Transfer Learning for Human & AI

What would it take to teach a machine to behave ethically? While broadethical rules may seem straightforward to state (“thou shalt not kill”), applying such rules to real-world situations is far more complex . We present Commonsense NormBank, a moral textbook customized for machines, which compiles 1.7M examples of people’s ethical judgments on a broad spectrum of everyday situations .…

Existing approaches to lifelong language learning rely on plenty of labeleddata for learning a new task, which is hard to obtain in most real scenarios . We expect the models also to be able to generalize well on new tasks without forgetting the previous ones .…

Conventional fine-tuning of pre-trained language models tunes all model parameters and stores a full model copy for each downstream task . UniPELT incorporates different PELT methods assub modules and learns to activate the ones that best suit the current data or setup .…

In visual interactive labeling, users iteratively assign labels to data items until the machine model reaches an acceptable accuracy . In scenarios with no or very little labeled data, visual inspection of the predictions is required . Users outperform traditionalautomated approaches when estimating a model’s accuracy .…

Beyond 5G networks will operate at high frequencies with wide bandwidths . Challenges include high path loss at operating frequencies, which requires network to be deployed densely to provide coverage . The dwell-time of the UEs in these cells is small, thus supporting mobility in these types of dense networks is a challenge .…

DeepOrder is a deep learning-based model that works on thebasis of regression machine learning . DeepOrder learns failed test cases based on multiple factors including theduration and execution status of test cases . The results show that DeepOrder outperforms the industry practice and state-of-the-art test prioritization approaches in terms of time-effectiveness and fault detection effectiveness of the DeepOrder approach .…

Pun location is to identify the punning word (usually a word or phrase that makes the text ambiguous) in a given short text . Pun interpretation is to find out two different meanings of the word . DANN (Dual-Attentive Neural Network) is proposed for pun location, effectively integrates word senses and pronunciation with context information to address two kinds of pun at the same time .…

Deep NLP models have been shown to learn spurious correlations, leaving them brittle to input perturbations . We develop a Retrieve-Generate-Filter-Filter technique to create counterfactualevaluation and training data with minimal human supervision . Using anopen-domain QA framework and question generation model trained on original task data, we create counterfactsuals that are fluent, semantically diverse, andautomatically labeled.…

Pun location is to identify the punning word (usually a word or phrase that makes the text ambiguous) in a given short text . Pun interpretation is to find out two different meanings of the word . DANN (Dual-Attentive Neural Network) is proposed for pun location, effectively integrates word senses and pronunciation with context information to address two kinds of pun at the same time .…

Recently, physics informed neural networks have successfully been applied to a broad variety of problems in applied mathematics and engineering . The principle idea is to use a neural network as a global ansatz function topartial differential equations . Due to the global approximation, physicsinformed neural networks struggle to display localized effects and strong non-linear solutions by optimization .…

In this paper, we propose a self-supervised audio representationlearning method and apply it to a variety of downstream non-speech audio tasks . We achieve a mean average precision(mAP) score of 0.415, which is a new state-of-the-art on this dataset . Our fine-tuned conformers also surpass ormatch the performance of previous systems pre-trained in a supervised way on several downstream tasks, we say .…

Automated audio captioning (AAC) is the task of automatically generating descriptions for general audio signals . We evaluate the performance of off-the-shelf models with aTransformer-based captioning approach . We show that sequences of audio embeddings can be processed using a Transformer encoder to produce higher-qualitycaptions .…

Some in industry and government now advocate a new technology to access targeted data: client-side scanning . CSS would enable on-device analysis of data in the clear . CSS by its nature createsserious security and privacy risks for all society while it can provide assistance for law enforcement is at best problematic, authors say .…

Power allocation is one of the fundamental problems in wireless networks . UWMMSE is a modern algorithm leveraging graph neuralnetworks . It is essential that the output power allocation of these algorithms is stable with respect to input perturbations . We focus on UWMM SE — a modern .…

Current open-domain conversational models can easily be made to talk inadequate ways . This work proposesSaFeRDialogues, a task and dataset of graceful responses to conversational feedback about safety failures . We collect a dataset of 10k dialoguesdemonstrating safety failures, feedback signaling them, and a responseacknowledging the feedback .…

We present a novel generative modeling method called diffusion normalizingflow based on stochastic differential equations . The algorithm consistsof two neural SDEs: a forward SDE that gradually adds noise to the data totransform the data into Gaussian random noise . The backward SDE converges to a diffusion process the starts with a Gaussian distribution and ends with the desired datadistribution .…

SGoLAM is ranked 2nd in the CVPR 2021 MultiON(Multi-Object Goal Navigation) challenge . It does not require any training of neural networks, it could be used in an off-the-shelfmanner, and amenable for fast generalization in new, unseen environments . The mapping module converts observations into an occupancy map, and the goal localization modulemarks the locations of goal objects .…

