d$ We derived a novel right asymptotic orderof orderof noncontinuous manifold $N-widths of the unit ball of $H$N- widths . Inconstructing approximation methods, the function decomposition by the tensorproduct Faber series plays a central role .…

Methods combine the well-established GMRES method for linear systems with algorithms for nonlineareigenvalue problems (NEPs) to generate a basis for the Krylov subspace . Weshow convergence factor bounds obtained similarly to those for the method GMRESfor linear systems. More specifically, a bound is obtained based on themagnitude of the parameter $mu$ and the spectrum of the linear companionmatrix, which corresponds to the reciprocal solutions to the corresponding NEP.…

Outcome prediction from clinical text can prevent doctors from overlookingpossible risks and help hospitals plan capacities . We simulate patients at admission time, when decision support can be especially valuable, and contribute a novel admission to discharge task . We show that our approach improves performance on the outcome tasks against several baselines .…

A client’s data is stored on a local data server and the clients communicate onlymodel parameters with a central server, and not their data . Data privacy and protection is a crucial issue for any automatic speechrecognition (ASR) service provider when dealing with clients .…

Machine learning is essential for advanced material discovery . We show how training aneural network to replace the exchange-correlation functional within afully-differentiable three-dimensional Kohn-Sham density functional theory framework can greatly improve simulation accuracy . Using only eight experimental data points on diatomic molecules, our trainedexchange-correlrelation network provided improved prediction of atomization andionization energies across a collection of 110 molecules when compared with commonly used DFT functionals and more expensive coupled clustersimulations .…

The classical Langevin Monte Carlo method looks for i.i.d. samples from atarget distribution by descending along the gradient of the target distribution . It is popular partially due to its fast convergence rate . However, the numerical cost is sometimes high because the gradient can be hard to obtain .…

Polar code lattices are formed from binary polar codes using Construction D . The dimension $n$ and target probability of decoding error areparameters for this design . At an error-rate of $10^{-4$ a lattice achieves a VNR of 2.5 dB, within 0.2 dB of the best-knownBCH code lattice .…

Markov chain Monte Carlo (MCMC) algorithms are standard techniques to solvesuch problems . MCMC techniques are computationally challenging as they require several thousands of forward PDE solves . The goal of this paper is to introduce a fractional deep neural network based approach for the forwardsolves within an MCMC routine .…

The effectiveness of the direction of arrival (DOA) estimation in array signal processing is animportant research area . In hybrid massive MIMO systems, the receivedsignals at the antennas are not sent to the receiver directly, and spatialcovariance matrix, which is essential in MUSIC algorithm, is thus unavailable .…

Deep learning has provided new ways of manipulating, processing and analyzing data . Another subfield of machine learning namedreinforcement learning, tries to find an optimal behavior strategy through interactions with the environment . Combining deep learning and reinforcementlearning permits resolving critical issues relative to the dimensionality and scalability of data in tasks with sparse reward signals .…

In this paper we present a general procedure for designing higher strong order methods for It\^o stochastic differential equations on matrix Lie groups . We show how our higher order schemes can be applied in amechanical engineering as well as in a financial mathematics setting .…

We analyze a class of stochastic processes, referred to as asynchronous andsemi-anonymous dynamics (ASD) over directed labeled random networks . These processes are a natural tool to describe general best-response and noisybest-response dynamics in network games where each agent, at random timesgoverned by independent Poisson clocks, can choose among a finite set ofactions .…

This paper introduces a class of mixed-integer formulations for trained ReLUneural networks . The approach balances model size and tightness by partitioning node inputs into groups and forming the convex hull over thepartitions via disjunctive programming . At one extreme, one partition per input covers the tightest possible formulation foreach node .…

Robot manipulation in cluttered scenes often requires contact-rich interactions with objects . For each object in a scene, depending on its properties, the robot mayor may not be allowed to make contact with, tilt, or topple it . To ensure that these constraints are satisfied during non-prehensile interactions, a planner can query a physics-based simulator to evaluate the complex multi-bodyinteractions caused by robot actions .…

