John J. Prevost

Michael G. DeMoor, John J. Prevost

Deep Metric Learning (DML) loss functions traditionally aim to control the forces of separability and compactness within an embedding space so that the same class data points are pulled together and different class ones are pushed apart. Within the context of DML, a softmax operation will typically normalize distances into a probability for optimization, thus coupling all the push/pull forces together. This paper proposes a potential new class of loss functions that operate within a euclidean domain and aim to take full advantage of the coupled forces governing embedding space formation under a softmax. These forces of compactness and separability can be boosted or mitigated within controlled locations at will by using a warping function. In this work, we provide a simple example of a warping function and use it to achieve competitive, state-of-the-art results on various metric learning benchmarks.

Edmond Adib, Fatemeh Afghah, John J. Prevost

ECG databases are usually highly imbalanced due to the abundance of Normal ECG and scarcity of abnormal cases. As such, deep learning classifiers trained on imbalanced datasets usually perform poorly, especially on minor classes. One solution is to generate realistic synthetic ECG signals using Generative Adversarial Networks (GAN) to augment imbalanced datasets. In this study, we combined conditional GAN with WGAN-GP and developed AC-WGAN-GP in 1D form for the first time to be applied on MIT-BIH Arrhythmia dataset. We investigated the impact of data augmentation on arrhythmia classification. We employed two models for ECG generation: (i) unconditional GAN; Wasserstein GAN with gradient penalty (WGAN-GP) is trained on each class individually; (ii) conditional GAN; one Auxiliary Classifier WGAN-GP (AC-WGAN-GP) model is trained on all classes and then used to generate synthetic beats in all classes. Two scenarios are defined for each case: (a) unscreened; all the generated synthetic beats were used, and (b) screened; only a portion of generated beats are selected and used, based on their Dynamic Time Warping (DTW) distance to a designated template. A state-of-the-art ResNet classifier (EcgResNet34) is trained on each of the augmented datasets and the performance metrics (precision/recall/F1-Score micro- and macro-averaged, confusion matrices, multiclass precision-recall curves) were compared with those of the unaugmented imbalanced case. We also used a simple metric Net Improvement. All the three metrics show consistently that net improvement (total and minor-class), unconditional GAN with raw generated data (not screened) creates the best improvements.

Edmond Adib, Fatemeh Afghah, John J. Prevost

Electrocardiogram (ECG) datasets tend to be highly imbalanced due to the scarcity of abnormal cases. Additionally, the use of real patients' ECGs is highly regulated due to privacy issues. Therefore, there is always a need for more ECG data, especially for the training of automatic diagnosis machine learning models, which perform better when trained on a balanced dataset. We studied the synthetic ECG generation capability of 5 different models from the generative adversarial network (GAN) family and compared their performances, the focus being only on Normal cardiac cycles. Dynamic Time Warping (DTW), Fréchet, and Euclidean distance functions were employed to quantitatively measure performance. Five different methods for evaluating generated beats were proposed and applied. We also proposed 3 new concepts (threshold, accepted beat and productivity rate) and employed them along with the aforementioned methods as a systematic way for comparison between models. The results show that all the tested models can, to an extent, successfully mass-generate acceptable heartbeats with high similarity in morphological features, and potentially all of them can be used to augment imbalanced datasets. However, visual inspections of generated beats favors BiLSTM-DC GAN and WGAN, as they produce statistically more acceptable beats. Also, with regards to productivity rate, the Classic GAN is superior with a 72% productivity rate. We also designed a simple experiment with the state-of-the-art classifier (ECGResNet34) to show empirically that the augmentation of the imbalanced dataset by synthetic ECG signals could improve the performance of classification significantly.

Sanchita Ghose, John J. Prevost

Deep learning based visual to sound generation systems essentially need to be developed particularly considering the synchronicity aspects of visual and audio features with time. In this research we introduce a novel task of guiding a class conditioned generative adversarial network with the temporal visual information of a video input for visual to sound generation task adapting the synchronicity traits between audio-visual modalities. Our proposed FoleyGAN model is capable of conditioning action sequences of visual events leading towards generating visually aligned realistic sound tracks. We expand our previously proposed Automatic Foley dataset to train with FoleyGAN and evaluate our synthesized sound through human survey that shows noteworthy (on average 81\%) audio-visual synchronicity performance. Our approach also outperforms in statistical experiments compared with other baseline models and audio-visual datasets.

Sanchita Ghose, John J. Prevost

In movie productions, the Foley Artist is responsible for creating an overlay soundtrack that helps the movie come alive for the audience. This requires the artist to first identify the sounds that will enhance the experience for the listener thereby reinforcing the Directors's intention for a given scene. In this paper, we present AutoFoley, a fully-automated deep learning tool that can be used to synthesize a representative audio track for videos. AutoFoley can be used in the applications where there is either no corresponding audio file associated with the video or in cases where there is a need to identify critical scenarios and provide a synthesized, reinforced soundtrack. An important performance criterion of the synthesized soundtrack is to be time-synchronized with the input video, which provides for a realistic and believable portrayal of the synthesized sound. Unlike existing sound prediction and generation architectures, our algorithm is capable of precise recognition of actions as well as inter-frame relations in fast moving video clips by incorporating an interpolation technique and Temporal Relationship Networks (TRN). We employ a robust multi-scale Recurrent Neural Network (RNN) associated with a Convolutional Neural Network (CNN) for a better understanding of the intricate input-to-output associations over time. To evaluate AutoFoley, we create and introduce a large scale audio-video dataset containing a variety of sounds frequently used as Foley effects in movies. Our experiments show that the synthesized sounds are realistically portrayed with accurate temporal synchronization of the associated visual inputs. Human qualitative testing of AutoFoley show over 73% of the test subjects considered the generated soundtrack as original, which is a noteworthy improvement in cross-modal research in sound synthesis.