Sensitivity of deep-neural models to input noise is known to be a challenging problem . In NLP, model performance often deteriorates with naturally occurring noise, such as spelling errors . We present a thoroughevaluation of several state-of-the-art NLP systems’ robustness in multiple languages .…

This work identifies the fundamental limits of Multi-Access Coded Caching . Each user is connected to multiple caches in a manner that follows a generalized combinatorial topology . This topology stands out as it allows forunprecedented coding gains, even with very modest cache resources .…

In recent years, unsupervised domain adaptation (UDA) for semanticsegmentation has brought many researchers’attention . The proposed SEAT is simple, easily implemented and easy to integrate into existing adversarial learning based UDA methods . We introduce multi level adaptation by adding thelower-level features to the higher-level ones before feeding them to the discriminator, without adding extra discriminator like others.…

Methods for model explainability have become increasingly critical fortesting the fairness and soundness of deep learning . In safety-critical applications, there is need for security around not only the machine learning pipeline but also the modelinterpretation process . We show that by perturbing the examples of the concept that is being investigated, we can radically change the output of the interpretability method, e.g.…

E-commerce companies face situations where they want to promote and sell new or discounted items . The MIPS (maximum inner product search) finds the item with the highest inner product with a given query user . We propose Simpfer, a simple, fast,and exact algorithm for reverse MIPS .…

With proliferation of deep learning (DL) applications in diverse domains, vulnerability of those PHM algorithms to adversarial attacks has become an increasingly important topic in the domains of Computer Vision (CV) and NaturalLanguage Processing (NLP) We use two real-world PHM applications as examples tovalidate our attack strategies and to demonstrate that PHM .…

This paper investigates unsupervised learning of Full-Waveform Inversion . The problem is mathematically formulated by a secondorder partial differential equation (PDE), but is hard to solve . We use finitedifference to approximate the forward modeling of PDE as a differentiable operator (from velocity map to seismic data) and model its inversion by CNN .…

Social concepts referring to non-physical objects are powerful tools to describe, index, and query the content of visual data . We propose the translation of recent theories about social concept representation into a software approach to represent them as multimodal frames, by integrating multisensory data .…

Sensor drift is a long-existing unpredictable problem that deteriorates the performance of gaseous substance recognition, calling for an antidrift domainadaptation algorithm . The main concept is that CNNs extract not only the domain-specific features of samples but alsothe domain-invariant features underlying both the source and target domains .…

Neural MMO is a computationally accessible research platform that combines large agent populations, long time horizons, open-ended tasks, and modular gamesystems . We present Neural MMO as free and opensource software with active support, ongoing development, documentation, and training, logging, and visualization tools to help users adapt to the new setting .…

A novel signaling design for secure transmission over two-user multiple-inputmultiple-output non-orthogonal multiple access channel using deep neuralnetworks (DNNs) is proposed . The goal of the DNN is to form the covariancematrix of users’ signals such that the message of each user is transmitted reliably while being confidential from its counterpart .…

Social concepts referring to non-physical objects are powerful tools to describe, index, and query the content of visual data . We propose the translation of recent theories about social concept representation into a software approach to represent them as multimodal frames, by integrating multisensory data .…

We present a simple, yet effective, approach for self-supervised 3D humanpose estimation . During training, we rely ontriangulating 2D body pose estimates of a multiple-view camera system . Atemporal convolutional neural network is trained with the generated 3Dground-truth and the geometric multi-view consistency loss, imposinggeometrical constraints on the predicted 3D body skeleton .…

We study a new family of inverse problems for recovering representations ofcorrupted data . We assume access to a pre-trained representation learningnetwork R(x) that operates on clean images . We propose a supervisedinversion method that uses a contrastive objective to obtain excellentrepresentations for highly corrupted images .…

In this paper, we present a pre-trained language model (PLM) based framework called RID for conversational recommender system (CRS) RID finetunes the large-scale PLMs such as DialoGPT, together with a pre trained Relational GraphConvolutional Network (RGCN) to encode the node representations of anitem-oriented knowledge graph .…

Pre-trained language models (PLMs) have revolutionized the field of NLP, not only in the general domain but also in the biomedical domain . In this work we introduceeHealth, a biomedical PLM in Chinese built with a new pre-training framework . This new framework trains eHealth as a discriminator through both token-leveland sequence-level discrimination .…

Modern audio-visual datasets contain biases that undermine the real-world performance of models trained on that data . We introduce Spoken ObjectNet to remove some of these biases . This dataset expands upon ObjectNet, which is a bias-controlled image dataset . We detail our datacollection pipeline, which features several methods to improve caption quality, including automated language model checks .…

