Transfer Learning for Human & AI

Memristors have attracted interest as neuromorphic computation elements because they show promise in enabling efficient hardware implementations of artificial neurons and synapses . We performed measurements on interface-typememristors to validate their use in neuromorphic hardware . Using this class of memristivedevices as the synaptic weight element in a spiking neural network yields, toour knowledge, one of the first models of this kind, capable of learning to be a universal function approximator, and strongly suggests the suitability of these memristors for usage in future computing platforms .…

There is a growing interest in the speech community in developing RecurrentNeural Network Transducer (RNN-T) models for automatic speech recognition (ASR) applications . We propose a modified loss function that does not enforcetemporal alignment of the training transcripts and audio .…

Deep Neural Networks are known to be very demanding in terms of computing and memory requirements . Using our proposed filter filterdamping, we achieved the 1st rank at the DCASE-2020 Challenge in the task of low-Complexity Acoustic Scene Classification . We propose afilter-damping technique for regularizing the RF of models, without altering their architecture and changing their parameter counts.…

This paper describes a method for overlap-aware speaker diarization . Given anoverlap detector and a speaker embedding extractor, our method performsspectral clustering of segments informed by the output of the overlap detector . Our method achieves a testdiarization error rate (DER) of 24.0% on the mixed-headset setting of the AMImeeting corpus, which is a relative improvement of 15.2% over a strongagglomerative hierarchical clustering baseline .…

Interface theories are powerful frameworks supporting incremental andcompositional design of systems through refinements and constructs forconjunction, and parallel composition . In this report we present a first Interface Theor — |Modal Mixed Interfaces — for systems exhibiting bothnon-determinism and randomness in their behaviour .…

When only a limited amount of accented speech data is available, the conventional approach isaccent-specific adaptation . To simplify the adaptation procedure, we exploreadapting the baseline model to multiple target accents simultaneously withmulti-accent mixed data . We propose using accent-specific top layer withgate mechanism (AST-G) to realize multi-acccent adaptation .…

In Machine Learning as a Service, a provider trains a deep neural network and provides many users access . The hosted (source) model is susceptible to model stealing attacks, where an adversary derives a \emph{surrogate model} from API access to the source model… For post hoc detection of such attacks, the provider needs a robust method to determine whether a suspect model is a surrogate of their model .…

The abstract of a scientific paper distills the contents of the paper into a short paragraph . Explicit categories could be helpful for more granular, that is, discourse-level search and recommendation . We train a deep neural network on structured abstracts from PubMed, then fine-tune it on a small hand-labeled corpus of computer science papers .…

Online end-to-end neural diarization system, BW-EDA-EEND, processes data incrementally for a variable number of speakers . System is based on the EDA architecture of Horiguchi et al., but utilizes theincremental Transformer encoder, attending only to its left contexts and usingblock-level recurrence in the hidden states to carry information from block to block, making the algorithm complexity linear in time .…

We design a framework for studying prelinguistic child voice from 3 to 24months based on state-of-the-art algorithms in di-arization . Our system consists of a time-invariant feature ex-tractor, a context-dependent embeddinggenerator, and a clas-sifier . We found that our best system achieved 43.8% DER ontestdataset, compared to 55.4% DERN achieved by LENA soft-ware .…

In visual guessing games, a Guesser has to identify a target object in ascene by asking questions to an Oracle . Existing models fail to learn truly multi-modal representations, relying instead on gold category labels . This provides an unnaturalperformance advantage when categories at inference time match those at trainingtime .…

Domain adaptation could be used to promote performance of deep neural network (DNN) acoustic model . However, sufferingfrom domain mismatch and data sparsity, domain adaptation is very challenging . This paper proposes a novel adaptation method for DNN acoustic model usingclass similarity .…

Deep learning has emerged as an effective approach for creating modern software systems, with neural networks often surpassing hand-crafted systems . Formal verification is a promising avenue for tackling this difficulty, by formally certifying that networks are correct . We propose anSMT-based technique for verifying \emph{binarized neural networks} – a popularkind of neural networks, where some weights have been binarized in order to make the network more memory and energy efficient, and quicker toevaluate .…

We present a unified network for voice separation of an unknown number of speakers . The proposed approach is composed of several separation headsoptimized together with a speaker classification branch . The separation is carried out in the time domain, together with parameter sharing between allseparation heads .…

Language models based on the Transformer architecture have achieved state-of-the-art performance on a wide range of NLP tasks such as text classification, question-answering, and token classification . We analyze these language models through extensive experiments on multiple downstream tasks in Hindi, Bengali, and Telugulanguage .…

In this work, we propose a hierarchical subspace model for acoustic unit discovery . We train the hyper-subspace on aset of transcribed languages and transfer it to the target language . In so doing, we simultaneously learn a subspace of units specific to that language and the units that dwell on it .…

