Accidental falls are the most frequent injury of old age and have dramatic implications on the individual, family, and the society as a whole. To date, fall prediction estimation is clinical, relying on the expertise ...
详细信息
ISBN:
(纸本)9781728193601
Accidental falls are the most frequent injury of old age and have dramatic implications on the individual, family, and the society as a whole. To date, fall prediction estimation is clinical, relying on the expertise of the physiotherapist for performing the diagnosis based on standard scales, such as the highly common and validated Berg Balance Scale (BBS). Unfortunately, the BBS is a time consuming subjective score, prone to variability and inconsistency between examiners. In this study, we developed an objective, computational tool, which automates the BBS fall assessment process and allows easy, efficient and accessible assessment of fall risk. The tool is based on a novel multi depth-camera human motion tracking system integrated with Machine learning algorithms. The system enables large scale screening of the general public at very little cost while significantly reducing physiotherapist resources. The system was pilot tested in the physiotherapy unit at a major hospital and showed high rates of fall risk predictions as well as correlation with physiotherapists BBS scores on individual BBS motion tasks.
In this paper we deal with a network of computing agents with local processing and neighboring communication capabilities that aim at solving (without any central unit) a submodular optimization problem. The cost func...
详细信息
In this paper we deal with a network of computing agents with local processing and neighboring communication capabilities that aim at solving (without any central unit) a submodular optimization problem. The cost function is the sum of many local submodular functions and each agent in the network has access to one function in the sum only. In this distributed set-up, in order to preserve their own privacy, agents communicate with neighbors but do not share their local cost functions. We propose a distributed algorithm in which agents resort to the Lovasz extension of their local submodular functions and perform local updates and communications in terms of single blocks of the entire optimization variable. Updates are performed by means of a greedy algorithm which is run only until the selected block is computed, thus resulting in a reduced computational burden. The proposed algorithm is shown to converge in expected value to the optimal cost of the problem, and an approximate solution to the submodular problem is retrieved by a thresholding operation. As an application, we consider a distributed image segmentation problem in which each agent has access only to a portion of the entire image. While agents cannot segment the entire image on their own, they correctly complete the task by cooperating through the proposed distributed algorithm. Copyright (C) 2020 The Authors.
A contextual bandit problem is studied in a highly non-stationary environment, which is ubiquitous in various recommender systems due to the time-varying interests of users. Two models with disjoint and hybrid payoffs...
详细信息
ISBN:
(纸本)9781577358350
A contextual bandit problem is studied in a highly non-stationary environment, which is ubiquitous in various recommender systems due to the time-varying interests of users. Two models with disjoint and hybrid payoffs are considered to characterize the phenomenon that users' preferences towards different items vary differently over time. In the disjoint payoff model, the reward of playing an arm is determined by an arm-specific preference vector, which is piecewise-stationary with asynchronous and distinct changes across different arms. An efficient learning algorithm that is adaptive to abrupt reward changes is proposed and theoretical regret analysis is provided to show that a sublinear scaling of regret in the time length T is achieved. The algorithm is further extended to a more general setting with hybrid payoffs where the reward of playing an arm is determined by both an arm-specific preference vector and a joint coefficient vector shared by all arms. Empirical experiments are conducted on real-world datasets to verify the advantages of the proposed learning algorithms against baseline ones in both settings.
Artificial neural networks are well-known to be susceptible to catastrophic forgetting when continually learning from sequences of tasks. Various continual (or "incremental") learning approaches have been pr...
详细信息
ISBN:
(纸本)9781728169262
Artificial neural networks are well-known to be susceptible to catastrophic forgetting when continually learning from sequences of tasks. Various continual (or "incremental") learning approaches have been proposed to avoid catastrophic forgetting, but they are typically adversary agnostic, i.e., they do not consider the possibility of a malicious attack. In this effort, we explore the vulnerability of Elastic Weight Consolidation (EWC), a popular continual learning algorithm for avoiding catastrophic forgetting. We show that an intelligent adversary can take advantage of EWC's continual learning capabilities to cause gradual and deliberate forgetting by introducing small amounts of mis-information to the model during training. We demonstrate such an adversary's ability to assume control of the model via injection of backdoor attack samples on both permuted and split benchmark variants of the MNIST dataset. Importantly, once the model has learned the adversarial misinformation, the adversary can then control the amount of forgetting of any task. Equivalently, the malicious actor can create a "false memory" about any task by inserting carefully-designed backdoor samples to any fraction of the test instances of that task. Perhaps most damaging, we show this vulnerability to be very acute;the model memory can be easily compromised with the addition of backdoor samples into as little as 1% of the training data of even a single task.
The World Health Organization estimates that 15 million infants are born preterm every year [1]. This is of concern because these infants have a significant chance of having neuromotor or cognitive developmental delay...
详细信息
ISBN:
(纸本)9781728159072
The World Health Organization estimates that 15 million infants are born preterm every year [1]. This is of concern because these infants have a significant chance of having neuromotor or cognitive developmental delays due to cerebral palsy or other developmental issues [2]. Our long-term goal is to determine the roles emotion and movement play in the diagnosis of atypical infants. In this paper, we examine how automated emotion assessment may have potential to classify typically and atypically developing infants. We compare a custom supervised machine learning algorithm that utilizes individual and grouped facial features for infant emotion classification with a state-of-the-art neural network. Our results show that only three concavity features are needed for the concavity algorithm, and the custom algorithm performed with relatively similar performance to the neural network. Automatic sentiment labels used in tandem with infant movement kinematics would be further investigated to determine if emotion and movement are interdependent and predictive of an infant's neurodevelopmental delay in disorders such as cerebral palsy.
