Deep reinforcement learning techniques have shown to be a promising path to solve very complex tasks that once were thought to be out of the realm of machines. However, while humans and animals learn incrementally during their lifetimes and exploit their experience to solve new tasks, standard deep learning methods specialize to solve only one task at a time and whatever information they acquire is hardly reusable in new situations. Given that any artificial agent would need such a generalization ability to deal with the complexities of the world, it is critical to understand what mechanisms give rise to this ability. We argue that one of the mechanisms humans rely on is the use of discrete conceptual representations to encode their sensory inputs. These representations group similar inputs in such a way that combined they provide a level of abstraction that is transverse to a wide variety of tasks, filtering out irrelevant information for their solution. Here, we show that it is possible to learn such concept-like representations by self-supervision, following an information-bottleneck approach, and that these representations accelerate the transference of skills by providing a prior that guides the policy optimization process. Our method is able to learn useful concepts in locomotive tasks that significantly reduce the number of optimization steps required, opening a new path to endow artificial agents with generalization abilities.

Multi-label sentences (text) in the clinical domain result from the rich description of scenarios during patient care. The state-of-theart methods for assertion detection mostly address this task in the setting of a single assertion label per sentence (text). In addition, few rules based and deep learning methods perform negation/assertion scope detection on single-label text. It is a significant challenge extending these methods to address multi-label sentences without diminishing performance. Therefore, we developed a convolutional neural network (CNN) architecture to localize multiple labels and their scopes in a single stage end-to-end fashion, and demonstrate that our model performs atleast 12% better than the state-of-the-art on multi-label clinical text.

The ability to detect anomalies in time series is considered as highly valuable within plenty of application domains. The sequential nature of time series objects is responsible for an additional feature complexity, ultimately requiring specialized approaches for solving the task. Essential characteristics of time series, laying outside the time domain, are often difficult to capture with state-of-the-art anomaly detection methods, when no transformations on the time series have been applied. Inspired by the success of deep learning methods in computer vision, several studies have proposed to transform time-series into image-like representations, leading to very promising results. However, most of the previous studies implementing time-series to image encodings have focused on the supervised classification. The application to unsupervised anomaly detection tasks has been limited. The paper has the following contributions: First, we evaluate the application of six time-series to image encodings to DL algorithms: Gramian Angular Field, Markov Transition Field, Recurrence Plot, Grey Scale Encoding, Spectrogram and Scalogram. Second, we propose modifications of the original encoding definitions, to make them more robust to the variability in large datasets. And third, we provide a comprehensive comparison between using the raw time series directly and the different encodings, with and without the proposed improvements. The comparison is performed on a dataset collected and released by Airbus, containing highly complex vibration measurements from real helicopters flight tests. The different encodings provide competitive results for anomaly detection.

Knowledge distillation has been widely used to compress existing deep learning models while preserving the performance on a wide range of applications. In the specific context of Automatic Speech Recognition (ASR), distillation from ensembles of acoustic models has recently shown promising results in increasing recognition performance. In this paper, we propose an extension of multi-teacher distillation methods to joint ctc-atention end-to-end ASR systems. We also introduce two novel distillation strategies. The core intuition behind both is to integrate the error rate metric to the teacher selection rather than solely focusing on the observed losses. This way, we directly distillate and optimize the student toward the relevant metric for speech recognition. We evaluated these strategies under a selection of training procedures on the TIMIT phoneme recognition task and observed promising error rate for these strategies compared to a common baseline. Indeed, the best obtained phoneme error rate of 16.4% represents a state-of-the-art score for end-to-end ASR systems.

Neural differential equations are a promising new member in the neural network family. They show the potential of differential equations for time series data analysis. In this paper, the strength of the ordinary differential equation (ODE) is explored with a new extension. The main goal of this work is to answer the following questions: (i)~can ODE be used to redefine the existing neural network model? (ii)~can Neural ODEs solve the irregular sampling rate challenge of existing neural network models for a continuous time series, i.e., length and dynamic nature, (iii)~how to reduce the training and evaluation time of existing Neural ODE systems? This work leverages the mathematical foundation of ODEs to redesign traditional RNNs such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The main contribution of this paper is to illustrate the design of two new ODE-based RNN models (GRU-ODE model and LSTM-ODE) which can compute the hidden state and cell state at any point of time using an ODE solver. These models reduce the computation overhead of hidden state and cell state by a vast amount. The performance evaluation of these two new models for learning continuous time series with irregular sampling rate is then demonstrated. Experiments show that these new ODE based RNN models require less training time than Latent ODEs and conventional Neural ODEs. They can achieve higher accuracy quickly, and the design of the neural network is simpler than, previous neural ODE systems.

