No description available.

Early time series classification is the task of predicting the class label of a time series before it is observed in its entirety. In time-sensitive domains where information is collected over time it is worth sacrificing some classification accuracy in favor of earlier predictions, ideally early enough for actions to be taken. However, since accuracy and earliness are contradictory objectives, a solution to this problem must find a task-dependent trade-off. There are two common state-of-the-art methods. The first involves an analyst selecting a timestep at which all predictions must be made. This does not capture earliness on a case-by-case basis, so if the selecting timestep is too early, all later signals are missed, and if a signal happens early, the classifier still waits to generate a prediction. The second method is the exhaustive search for signals, which encodes no timing information and is not scalable to high dimensions or long time series. We design the first early classification model called EARLIEST to tackle this multi-objective optimization problem, jointly learning (1) to decide at which time step to halt and generate predictions and (2) how to classify the time series. Each of these is learned based on the task and data features. We achieve an analyst-controlled balance between the goals of earliness and accuracy by pairing a recurrent neural network that learns to classify time series as a supervised learning task with a stochastic controller network that learns a halting-policy as a reinforcement learning task. The halting-policy dictates sequential decisions, one per timestep, of whether or not to halt the recurrent neural network and classify the time series early. This pairing of networks optimizes a global objective function that incorporates both earliness and accuracy. We validate our method via critical clinical prediction tasks in the MIMIC III database from the Beth Israel Deaconess Medical Center along with another publicly available time series classification dataset. We show that EARLIEST out-performs two state-of-the-art LSTM-based early classification methods. Additionally, we dig deeper into our model's performance using a synthetic dataset which shows that EARLIEST learns to halt when it observes signals without having explicit access to signal locations. The contributions of this work are three-fold. First, our method is the first neural network-based solution to early classification of time series, bringing the recent successes of deep learning to this problem. Second, we present the first reinforcement-learning based solution to the unsupervised nature of early classification, learning the underlying distributions of signals without access to this information through trial and error. Third, we propose the first joint-optimization of earliness and accuracy, allowing learning of complex relationships between these contradictory goals.

Robotics agents are meant to play an increasingly larger role in our everyday lives. To be successfully integrated in our environment, robots will need to develop and display adaptive, robust, and socially suitable behaviours. To tackle these issues, the robotics research community has invested a considerable amount of efforts in modelling robotic architectures inspired by research on living systems, from ethology to developmental psychology. Following a similar approach, this thesis presents the research results of the modelling and experimental testing of robotic architectures based on affective and attachment bonds between young infants and their primary caregiver. I follow a bottom-up approach to the modelling of such bonds, examining how they can promote the situated development of an autonomous robot. Specifically, the models used and the results from the experiments carried out in laboratory settings and with naive users demonstrate the impact such affective bonds have on the learning outcomes of an autonomous robot and on the perception and behaviour of humans. This research leads to the emphasis on the importance of the interplay between the dynamics of the regulatory behaviours performed by a robot and the responsiveness of the human partner. The coupling of such signals and behaviours in an attachment-like dyad determines the nature of the outcomes for the robot, in terms of learning or the satisfaction of other needs. The experiments carried out also demonstrate of the attachment system can help a robot adapt its own social behaviour to that of the human partners, as infants are thought to do during their development.

Deep neural network (DNN) hardware accelerators are critical enablers of the current resurgence in machine learning technologies. Adopting machine learning in safety-critical systems imposes additional reliability requirements on hardware design. Addressing these requirements mandates an accurate assessment of the impact caused by permanent faults in the processing engines (PE). Carrying out this reliability assessment early in the design process allows for addressing potential reliability concerns when it is less costly to perform design revisions. However, the large size of modern DNN hardware and the complexity of the DNN applications running on it present barriers to efficient reliability evaluation before proceeding with the design implementation. Considering these barriers, this dissertation proposes two methodologies to assess fault resiliency in integer arithmetic units in DNN hardware. Using the information from the data streaming patterns of the DNN accelerators, which are known before the register-transfer level (RTL) implementation, the first methodology enables fault injection experiments to be carried out in PE units at the pre-RTL stage during architectural design space exploration. This is achieved in a DNN simulation framework that captures the mapping between a model's operations and the hardware's arithmetic units. This facilitates a fault resiliency comparison of state-of-the-art DNN accelerators comprising thousands of PE units. The second methodology introduces accurate and efficient modelling of the impact of permanent faults in integer multipliers. It avoids the need for computationally intensive circuit models, e.g., netlists, to inject faults in integer arithmetic units, thus scaling the fault resiliency assessment to accelerators with thousands of PE units with negligible simulation time overhead. As a first step, we formally analyze the impact of permanent faults affecting the internal nodes of two integer multiplier architectures. This analysis indicates that, for most internal faults, the impact on the output is independent of the operands involved in the arithmetic operation. As the second step, we develop a statistical fault injection approach based on the likelihood of a fault being triggered in the applications that run on the target DNN hardware. By modelling the impact of faults in internal nodes of arithmetic units using fault-free operations, fault injection campaigns run three orders of magnitude faster than using arithmetic circuit models in the same simulation environment. The experiments also show that the proposed method's accuracy is on par with that of using netlists to model arithmetic circuitry in which faults are injected. Using the proposed methods, one can conduct fault assessment experiments for various DNN models and hardware architectures, examining the sensitivity of DNN model-related and hardware architecture-related features on the DNN accelerator's reliability. In addition to understanding the impact of permanent hardware faults on the accuracy of DNN models running on defective hardware, the outcomes of these experiments can yield valuable insights for designers seeking to balance fault criticality and performance, thereby facilitating the development of more reliable DNN hardware in the future. / Thesis / Doctor of Philosophy (PhD) / The reliability of Deep Neural Network (DNN) hardware has become critical in recent years, especially for the adoption of machine learning in safety-critical applications. Evaluating the reliability of DNN hardware early in the design process enables addressing potential reliability concerns before committing to full implementation. However, the large size and complexity of DNN hardware impose challenges in evaluating its reliability in an efficient manner. In this dissertation, two novel methodologies are proposed to address these challenges. The first methodology introduces an efficient method to describe the mapping of operations of DNN applications to the processing engines of a target DNN hardware architecture in a high-performance computing DNN simulation environment. This approach allows for assessing the fault resiliency of large hardware architectures, incorporating thousands of processing engines while using fewer simulation resources compared to existing methods. The second methodology introduces an accurate and efficient approach to modelling the impact of permanent faults in integer arithmetic units of DNN hardware during inference. By leveraging the special characteristics of integer arithmetic units, this method achieves fault assessment at negligible computational overhead relative to running DNN inference in the fault-free mode in state-of-the-art DNN frameworks.