Probabilistic Face Embeddings (PFE) can improve face recognition performance by integrating data uncertainty into the featurerepresentation . However, existing PFE methods tend to be over-confident inestimating uncertainty and is too slow to apply to large-scale face matching . This paper proposes a regularized probabilistic face embedding method to improve the robustness and speed of PFE .…

BPlag is designed to be both robust to pervasive plagiarism-hidingtransformations, and accurate in the detection of plagiarised source code . It is evaluated for robustness, accuracy and efficiency against 5 commonly used source code plagiarism detection tools . It was shown to be more robust to the plagiarism of source code, but is lessefficient than compared tools .…

This paper investigates task-oriented communication for edge inference . It is critical to encode the data into an informative and compact representation for low-latency inference . We propose a learning-based communication schemethat jointly optimizes feature extraction, source coding, and channel coding in a task-orientated manner, i.e.,…

This paper considers the uncertainty in nodeaccess and download service . In one access model, a user can access each node with a fixedprobability, and the other, a random fixed-size subset of nodes . For afixed redundancy level, the systems’ service rate is determined by the allocation of coded chunks over storage nodes .…

Most speech separation methods are still far from having enough general-ization capabilities for real scenarios where the number of input sounds is usually uncertain and dynamic . In this work, we employ ideas from auditory attention with twoears and propose a speaker and direction inferred speech separation network .…

Arbitrary Conditioning with Energy (ACE) uses an energy function to specify densities . ACE is state-of-the-art for arbitrary conditional and marginal likelihoodestimation and for tabular data imputation . We also simplify the learningproblem by only learning one-dimensional conditionals, from which more complexdistributions can be recovered during inference .…

Black-box optimization (BBO) algorithms are concerned with finding the bestsolutions for the problems with missing analytical details . Most classical methods for such problems are based on strong and fixed \emph{a priori}assumptions such as Gaussian distribution . But lots of complex real-world problems are far from the distribution of the Gaussian .…

We consider the problem of computing the k-means centers for a largehigh-dimensional dataset in the context of edge-based machine learning . Wepropose to let the data sources send small summaries, generated by jointdimensionality reduction (DR) and cardinality reduction (CR) to support approximating k-Means computation at reduced complexity and communication cost .…

This paper deals with the polynomial linear system solving with errors(PLSwE) problem . The number of evaluations needed to recover the solution of the linear system is crucial since it affects the number of computations . Our work is part of a series of papers about PLSwE aiming to reduce thisnumber of evaluations .…

It remains a tough challenge to recover the speech signals contaminated by various noises under real acoustic environments . To this end, we propose anovel system for denoising in the complicated applications . The first pipeline is proposed to decouple the optimization problemw:r:t: magnitude and phase, i.e.,…

This paper studies model-based bandit and reinforcement learning (RL) with nonlinear function approximations . Global convergence is intractable even for one-layer neural net bandit with adeterministic reward . On the other hand, for convergence to localmaxima, it suffices to maximize the virtual return if the model can alsoreasonably predict the size of the gradient and Hessian of the real return .…

This paper develops a grey-box approach to stability analysis of complex power systems that facilitates root-cause tracing without requiring disclosure of the full details of the internal control structure of apparatus connected to the system . The approach is favoured by manufacturers of wind and solar systems for the limited disclosure required .…

Quality assessment metrics may allow a wide range of processes to be guided and evaluated, such as level of detail creation,compression, filtering, and so on . Almost no research has been conducted on the evaluation of texture-mapped 3D models . We introduce both texture and geometry distortions on aset of 5 reference models to produce a database of 136 distorted models .…

