From System Level Synthesis to Robust Closed loop Data enabled Predictive Control

The Willem’s fundamental lemma and the system level synthesis characterizetrajectories generated by a linear system . While the former method is valid fordeterministic LTI systems, the latter one is further effective for stochasticlinear time varying systems . A causal feedback structure is further derived, leading to an computational cost similar to standard robust MPC with full statemeasurements .…

Multiversal views on language models

The virtuosity of language models like GPT-3 opens a new world of possibility for human-AI collaboration in writing . In this paper, we present a framework inwhich generative language models are conceptualized as multiverse generators . We call for exploration into this commonality through new formsof interfaces which allow humans to couple their imagination to AI to write,explore, and understand non-linear fiction .…

The Structure of Minimum Vertex Cuts

In this paper we continue a long line of work on representing the cutstructure of graphs . We classify the types minimum vertex cuts, and the possible relationships between multiple minimum vertices . We exhibit a simple $O(\kappan)-space data structure that can quickly answer pairwise$(\kappa+1)-connectivity queries in a $kappa$-connected graph .…

Leveraging Artificial Intelligence to Analyze Citizens Opinions on Urban Green Space

The quality of urban green space has been measured by expertassessments, including in-situ observations, surveys, and remote sensinganalyses . Location data platforms, such as TripAdvisor, can provide people’sopinion on many destinations and experiences, including UGS . Such an application can support local authorities andstakeholders in their understanding of and justification for future investments in urban green spaces, says the author of a new paper on the subject of urban greenspace quality assessments .…

A more accurate view of the Flat Wall Theorem

We introduce a supporting combinatorial framework for the Flat Wall Theorem . We suggest two variants of the theorem and introduce a new,more versatile, concept of wall homogeneity as well as the notion of regularity in flat walls . All proposed concepts and results aim at facilitating the use of the irrelevant vertex technique in future algorithmic applications .…

A Little Pretraining Goes a Long Way A Case Study on Dependency Parsing Task for Low resource Morphologically Rich Languages

The bottleneck of massive labeled data limits the effectiveness of these approaches for low resource languages . We perform experiments on 10 MRLs in low-resource settings to measure the efficacy of our proposed pretraining method . We observe an average absolute gain of 2 points (UAS) and 3.6 points (LAS) Code and data available at:https://://://github.com/jivnesh/LCM.LCM…

Improving Zero shot Neural Machine Translation on Language specific Encoders Decoders

Recently, universal neural machine translation (NMT) with sharedencoder-decoder gained good performance on zero-shot translation . Unlike universal NMT, jointly trained language-specific encoders-decoders aim toachieve universal representation across non-shared modules, each of which isfor a language or language family . We propose to generalizethe non-shottranslation architecture by differentiating theTransformer layers between language-Specific and interlingua .…

ReLU Neural Networks for Exact Maximum Flow Computation

Understanding the great empirical success of artificial neural networks is currently one of the hottest research topics in computer science . In this paper we study the expressive power of NNswith rectified linear units from a combinatorial optimization perspective . We show that, given a directed graph with $n$ nodes and $m$ arcs,there exists an NN of polynomial size that computes a maximum flow from any possible real-valued arc capacities as input .…

Optimizing Inference Performance of Transformers on CPUs

The Transformer architecture revolutionized the field of natural languageprocessing (NLP) Transformers-based models (e.g., BERT) power many importantWeb services, such as search, translation, question-answering, etc. The optimizations are evaluated using theinference benchmark from HuggingFace, and are shown to achieve the speedup of up to x2.36 .…

VARA TTS Non Autoregressive Text to Speech Synthesis based on Very Deep VAE with Residual Attention

This paper proposes VARA-TTS, a non-autoregressive (non-AR) text-to-speech(TTS) model . It uses a very deep Variational Autoencoder (VDVAE) with ResidualAttention mechanism, which refines the textual-to.-acoustic alignmentlayer-wisely . An utterance-level speaking speed factor is computed by a jointly-trained speaking speed predictor, which takes the mean-pooled latent variables of the coarsest layer as input, to determine number of acousticframes at inference .…

