This paper presents a novel unifying framework of bilinear L STMs that can represent and utilize the nonlinear interaction of the input feat ures present in sequence datasets for achieving superior performance over a linear L STM and yet not incur more parameters to be learned. To realize this, our unifying framework allows the expressivity of the linear vs. bilinear terms to be balan ced by correspondingly trading off between the hidden state vector size vs. approxi mation quality of the weight matrix in the bilinear term so as to optimize the perfo rmance of our bilinear LSTM, while not incurring more parameters to be learned. W e e mpirically evaluate the performance of our bilinear LSTM in several languag e-based sequence learning tasks to demonstrate its general applicability. Recurrent neural networks (RNNs) are popularized by their impressive performance in a wide variety of supervised and unsupervised sequence learning t asks, which include language modeling (Merity et al., 2018), statistical machine translation (Bahdanau et al., 2015), and coreference resolution (Lee et al., 2017). Different variants of RNNs su ch as long short-term memory (LSTM) networks (Hochreiter & Schmidhuber, 1997) and gated recurr ent units (Cho et al., 2014) share a common architectural trait of being built by feedforward ne ural networks connected in a recurrent manner. Typically, a RNN is instantiated by linear neurons coupled w ith a nonlinear activation function, which constitute its basic building blocks; to be consisten t with the literature (Park & Zhu, 1994), we refer to such neurons as linear . This should naturally affect the processin g of adjacent words based on context in a nonlinear manner (see Table 2 in Section 4.2). A natural language is usually well-defined by a grammar full of complex context-sensitive interaction s (Table 3, Section 4.4).

We develop a framework for estimating unknown partial differential equations from noisy data, using a deep learning approach. Given noisy samples of a solution to an unknown PDE, our method interpolates the samples using a neural network, and extracts the PDE by equating derivatives of the neural network approximation. Our method applies to PDEs which are linear combinations of user-defined dictionary functions, and generalizes previous methods that only consider parabolic PDEs. We introduce a regularization scheme that prevents the function approximation from overfitting the data and forces it to be a solution of the underlying PDE. We validate the model on simulated data generated by the known PDEs and added Gaussian noise, and we study our method under different levels of noise. We also compare the error of our method with a Cramer-Rao lower bound for an ordinary differential equation. Our results indicate that our method outperforms other methods in estimating PDEs, especially in the low signal-to-noise regime.

Federated learning is a distributed, on-device computation framework that enables training global models without exporting sensitive user data to servers. In this work, we describe methods to extend the federation framework to evaluate strategies for personalization of global models. We present tools to analyze the effects of personalization and evaluate conditions under which personalization yields desirable models. We report on our experiments personalizing a language model for a virtual keyboard for smartphones with a population of tens of millions of users. We show that a significant fraction of users benefit from personalization.

ABSTRACT While deep learning has received a surge of interest in a variety of fields in recent years, major deep learning models barely use complex numbers. However, speech, signal and audio data are naturally complex-valued after Fourier Transform, and studies have shown a potentially richer representation of complex nets. In this paper, we propose a Complex Transformer, which incorporates the transformer model as a backbone for sequence modeling; we also develop attention and encoder-decoder network operating for complex input. The model achieves state-of-the-art performance on the MusicNet dataset and an In-phase Quadrature (IQ) signal dataset. The GitHub implementation to reproduce the experimental results is available at https://github.com/

This paper shows the susceptibility of spectrogram-based audio classifiers to adversarial attacks and the transferability of such attacks to audio waveforms. Some commonly adversarial attacks to images have been applied to Mel-frequency and short-time Fourier transform spectrograms and such perturbed spectrograms are able to fool a 2D convolutional neural network (CNN) for music genre classification with a high fooling rate and high confidence. Such attacks produce perturbed spectrograms that are visually imperceptible by humans. Experimental results on a dataset of western music have shown that the 2D CNN achieves up to 81.87% of mean accuracy on legitimate examples and such a performance drops to 12.09% on adversarial examples. Furthermore, the audio signals reconstructed from the adversarial spectrograms produce audio waveforms that perceptually resemble the legitimate audio.