Modern software systems rely on machine learning models to make data-driven decisions based on interactions with users and systems . For broader adoption, this practice must accommodate software engineers without ML backgrounds, and (ii) providemechanisms to optimize for product goals .…

A unified-modalSpeechT5 framework explores the encoder-decoder pre-training for self-supervised speech/text representation learning . SpeechT5 canpre-train on a large scale of unlabeled speech and text data to improve the ability of the speech and textual modeling . To align the textual and speech information into a unified semantic space, we propose a random mixing-up method .…

A properlypre-trained model endows an important property: transferability . A highertransferability of the learned representations indicates a bettergeneralizability across domains of different distributions . Transferability has become the key to enable data-efficient deep learning, however, existing pre-training methods focus only on the domaintransferability .…

Image- and video-based 3D human recovery (i.e. pose and shape estimation)have achieved substantial progress . However, due to the prohibitive cost ofmotion capture, existing datasets are often limited in scale and diversity . In this work, we obtain massive human sequences as well as their 3D ground truths by playing video games .…

Deep NLP models have been shown to learn spurious correlations, leaving them brittle to input perturbations . We develop a Retrieve-Generate-Filter-Filter technique to create counterfactualevaluation and training data with minimal human supervision . Using anopen-domain QA framework and question generation model trained on original task data, we create counterfactsuals that are fluent, semantically diverse, andautomatically labeled.…

A classical computer works with ones and zeros, whereas a quantum computer uses ones, zeros and superpositions of ones and zero . In a cloud-supported cyber-physical system environment, running a machine learning application in quantum computers is often difficult . With the combination of quantum-classical neural networks (NN), complex and high-dimensional features can be extracted by the classical NN to a reduced butmore informative feature space to be processed by the existing quantumcomputers .…

We consider casting the ACSAtasks into natural language generation tasks, using natural language sentencesto represent the output . Our method allows more direct use of pre-trained knowledge in seq2seq language models by directly following the task setting during pre-training . Experiments on several benchmarks show that our methodgives the best reported results, having large advantages in few-shot and zero-shot settings .…

Using vibration data in conjunction with pattern recognition methods is one of the most common fault detection strategies for structures . Deep learning facilitates the fault detectionprocedure by automating the feature extraction and selection, and classification procedure . This study proposes an ensemble deeplearning framework based on a convolutional neural network (CNN) andDempster-Shafer theory (DST) Called CNN-DST, it classifies turbine blades with an average prediction accuracy of 97.19% .…

The confluence of 5G and AI is transforming wireless networks to deliver diverse services at the Edge, driving towards a vision of pervasive distributedintelligence . Future 6G networks will need to deliver quality of . experiencethrough seamless integration of communication, computation and AI .…

Quantum annealing (QA) has gained considerable attention because it can be applied to combinatorial optimization problems . In recent years, research on solving practical combinatorials optimization problems using them hasaccelerated . In our study, wedetermine that QA demonstrates better performance than the solvers in the breakminimization problem in a mirrored double round-robin tournament (MDRRT) We also explain the desirable performance of QA for the sparse interaction betweenvariables and a problem without constraints .…

Bayesian optimization (BO) has been effective for improving closed-loop performance by automatically tuning controller gains or reference setpoints in a model-free manner . However,BO methods have rarely been tested on dynamical systems with unmodeled dynamics . In this paper, we propose a violation-aware BO algorithm (VABO) that optimizes performance while simultaneously learningconstraint-feasible solutions .…

Adversarial attacks and backdoor attacks harness task-irrelevant features of datain their implementation . Text style is a feature that is naturally irrelevant to most NLP tasks, and thus suitable for adversarial attacks . The attack success rates can exceed 90% without much effort .…

Automated audio captioning (AAC) is the task of automatically generating descriptions for general audio signals . We evaluate the performance of off-the-shelf models with aTransformer-based captioning approach . We show that sequences of audio embeddings can be processed using a Transformer encoder to produce higher-qualitycaptions .…

Markov chains with variable length are useful parsimonious stochastic modelsable to generate most stationary sequence of discrete symbols . The idea is to identify suffixes of the past, called contexts, that are relevant topredict the future symbol . Sometimes a single state is a context, and lookingat the past and finding this specific state makes the further past irrelevant .…

Dense retrieval (DR) methods conduct text retrieval by first encoding texts in the embedding space and then matching them by nearest neighbor search . This requires strong locality properties from the representation space, i.e, theclose allocations of each small group of relevant texts, which are hard to generalize to domains without sufficient training data .…

We propose aCLIP-Enhanced method to incorporate the image-text pretrained knowledge intodownstream video-text tasks . Combined with several other improved designs, ourmethod outperforms the state-of-the-art by $2.4\%$ ($57.58$ to $60.00$)…