Impact Of Semantics, Physics And Adversarial Mechanisms In Deep Learning

Loading...
Thumbnail Image

Publication or External Link

Date

2020

Citation

Abstract

Deep learning has greatly advanced the performance of algorithms on tasks such as image classification, speech enhancement, sound separation, and generative image models. However many current popular systems are driven by empirical rules that do not fully exploit the underlying physics of the data. Many speech and audio systems fix STFT preprocessing before their networks. Hyperspectral Image (HSI) methods often don't deliberately consider the spectral spatial trade off that is not present in normal images. Generative Adversarial Networks (GANs) that learn a generative distribution of images don't prioritize semantic labels of the training data. To meet these opportunities we propose to alter known deep learning methods to be more dependent on the semantic and physical underpinnings of the data to create better performing and more robust algorithms for sound separation and classification, image generation, and HSI segmentation. Our approaches take inspiration from from Harmonic Analysis, SVMs, and classical statistical detection theory, and further the state-of-the art in source separation, defense against audio adversarial attacks, HSI classification, and GANs.

Recent deep learning approaches have achieved impressive performance on speech enhancement and separation tasks. However, these approaches have not been investigated for separating mixtures of arbitrary sounds of different types, a task we refer to as universal sound separation. To study this question, we develop a dataset of mixtures containing arbitrary sounds, and use it to investigate the space of mask-based separation architectures, varying both the overall network architecture and the framewise analysis-synthesis basis for signal transformations. We compare using a short-time Fourier transform (STFT) with a learnable basis at variable window sizes for the feature extraction stage of our sound separation network. We also compare the robustness to adversarial examples of speech classification networks that similarly hybridize established Time-frequency (TF) methods with learnable filter weights.

We analyze HSI images for material classification. For hyperspectral image cubes TF methods decompose spectra into multi-spectral bands, while Neural Networks (NNs) incorporate spatial information across scales and model multiple levels of dependencies between spectral features. The Fourier scattering transform is an amalgamation of time-frequency representations with neural network architectures. We propose and test a three dimensional Fourier scattering method on hyperspectral datasets, and present results that indicate that the Fourier scattering transform is highly effective at representing spectral data when compared with other state-of-the-art methods. We study the spectral-spatial trade-off that our Scattering approach allows.We also use a similar multi-scale approach to develop a defense against audio adversarial attacks. We propose a unification of a computational model of speech processing in the brain with commercial wake-word networks to create a cortical network, and show that it can increase resistance to adversarial noise without a degradation in performance.

Generative Adversarial Networks are an attractive approach to constructing generative models that mimic a target distribution, and typically use conditional information (cGANs) such as class labels to guide the training of the discriminator and the generator. We propose a loss that ensures generator updates are always class specific, rather than training a function that measures the information theoretic distance between the generative distribution and one target distribution, we generalize the successful hinge-loss that has become an essential ingredient of many GANs to the multi-class setting and use it to train a single generator classifier pair. While the canonical hinge loss made generator updates according to a class agnostic margin a real/fake discriminator learned, our multi-class hinge-loss GAN updates the generator according to many classification margins. With this modification, we are able to accelerate training and achieve state of the art Inception and FID scores on Imagenet128. We study the trade-off between class fidelity and overall diversity of generated images, and show modifications of our method can prioritize either each during training. We show that there is a limit to how closely classification and discrimination can be combined while maintaining sample diversity with some theoretical results on K+1 GANs.

Notes

Rights