We present a test platform for visual in-cabin scene analysis and occupant monitoring functions. The test platform is based on a driving simulator developed at the DFKI, consisting of a realistic in-cabin mock-up and a wide-angle projection system for a realistic driving experience. The platform has been equipped with a wide-angle 2D/3D camera system monitoring the entire interior of the vehicle mock-up of the simulator. It is also supplemented with a ground truth reference sensor system that allows to track and record the occupant's body movements synchronously with the 2D and 3D video streams of the camera. Thus, the resulting test platform will serve as a basis to validate numerous incabin monitoring functions, which are important for the realization of novel human-vehicle interfaces, advanced driver assistant systems, and automated driving. Among the considered functions are occupant presence detection, size and 3D-pose estimation, and driver intention recognition. In addition, our platform will be the basis for the creation of large-scale in-cabin benchmark datasets.

While neural networks have achieved impressive successes in domains like image classification or machine translation, standard models still struggle with tasks that require very longterm memory without interference and would thus benefit from a separation of memory and computation [Graves et al., 2016, Collier and Beel, 2018]. Neural Turing Machines (NTM) attempt to address these tasks by augmenting recurrent neural networks with an explicit memory to which the network has read and write access [Graves et al., 2016, Collier and Beel, 2018]. Unfortunately, such models are notoriously hard to train, even compared to other deep learning models [Collier and Beel, 2018]. In our contribution, we propose to address this training problem by replacing the learned recurrent neural network controller of a NTM with an echo state network (ESN) [Jaeger and Haas, 2004]. In other words, we only learn the controller for the read and write head of our memory access as well as the output mapping, all of which is possible via standard linear regression.

We introduce a data-driven approach to automatic pitch correction of solo singing performances. The proposed approach predicts note-wise pitch shifts from the relationship between the respective spectrograms of the singing and accompaniment. This approach differs from commercial systems, where vocal track notes are usually shifted to be centered around pitches in a user-defined score, or mapped to the closest pitch among the twelve equal-tempered scale degrees. The proposed system treats pitch as a continuous value rather than relying on a set of discretized notes found in musical scores, thus allowing for improvisation and harmonization in the singing performance. We train our neural network model using a dataset of 4,702 amateur karaoke performances selected for good intonation. Our model is trained on both incorrect intonation, for which it learns a correction, and intentional pitch variation, which it learns to preserve. The proposed deep neural network with gated recurrent units on top of convolutional layers shows promising performance on the real-world score-free singing pitch correction task of autotuning.

Existing graph convolutional networks focus on the neighborhood aggregation scheme. When applied to semi-supervised learning, they often suffer from the overfitting problem as the networks are trained with the cross-entropy loss on a small potion of labeled data. In this paper, we propose an unsupervised manifold smoothness loss defined with respect to the graph structure, which can be added to the loss function as a regularization. We draw connections between the proposed loss with an iterative diffusion process, and show that minimizing the loss is equivalent to aggregate neighbor predictions with infinite layers. We conduct experiments on multi-layer perceptron and existing graph networks, and demonstrate that adding the proposed loss can improve the performance consistently.

Machine learning (ML) methods have the potential to automate clinical EEG analysis. They can be categorized into feature-based (with handcrafted features), and end-to-end approaches (with learned features). Previous studies on EEG pathology decoding have typically analyzed a limited number of features, decoders, or both. For a I) more elaborate feature-based EEG analysis, and II) in-depth comparisons of both approaches, here we first develop a comprehensive feature-based framework, and then compare this framework to state-of-the-art end-to-end methods. To this aim, we apply the proposed feature-based framework and deep neural networks including an EEG-optimized temporal convolutional network (TCN) to the task of pathological versus non-pathological EEG classification. For a robust comparison, we chose the Temple University Hospital (TUH) Abnormal EEG Corpus (v2.0.0), which contains approximately 3000 EEG recordings. The results demonstrate that the proposed feature-based decoding framework can achieve accuracies on the same level as state-of-the-art deep neural networks. We find accuracies across both approaches in an astonishingly narrow range from 81--86\%. Moreover, visualizations and analyses indicated that both approaches used similar aspects of the data, e.g., delta and theta band power at temporal electrode locations. We argue that the accuracies of current binary EEG pathology decoders could saturate near 90\% due to the imperfect inter-rater agreement of the clinical labels, and that such decoders are already clinically useful, such as in areas where clinical EEG experts are rare. We make the proposed feature-based framework available open source and thus offer a new tool for EEG machine learning research.