Fair Robust Assignment using Redundancy

We study the consideration of fairness in redundant assignment formulti-agent task allocation . Solving this problem optimally is NP-hard . We exploit properties ofsupermodularity to propose a polynomial-time, near-optimal solution . Empirically, our algorithm outperforms benchmarks, scales to large problems, and provides improvements in both fairness and average utility.…

Supporting search engines with knowledge and context

We aim to support search engines in leveraging knowledge while accounting for context . We study how to make structured knowledge moreaccessible to the user when the search engine proactively provides such knowledge as context to enrich search results . We propose to model query resolution as a term classification task and propose a method to address it .…

From perspective maps to epigraphical projections

The projection onto the epigraph or a level set of a closed proper convexfunction can be achieved by finding a root of a scalar equation . The approach is based on the variational-analytic properties of general convexoptimization problems that are (partial) infimal projections of the sum of the function in question and the perspective map of a convex kernel .…

Do as I mean not as I say Sequence Loss Training for Spoken Language Understanding

Spoken language understanding (SLU) systems extract transcriptions, as well as semantics of intent or named entities from speech . We propose non-differentiable sequence losses based on SLU metrics as a proxy for semantic error . We show that custom sequence loss training is the state-of-the-art on open SLU datasets and leads to 6% relative improvement in both ASR and NLU performance metrics on large proprietary datasets .…

Well posedness theory for nonlinear scalar conservation laws on networks

We consider nonlinear scalar conservation laws posed on a network . We establish $L^1$ stability, and thus uniqueness, for weak solutions satisfying the entropy condition . We demonstrate themethod’s properties through several numerical experiments . In one important case — for monotone fluxes with an upwinddifference scheme — we show that the set of (discrete) stationary solutions is sufficiently large to suit our general theory .…

Contrastive Unsupervised Learning for Speech Emotion Recognition

Speech emotion recognition (SER) has long suffered from a lack of large-scale labeled datasets . We show that the contrastive predictive coding (CPC) method can learn salientrepresentations from unlabeled datasets . In our experiments, this method achieved state-of-the-art correlation coefficient (CCC) performance for all emotionprimitives (activation, valence, and dominance) on IEMOCAP.…

Low precision logarithmic number systems Beyond base 2

Logarithmic number systems (LNS) are used to represent real numbers in many applications using a constant base raised to a fixed-point exponent making its distribution exponential . This greatly simplifies hardware multiply, divide andsquare root . LNS with base-2 is most common, but in this paper we show that for low-precision LNS the choice of base has a significant impact .…

Barriers for recent methods in geodesic optimization

We study a class of optimization problems including matrix scaling, matrixbalancing, multidimensional array scaling, operator scaling, and tensor scaling . Trust region methods, which includebox-constrained Newton’s method, are known to produce high precision solutionsvery quickly for matrix scaling and matrix balancing .…

High Order Control Lyapunov Barrier Functions for Temporal Logic Specifications

Recent work has shown that stabilizing an affine control system to a desired state while optimizing a quadratic cost subject to state and controlconstraints can be reduced to a sequence of Quadratic Programs . In this paper, we generalize HOCBFs to High OrderControl Lyapunov-Barrier Functions (HOCLBFs) We also show that the proposed HOCLFs can be used to guarantee the Boolean satisfaction of Signal TemporalLogic (STL) formulae over the state of the system .…

Multi source Pseudo label Learning of Semantic Segmentation for the Scene Recognition of Agricultural Mobile Robots

In conventionalsemantic segmentation methods, the labels are given by manual annotation, which is a tedious and time-consuming task . Unsupervised domain adaptation (UDA) transferknowledge from labeled source datasets to unlabeled target datasets . We propose to use multiple publicly available datasets of outdoor images as source datasets, and also propose asimple yet effective method of generating pseudo-labels by transferring knowledge from the source datasets that have different appearance and a labelset from the target datasets.…