Attention architectures are widely used; they recently gained renewed popularity with Transformers yielding a streak of state of the art results. Yet, the geometrical implications of softmax-attention remain largely unexplored. In this work we highlight the limitations of constraining attention weights to the probability simplex and the resulting convex hull of value vectors. We show that Transformers are sequence length dependent biased towards token isolation at initialization and contrast Transformers to simple max- and sum-pooling - two strong baselines rarely reported. We propose to replace the softmax in self-attention with normalization, yielding a hyperparameter and data-bias robust, generally applicable architecture. We support our insights with empirical results from more than 25,000 trained models. All results and implementations are made available.

Motivated by the attention mechanism of the human visual system and recent developments in the field of machine translation, we introduce our attention-based and recurrent sequence to sequence autoencoders for fully unsupervised representation learning from audio files. In particular, we test the efficacy of our novel approach on the task of speech-based sleepiness recognition. We evaluate the learnt representations from both autoencoders, and then conduct an early fusion to ascertain possible complementarity between them. In our frameworks, we first extract Mel-spectrograms from raw audio files. Second, we train recurrent autoencoders on these spectrograms which are considered as time-dependent frequency vectors. Afterwards, we extract the activations of specific fully connected layers of the autoencoders which represent the learnt features of spectrograms for the corresponding audio instances. Finally, we train support vector regressors on these representations to obtain the predictions. On the development partition of the data, we achieve Spearman's correlation coefficients of .324, .283, and .320 with the targets on the Karolinska Sleepiness Scale by utilising attention and non-attention autoencoders, and the fusion of both autoencoders' representations, respectively. In the same order, we achieve .311, .359, and .367 Spearman's correlation coefficients on the test data, indicating the suitability of our proposed fusion strategy.

Despite the significant progress in automatic speech recognition (ASR), distant ASR remains challenging due to noise and reverberation. A common approach to mitigate this issue consists of equipping the recording devices with multiple microphones that capture the acoustic scene from different perspectives. These multi-channel audio recordings contain specific internal relations between each signal. In this paper, we propose to capture these inter- and intra- structural dependencies with quaternion neural networks, which can jointly process multiple signals as whole quaternion entities. The quaternion algebra replaces the standard dot product with the Hamilton one, thus offering a simple and elegant way to model dependencies between elements. The quaternion layers are then coupled with a recurrent neural network, which can learn long-term dependencies in the time domain. We show that a quaternion long-short term memory neural network (QLSTM), trained on the concatenated multi-channel speech signals, outperforms equivalent real-valued LSTM on two different tasks of multi-channel distant speech recognition.

Simplicity is the ultimate sophistication. Differentiable Architecture Search (DARTS) has now become one of the mainstream paradigms of neural architecture search. However, it largely suffers from several disturbing factors of the optimization process whose results are unstable to reproduce. Chu et al. [2019a] point out that skip connections natively have an unfair advantage in exclusive competition which primarily leads to dramatic performance collapse. While Fair-DARTS (Chu et al. [2019a]) turns the unfair competition into a collaborative one, we instead impede such unfair advantage by injecting unbiased random noise into skip operations' output. In effect, the optimizer should perceive this difficulty at each training step and refrain from overshooting on skip connections, but in a long run, it still converges to the right solution area since no bias is added to the gradient in terms of expectation. We name this novel approach as NoisyDARTS. Our experiments on CIFAR-10 and ImageNet attest that it can effectively break the skip connections' unfair advantage and yield better performance. It generates a series of models that achieve state-of-the-art results on both datasets.

Their importance has been highlighted in recent times, when ICUs around the world have been overrun by the COVID-19 pandemic [1, 2]. It is in times like these when research into ways to adequately manage scarce critical care resources must be even more vigorously pursued, in order to offer additional tools that support medical decisions and allow for the effective benchmark of clinical practice. The issue of mortality prediction in the ICU has been approached from a statistical standpoint by means of risk prediction models like APACHE, SAPS, MODS, among others [3]. These models use a set of physiological predictors, demographic factors, and the occurrence of certain chronic conditions, to estimate a score that serves as a proxy for the likelihood of death of ICU patients. Because of the relatively straightforward way of interpreting results, simple statistical approaches such as logistic regression are the go-to modeling techniques used to estimate mortality probability and the importance of the predictors involved. On the other hand, the simplicity of the models also mean that their limited expressiveness may not accurately represent the possibly nonlinear dynamics of mortality prediction. Given this, high-capacity machine learning models might be useful to increase predictive performance.