Deep learning (DL) models have become core modules for many applications . Deploying these models without careful performance benchmarking that considers both hardware and software’s impact often leads to poor service and operational expenditure . To facilitate DL models’ deployment, we implement an automatic and comprehensive benchmark system for DL developers .…

We develop a theory to elucidate the reason that deep neural networks perform better than other methods . We reveal that a certain class of common models are sub-optimal, including linear estimators and other harmonicanalysis methods such as wavelets and curvelets .…

Current supervised learning can learn spurious correlation during the data-fitting process, imposing issues regarding interpretability,out-of-distribution (OOD) generalization, and robustness . To avoid spurious correlation, we propose a Latent Causal Invariance Model (LaCIM) which pursuescausal prediction . We introduce latent variables that areseparated into (a) output-causative factors and (b) others that are spuriouslycorrelated to the output via confounders, to model the underlying causalfactors .…

Sports Analytics is the study of this datato provide a constructive advantage to the team and its players . Here we propose a novel Recurrent Neural Network model which can predict the win probability of a match at regular intervals .…

Automatic MRI brain tumor segmentation is of vital importance for the diseasediagnosis, monitoring, and treatment planning . In this paper, we propose atwo-stage encoder-decoder based model for brain tumor subregional segmentation . Variational autoencoder regularization is utilized in both stages to prevent overfitting issue .…

Data-driven graph learning models a network by determining the strength ofconnections between its nodes . Ourgraph learning method modifies the embeddings to compute the graphsimilarity matrix . In this work, graph learning is used to examine brainnetworks for brain state identification .…

Differentiable Annealing Indicator Search (DAIS) leverages the strength of neural architecture search in the channelpruning and automatically searches for the effective pruned model with given constraints on computation overhead . Experimental results show that DAIS outperforms state-of-the-art pruning methods on CIFAR-10 and ImageNet .…

Recent studies have shown that federated learning (FL) is vulnerable topoisoning attacks which aim at injecting a backdoor into the global model . These attacks are effective, even when performed by a single client, andundetectable by most existing defensive techniques .…

Large-scale deep-learning recommendation systems consume 79% of all inference cycles in the data center . Distributed serving, or distributed inference, divides the memory requirements of a single large model across multiple servers . This work is a first-step for the systems research community to develop novel model-serving solutions, given the huge system design space .…

ResidualLikelihood Forests (RLF) produces conditional likelihoods that are sequentially optimized using global loss in the context of previouslearners within a boosting-like framework . This increases the efficiency of our strong classifier, allowing for the design of classifiers which are more compact in terms of model capacity .…

Open set domain adaptation refers to the scenario that the target domaincontains categories that do not exist in the source domain . It is a more commonsituation in the reality compared with the typical closed set domains adaptation . We propose an “againstadversarial learning” method that can distinguish unknown target data and knowndata naturally without setting any additional hyper parameters and the target data predicted to the known classes can be classified at the same time .Experimental…

Language pairs with limited amounts of parallel data remain a challenge for neural machine translation . Using an optimized Transformer for low-resource conditions improves the translation quality up to 7.3 BLEU points compared to using the Transformer default settings . Our experiments on different subsets of the IWSLT14 training data show that theeffectiveness of Transformer under low-Resource conditions is highly dependent on the hyper-parameter settings .…

This paper gives a review of concentration inequalities which are widelyemployed in analyzes of mathematical statistics in a wide range of settings . From sub-Gaussian to sub-exponential, sub-Weibull random variables, we aim to illustrate theconcentration inequalities with known constants and to improve existing boundswith sharper constants .…

Local ancestry inference (LAI) identifies the ancestry of each segment of an individual’s genome and is an important step in medical and population geneticstudies . We formulate the LAItask as an image segmentation problem and develop a new LAI tool using a deepconvolutional neural network with an encoder-decoder architecture .…

Classical linear discriminant analysis (LDA) is based on squared Frobeniousnorm and hence is sensitive to outliers and noise . To improve the robustness of LDA, we introduce capped l_{2,1}-norm of a matrix, which employsnon-squared l_2-norm and “capped” operation, we propose a novel cappedl_{2,.1}norm linear discriminative analysis .…

We explore role of Conditional Generative Adversarial Networks (GAN) inimputing missing data . We apply GAN imputation on a novel use case ine-commerce: a learning-to-rank problem with incomplete training data . Using an AmazonSearch ranking dataset, we produce standard ranking models trained onGAN-imputed data that are comparable to training on ground-truth data based on NDCG and MRR .…

Patients with respiratory conditions typically exhibit adventitiousrespiratory sounds, such as wheezes . Wheeze events have variable duration . We studied the influence of event duration on wheeze classification, including how the creation of the non-wheeze class affected the classifiers’ performance .…