Mind the beat detecting audio onsets from EEG recordings of music listening

We propose a deep learning approach to predicting audio event onsets inelectroencephalogram (EEG) recorded from users as they listen to music . We use a publicly available dataset containing ten contemporary songs and concurrentlyrecorded EEG . We generate a sequence of onset labels for the songs in ourdataset and trained neural networks (a fully connected network (FCN) and arecurrent neural network (RNN) to parse one second windows of input EEG topredict .…

Numerical investigation of Mach number consistent Roe solvers for the Euler equations of gas dynamics

Traditional approaches to prevent the carbuncle phenomenon in gasdynamics simulations increase the viscosity on entropy and shear waves nearshocks . The goal is to achieve what, in thispaper, we call Mach number consistency . We take the simple approach that was used for the proof of concepttogether with the simple model for the increased numerical viscoity on linearwaves .…

Neural Architecture Search as Program Transformation Exploration

Improving the performance of deep neural networks is important to both the compiler and neural architecture search (NAS) communities . We express neural architectureoperations as program transformations whose legality depends on a notion ofrepresentational capacity . This allows them to be combined with existing transformations into a unified optimization framework .…

Robust Hybrid High Order method on polytopal meshes with small faces

We design a Hybrid High-Order scheme for the Poisson problem that is robust on polytopal meshes in the presence of small edges/faces . We stategeneral assumptions on the stabilisation terms involved in the scheme, underwhich optimal error estimates are established with multiplicative constants that do not depend on the maximum number of faces in each element, or the relative size between an element and its faces .…

Universal Adversarial Perturbations for Malware

Universal Adversarial Perturbations (UAPs) identify noisypatterns that generalize across the input space . UAPs allow attackers to greatlyscale up the generation of these adversarial examples . Ourexperiments limit the effectiveness of a white box Android evasion attack to~20% at the cost of 3% TPR at 1% FPR.…

Visualizing hierarchies in scRNA seq data using a density tree biased autoencoder

Single cell RNA sequencing (scRNA-seq) data makes studying the development of cells possible at unparalleled resolution . Given that many cellulardifferentiation processes are hierarchical, their scRNA-sequencing data is expected to be approximately tree-shaped in gene expression space . Inference andrepresentation of this tree-structure in two dimensions is highly desirable forbiological interpretation and exploratory analysis .…

Customizable Stochastic High Fidelity Model of the Sensors and Camera onboard a Low SWaP Fixed Wing Autonomous Aircraft

The navigation systems of autonomous aircraft rely on the readings provided by a suite of onboard sensors to estimate the aircraft state . Anaccurate representation of the behavior and error sources of each of thesesensors, together with the images generated by the cameras, in indispensable for flight simulation and the evaluation of novel inertial or visual navigational algorithms .…

Exactness and Convergence Properties of Some Recent Numerical Quadrature Formulas for Supersingular Integrals of Periodic Functions

In a recent work, we developed three compact numerical quadratureformulas for finite-range periodic supersingular integrals $I[f[f]=t/n$ with $h=T/n . We prove that these formulas have spectral accuracy . We also prove that, when $u(z)$ is analytic in astrip $Im, the errors in all three formulas are $O(e^{-2n\pi\sigma/T)$ as$n\to\infty$ for all practical purposes .…

Which Regular Languages can be Efficiently Indexed

In the present work, we study the hierarchy of $p$-sortable languages: regular languages accepted by automata of width $p$. In this hierarchy, regular languages are sorted according to the new fundamental measure of NFA complexity$p$. Our main contributions are: (i) we show that the hierarchyis strict and does not collapse, (ii) we provide (exponential) upper and lowerbounds relating the minimum widths of equivalent NFAs and DFAs, and (iii) wecharacterize DFAs of minimum $p $ for a given $\mathcal L$ via aco-lexicographic variant of the Myhill-Nerode .…