Deep neural networks, while generalize well, are known to be sensitive to small adversarial perturbations. This phenomenon poses severe security threat and calls for in-depth investigation of the robustness of deep learning models. With the emergence of neural networks for graph structured data, similar investigations are urged to understand their robustness. It has been found that adversarially perturbing the graph structure and/or node features may result in a significant degradation of the model performance. In this work, we show from a different angle that such fragility similarly occurs if the graph contains a few bad-actor nodes, which compromise a trained graph neural network through flipping the connections to any targeted victim. Worse, the bad actors found for one graph model severely compromise other models as well. We call the bad actors "anchor nodes" and propose an algorithm, named GUA, to identify them. Thorough empirical investigations suggest an interesting finding that the anchor nodes often belong to the same class; and they also corroborate the intuitive trade-off between the number of anchor nodes and the attack success rate. For the data set Cora which contains 2708 nodes, as few as six anchor nodes will result in an attack success rate higher than 80\% for GCN and other three models.

Deep learning is increasingly common in healthcare, yet transfer learning for physiological signals (e.g., temperature, heart rate, etc.) is under-explored. Here, we present a straightforward, yet performant framework for transferring knowledge about physiological signals. Our framework is called PHASE (PHysiologicAl Signal Embeddings). It i) learns deep embeddings of physiological signals and ii) predicts adverse outcomes based on the embeddings. PHASE is the first instance of deep transfer learning in a cross-hospital, cross-department setting for physiological signals. We show that PHASE's per-signal (one for each signal) LSTM embedding functions confer a number of benefits including improved performance, successful transference between hospitals, and lower computational cost.

While current machine learning models have impressive performance over a wide range of applications, their large size and complexity render them unsuitable for tasks such as remote monitoring on edge devices with limited storage and computational power. A naive approach to resolve this on the model level is to use simpler architectures, but this sacrifices prediction accuracy and is unsuitable for monitoring applications requiring accurate detection of the onset of adverse events. In this paper, we propose an alternative solution to this problem by decomposing the predictive model as the sum of a simple function which serves as a local monitoring tool, and a complex correction term to be evaluated on the server. A sign requirement is imposed on the latter to ensure that the local monitoring function is safe, in the sense that it can effectively serve as an early warning system. Our analysis quantifies the trade-offs between model complexity and performance, and serves as a guidance for architecture design. We validate our proposed framework on a series of monitoring experiments, where we succeed at learning monitoring models with significantly reduced complexity that minimally violate the safety requirement. More broadly, our framework is useful for learning classifiers in applications where false negatives are significantly more costly compared to false positives.

The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.

Local robustness ensures that a model classifies all inputs within an $\epsilon$-ball consistently, which precludes various forms of adversarial inputs. In this paper, we present a fast procedure for checking local robustness in feed-forward neural networks with piecewise linear activation functions. The key insight is that such networks partition the input space into a polyhedral complex such that the network is linear inside each polyhedral region; hence, a systematic search for decision boundaries within the regions around a given input is sufficient for assessing robustness. Crucially, we show how these regions can be analyzed using geometric projections instead of expensive constraint solving, thus admitting an efficient, highly-parallel GPU implementation at the price of incompleteness, which can be addressed by falling back on prior approaches. Empirically, we find that incompleteness is not often an issue, and that our method performs one to two orders of magnitude faster than existing robustness-certification techniques based on constraint solving.