In this thesis I attempt to explain mechanisms of neuronal coding in the auditory system as a form of adaptation to statistics of natural stereo sounds. To this end I analyse recordings of real-world auditory environments and construct novel statistical models of these data. I further compare regularities present in natural stimuli with known, experimentally observed neuronal mechanisms of spatial hearing. In a more general perspective, I use binaural auditory system as a starting point to consider the notion of function implemented by sensory neurons. In particular I argue for two, closely-related tenets: 1. The function of sensory neurons can not be fully elucidated without understanding statistics of natural stimuli they process. 2. Function of sensory representations is determined by redundancies present in the natural sensory environment. I present the evidence in support of the first tenet by describing and analysing marginal statistics of natural binaural sound. I compare observed, empirical distributions with knowledge from reductionist experiments. Such comparison allows to argue that the complexity of the spatial hearing task in the natural environment is much higher than analytic, physics-based predictions. I discuss the possibility that early brain stem circuits such as LSO and MSO do not \"compute sound localization\" as is often being claimed in the experimental literature. I propose that instead they perform a signal transformation, which constitutes the first step of a complex inference process. To support the second tenet I develop a hierarchical statistical model, which learns a joint sparse representation of amplitude and phase information from natural stereo sounds. I demonstrate that learned higher order features reproduce properties of auditory cortical neurons, when probed with spatial sounds. Reproduced aspects were hypothesized to be a manifestation of a fine-tuned computation specific to the sound-localization task. Here it is demonstrated that they rather reflect redundancies present in the natural stimulus. Taken together, results presented in this thesis suggest that efficient coding is a strategy useful for discovering structures (redundancies) in the input data. Their meaning has to be determined by the organism via environmental feedback.

Deep learning has in recent years come to dominate the previously separate fields of research in machine learning, computer vision, natural language understanding and speech recognition. Despite breakthroughs in training deep networks, there remains a lack of understanding of both the optimization and structure of deep networks. The approach advocated by many researchers in the field has been to train monolithic networks with excess complexity, and strong regularization --- an approach that leaves much to desire in efficiency. Instead we propose that carefully designing networks in consideration of our prior knowledge of the task and learned representation can improve the memory and compute efficiency of state-of-the art networks, and even improve generalization --- what we propose to denote as structural priors. We present two such novel structural priors for convolutional neural networks, and evaluate them in state-of-the-art image classification CNN architectures. The first of these methods proposes to exploit our knowledge of the low-rank nature of most filters learned for natural images by structuring a deep network to learn a collection of mostly small, low-rank, filters. The second addresses the filter/channel extents of convolutional filters, by learning filters with limited channel extents. The size of these channel-wise basis filters increases with the depth of the model, giving a novel sparse connection structure that resembles a tree root. Both methods are found to improve the generalization of these architectures while also decreasing the size and increasing the efficiency of their training and test-time computation. Finally, we present work towards conditional computation in deep neural networks, moving towards a method of automatically learning structural priors in deep networks. We propose a new discriminative learning model, conditional networks, that jointly exploit the accurate representation learning capabilities of deep neural networks with the efficient conditional computation of decision trees. Conditional networks yield smaller models, and offer test-time flexibility in the trade-off of computation vs. accuracy.

This work deals with automatic chord recognition using neural networks. The problem was separated into two subproblems. The first subproblem aims to experimental finding of most suitable solution for a acoustic model and the second one aims to experimental finding of most suitable solution for a language model. The problem was solved by iterative method. First a suboptimal solution of the first subproblem was found and then the second one. A total of 19 acoustic and 12 language models were made. Ten training datasets was created for acoustic models and three for language models. In total, over 200 models were trained. The best results were achieved on acoustic models represented by convolutional networks together with language models represented by recurent networks with LSTM modules.