Sparsity in Deep Neural Networks (DNNs) has been widely studied to compress and accelerate the models on resource-constrained environments . Fine-grained sparsity can achieve a high compression ratio but is not hardware friendly and hence receives limited speed gains . We propose a noveland effective ingredient, sparse-refined straight-through estimator (SR-STE) We also define a metric, Sparse ArchitectureDivergence (SAD), to measure the sparse network’s topology change during the training process .…

This paper revisits the classical Linear Quadratic Gaussian (LQG) control from a modern optimization perspective . We analyze two aspects of the optimization landscape of the LQG problem: connectivity of the set ofstabilizing controllers and structure of stationary points . These results shed some light on the performanceanalysis of direct policy gradient methods for solving the problem .…

Virtual Element Method (VEM) was shown to be able to deal with a large variety of polygons while achieving good convergencerates . The regularity assumptions proposed in the VEM literature to guaranteethe convergence on a theoretical basis are quite general .…

The paper proposes a novel concept of docking drones to make this process assafe and fast as possible . The idea behind the project is that a robot with thegripper grasps the drone in midair . The soft hand has a unique technology of providing force information through the color of the fingers to the remote computer vision (CV)system .…

After a disaster, locating and extracting victims quickly is critical because mortality rises rapidly after the first two days . To assist search and rescueteams and improve response times, teams of camera-equipped aerial robots canengage in tasks such as mapping buildings and locating victims .…

RL has made groundbreaking advancements in robotic, datacenter managements and other applications . System-level bottlenecks in RL workloads are poorly understood . We observe fundamental structural differences in RLworkloads that make them inherently less GPU-bound than supervised learning . RL-Scope is an open-source tool available at https://://://github.com/UofT-EcoSystem/rlscope…

DeepReinforcement Learning (DRL) is one of the leading robotic automation techniquethat has been able to achieve dexterous manipulation and locomotion roboticsskills . We propose a multi-subtask reinforcement learning method where complex tasks can be decomposed into low-level subtasks . Thesesubtasks can be parametrised as expert networks and learnt via existing DRL methods .…

Deep neural networks have been successfully applied in channelcoding to improve the decoding performance . However, state-of-the-artneural channel decoders cannot achieve high decoding performance and lowcomplexity simultaneously . To overcome this challenge, in this paper we proposedoubly residual neural (DRN) decoder .…

Scenariobased testing is an approach to tackle this problem and reduce necessary test drives by replacing driven kilometers with simulations of relevant or interesting scenarios . These scenarios can be begenerated or extracted from recorded data with machine learning algorithms or created by experts .…

DeepEST looks for failing test cases in the operational dataset of a DNN, with the goal of assessing the DNN expected accuracy by a small and ”informative” test suite . The results show that DeepEST provides DNNaccuracy estimates with precision close to (and often better than) those of existing sampling-based DNN testing techniques, while detecting from 5 to 30times more mispredictions with the same test suite size .…

Identifying the origin of a sample image in biometric systems can be beneficial for data authentication in case of attacks against the system and initiating sensor-specific processing pipelines . Motivated by shortcomings of the photo response non-uniformity(PRNU) based method in the biometric context, we use a texture classificationapproach to detect the origin .…

The performance of algorithms for neural architecture search strongly depend on the parametrization of the search space . We use contrastive learning to identify networks across different initializations based on their dataJacobians . We show that traditional black-box optimization algorithms, without modification, can reach state-of-the-art performance in Neural ArchitectureSearch .…

Uncertainty quantification of groundwater (GW) aquifer parameters is critical for efficient management and sustainable extraction of GW resources . Here we develop a Bayesian inversion framework that usesInterferometric Synthetic Aperture Radar (InSAR) surface deformation data to ferry the laterally heterogeneous permeability of a confined GW aquifer .…

The research presents an overhead view of 10 important objects and followsthe general formatting requirements of the most popular machine learning task:digit recognition with MNIST . A prototype deep learning approach with transfer learning and convolutional neural networks (MobileNetV2)correctly identifies the ten overhead classes with an average accuracy of96.7%.…