A database-driven PID (DD-PID) control method is one of the effective control methods for nonlinear systems. In the conventional DD-PID control method, there is a problem that the calculation cost and required memory ...
详细信息
A database-driven PID (DD-PID) control method is one of the effective control methods for nonlinear systems. In the conventional DD-PID control method, there is a problem that the calculation cost and required memory for creating an optimal database are large. For the above problem, this paper proposes a method to implement the DD-PID controller with small-sized sub-databases. In the proposed method, one database that includes past I/O data and PID gains are created, and the database is updated in an offline manner. Moreover, sub-databases are constructed by clustering the created database using the k-means clustering method. The number of clusters for k-means clustering is determined automatically based on kernel functions. The effectiveness of the proposed method is presented by numerical examples. Copyright (C) 2020 The Authors.
We propose a new symbolic trace semantics for register automata (extended finite state machines) which records both the sequence of input symbols that occur during a run as well as the constraints on input parameters ...
详细信息
ISBN:
(纸本)9783030642761;9783030642754
We propose a new symbolic trace semantics for register automata (extended finite state machines) which records both the sequence of input symbols that occur during a run as well as the constraints on input parameters that are imposed by this run. Our main result is a generalization of the classical Myhill-Nerode theorem to this symbolic setting. Our generalization requires the use of three relations to capture the additional structure of register automata. Location equivalence (l) captures that symbolic traces end in the same location, transition equivalence (t )captures that they share the same final transition, and a partial equivalence relation (r )captures that symbolic values v and v' are stored in the same register after symbolic traces w and w', respectively. A symbolic language is defined to be regular if relations (l,)(t )and (r )exist that satisfy certain conditions, in particular, they all have finite index. We show that the symbolic language associated to a register automaton is regular, and we construct, for each regular symbolic language, a register automaton that accepts this language. Our result provides a foundation for grey-box learning algorithms in settings where the constraints on data parameters can be extracted from code using e.g. tools for symbolic/concolic execution or tainting. We believe that moving to a grey-box setting is essential to overcome the scalability problems of state-of-the-art black-box learning algorithms.
In many fields such as digital marketing, healthcare, finance, and robotics, it is common to have a well-tested and reliable baseline policy running in production (e.g., a recommender system). Nonetheless, the baselin...
详细信息
ISBN:
(纸本)9781577358350
In many fields such as digital marketing, healthcare, finance, and robotics, it is common to have a well-tested and reliable baseline policy running in production (e.g., a recommender system). Nonetheless, the baseline policy is often suboptimal. In this case, it is desirable to deploy online learning algorithms (e.g., a multi-armed bandit algorithm) that inter-act with the system to learn a better/optimal policy under the constraint that during the learning process the performance is almost never worse than the performance of the baseline itself. In this paper, we study the conservative learning problem in the contextual linear bandit setting and introduce a novel algorithm, the Conservative Constrained LINUCB (CLUCB2). We derive regret bounds for CLUCB2 that match existing results and empirically show that it outperforms state-of-the-art conservative bandit algorithms in a number of synthetic and real-world problems. Finally, we consider a more realistic constraint where the performance is verified only at predefined checkpoints (instead of at every step) and show how this relaxed constraint favorably impacts the regret and empirical performance of CLUCB2.
The k-Nearest Neighbors (kNN) classifier is a fundamental non-parametric machine learning algorithm. However, it is well known that it suffers from the curse of dimensionality, which is why in practice one often appli...
详细信息
ISBN:
(纸本)9781713829546
The k-Nearest Neighbors (kNN) classifier is a fundamental non-parametric machine learning algorithm. However, it is well known that it suffers from the curse of dimensionality, which is why in practice one often applies a kNN classifier on top of a (pre-trained) feature transformation. From a theoretical perspective, most, if not all theoretical results aimed at understanding the kNN classifier are derived for the raw feature space. This leads to an emerging gap between our theoretical understanding of kNN and its practical applications. In this paper, we take a first step towards bridging this gap. We provide a novel analysis on the convergence rates of a kNN classifier over transformed features. This analysis requires in-depth understanding of the properties that connect both the transformed space and the raw feature space. More precisely, we build our convergence bound upon two key properties of the transformed space: (1) safety - how well can one recover the raw posterior from the transformed space, and (2) smoothness - how complex this recovery function is. Based on our result, we are able to explain why some (pre-trained) feature transformations are better suited for a kNN classifier than others. We empirically validate that both properties have an impact on the kNN convergence on 30 feature transformations with 6 benchmark datasets spanning from the vision to the text domain.
Recently, the subject-specific surface electromyography (sEMG)-based gesture classification with deep learning algorithms has been widely researched. However, it is not practical to obtain the training data by requiri...
详细信息
ISBN:
(纸本)9781728119908
Recently, the subject-specific surface electromyography (sEMG)-based gesture classification with deep learning algorithms has been widely researched. However, it is not practical to obtain the training data by requiring a user to perform hand gestures many times in real life. This problem can be alleviated to a certain extent if sEMG from many other subjects could be used to train the classifier. In this paper, we propose a normalisation approach that allows implementing real-time subject-independent sEMG based hand gesture classification without training the deep learning algorithm subject specifically. We hypothesed that the amplitude ranges of sEMG across channels between forearm muscle contractions for a hand gesture recorded in the same condition do not vary significantly within each individual. Therefore, the min-max normalisation is applied to source domain data but the new maximum and minimum values of each channel used to restrict the amplitude range are calculated from a trial cycle of a new user (target domain) and assigned by the class label. A convolutional neural network (ConvNet) trained with the normalised data achieved an average 87.03% accuracy on our G. dataset (12 gestures) and 94.53% on M. dataset (7 gestures) by using the leave-one-subject-out cross-validation.
暂无评论