Speech enhancement (SE) based on deep speech prior has attracted attention, such as the variational auto-encoder with non-negative matrixfactorization (VAE-NMF) architecture . Despite the clear advantage in theory, we argue that deep priors must be used with much caution, since the likelihood produced by a deep generative model does not always coincide with the speech quality .…

In the settings of conventional domain adaptation, categories of the sourcedataset are from the same domain, which is not always true in reality . In this paper, we propose a mixed set domain adaptation method that can reduce distribution discrepancy between different categories .…

Previous multi-agent pathfinding (MAPF) methods hardly take formation into consideration . Other decentralized partially observable approaches to MAPF arereinforcement learning (RL) methods . However, these RL methods encounterdifficulties when learning path finding and formation problem at the same time . In this paper, we propose a novel algorithm that uses a hierarchical structure to decompose the multi objective task into unrelated ones .…

Many adaptive optimization methods have been proposed and used in deeplearning . Adam is regarded as the default algorithm and widely used in many deep learning frameworks . Recently, many variants of Adam, such asAdabound, RAdam and Adabelief, show better performancethan Adam .…

Vision-based deep learning using convolutional neural networks is apromising approach for providing useful force estimates . The network with both state and vision inputs outperformed aphysics-based baseline model in accuracy . It showed comparable accuracy butfaster computation times than a baseline recurrent neural network, making itbetter suited for real-time applications.…

Learned optimizers are algorithms that can themselves be trained to solve optimization problems . In contrast to baseline optimizers (such as momentum orAdam) that use simple update rules, learnedoptimizers use flexible, high-dimensional, nonlinear parameterizations . We study learned optimizers trained from scratch on threedisparate tasks, and discover that they have learned interpretable mechanisms, including: momentum, gradient clipping, learning rate schedules, and a new form of learning rate adaptation .…

Retrosynthesis is the process of recursively decomposing target molecules into available building blocks . It plays an important role in solving problems in organic synthesis planning . The method disassembles retrosyntheses into twosteps: i. identify the potential reaction center of the target molecule through a novel graph neural network and generate intermediate synthons .…

This study is focused on applying genetic algorithms (GA) to model and bandselection in hyperspectral image classification . We use a forensic-inspired data set of seven hyperspectrals images with blood and five visually similarsubstances to test GA-optimised classifiers in two scenarios: when the trainingand test data come from the same image and when they come from different images, which is a more challenging task due to significant spectradifferences .…

The teacher-student knowledge distillation framework hasdemonstrated its potential in training Graph Neural Networks (GNNs) However,due to the difficulty of training deep and wide GNN models, one can not alwaysobtain a satisfactory teacher model for distillation . We propose the first GNNSelf-Distillation framework for GNNs, termed GNN self-distillation, that serves as a drop-in replacement for improving the training process .…

This paper considers the minimization of a sum of smooth and strongly convexfunctions dispatched over the nodes of a communication network . We introduce a new communication scheme, based on Loss-Networks, that is programmable in a fully asynchronous anddecentralized fashion .…

In this paper we propose a 2D deep residual Unet with 104 convolutionallayers (DR-Unet104) for lesion segmentation in brain MRIs . We make multipleadditions to the Unet architecture, including adding the ‘bottleneck’ residualblock to Unet encoder and adding dropout after each convolution blockstack .…

Deep learning had already demonstrated its power in medical images, including denoising, classification, segmentation, etc. All these applications are proposed to automatically analyze medical images beforehand, which brings more information to radiologists during clinical assessment for accuracyimprovement . In experiments, we apply the proposed framework to differentdatasets, models, and use cases .…

BASS, with its lowcomputational cost and fast convergence, obtained SPs 100 times faster than the current best greedy approaches . BASS wastested with four reconstruction methods for parallel MRI based on low-rankness and sparsity that allow a free choice of the SP .…

In recent years, deeper and wider neural networks have shown excellent performance in computer vision tasks . Several methods have been proposed to compress the size of the networks without reducing networkperformance . We present a novel incremental training algorithm for deepneural networks called planting .…

Translation of imagined speech electroencephalogram(EEG) into humanunderstandable commands greatly facilitates the design of naturalistic braincomputer interfaces . To achieve improved imagined speech unit classification,this work aims to profit from the parallel information contained in multi-phasal EEG data recorded while speaking, imagining and performingarticulatory movements corresponding to specific speech units .…

This paper presents a recurrent hybrid model and training procedure fortask-oriented dialogue systems based on Deep Recurrent Q-Networks (DRQN) The model copes with both tasks required for Dialogue Management: State Trackingand Decision Making . It is based on modeling Human-Machine interaction into alatent representation embedding an interaction context to guide the discussion .…

Adaptive meshing includes local refinement as well as coarsening of meshes . In ameshref, no explicit information about the refinement process is stored, but is instead implicit in the datastructure . Thereby, the desired properties are guaranteed andcomputational efficiency is maintained .…