Machine Learning systems are vulnerable to adversarial attacks and will highly likely produce incorrect outputs under these attacks. There are white-box and black-box attacks regarding to adversary's access level to the victim learning algorithm. To defend the learning systems from these attacks, existing methods in the speech domain focus on modifying input signals and testing the behaviours of speech recognizers. We, however, formulate the defense as a classification problem and present a strategy for systematically generating adversarial example datasets: one for white-box attacks and one for black-box attacks, containing both adversarial and normal examples. The white-box attack is a gradient-based method on Baidu DeepSpeech with the Mozilla Common Voice database while the black-box attack is a gradient-free method on a deep model-based keyword spotting system with the Google Speech Command dataset. The generated datasets are used to train a proposed Convolutional Neural Network (CNN), together with cepstral features, to detect adversarial examples. Experimental results show that, it is possible to accurately distinct between adversarial and normal examples for known attacks, in both single-condition and multi-condition training settings, while the performance degrades dramatically for unknown attacks. The adversarial datasets and the source code are made publicly available.

Matching two sets of items, called set-to-set matching problem, is being recently raised. The difficulties of set-to-set matching over ordinary data matching lie in the exchangeability in 1) set-feature extraction and 2) set-matching score; the pair of sets and the items in each set should be exchangeable. In this paper, we propose a deep learning architecture for the set-to-set matching that overcomes the above difficulties, including two novel modules: 1) a cross-set transformation and 2) cross-similarity function. The former provides the exchangeable set-feature through interactions between two sets in intermediate layers, and the latter provides the exchangeable set matching through calculating the cross-feature similarity of items between two sets. We evaluate the methods through experiments with two industrial applications, fashion set recommendation, and group re-identification. Through these experiments, we show that the proposed methods perform better than a baseline given by an extension of the Set Transformer, the state-of-the-art set-input function.

In federated learning systems, clients are autonomous in that their behaviors are not fully governed by the server. Consequently, a client may intentionally or unintentionally deviate from the prescribed course of federated model training, resulting in abnormal behaviors, such as turning into a malicious attacker or a malfunctioning client. Timely detecting those anomalous clients is therefore critical to minimize their adverse impacts. In this work, we propose to detect anomalous clients at the server side. In particular, we generate low-dimensional surrogates of model weight vectors and use them to perform anomaly detection. We evaluate our solution through experiments on image classification model training over the FEMNIST dataset. Experimental results show that the proposed detection-based approach significantly outperforms the conventional defense-based methods.

The high computational complexity associated with training deep neural networks limits online and real-time training on edge devices. This paper proposed an end-to-end training and inference scheme that eliminates multiplications by approximate operations in the log-domain which has the potential to significantly reduce implementation complexity. We implement the entire training procedure in the log-domain, with fixed-point data representations. This training procedure is inspired by hardware-friendly approximations of log-domain addition which are based on look-up tables and bit-shifts. We show that our 16-bit log-based training can achieve classification accuracy within approximately 1% of the equivalent floating-point baselines for a number of commonly used datasets.

Automatic cover detection -- the task of finding in a audio dataset all covers of a query track -- has long been a challenging theoretical problem in MIR community. It also became a practical need for music composers societies requiring to detect automatically if an audio excerpt embeds musical content belonging to their catalog. In a recent work, we addressed this problem with a convolutional neural network mapping each track's dominant melody to an embedding vector, and trained to minimize cover pairs distance in the embeddings space, while maximizing it for non-covers. We showed in particular that training this model with enough works having five or more covers yields state-of-the-art results. This however does not reflect the realistic use case, where music catalogs typically contain works with zero or at most one or two covers. We thus introduce here a new test set incorporating these constraints, and propose two contributions to improve our model's accuracy under these stricter conditions: we replace dominant melody with multi-pitch representation as input data, and describe a novel prototypical triplet loss designed to improve covers clustering. We show that these changes improve results significantly for two concrete use cases, large dataset lookup and live songs identification.