Time series classification from scratch with deep neural networks: A strong baseline

Wang, Zhiguang; Yan, Weizhong; Oates, Tim

doi:10.1109/ijcnn.2017.7966039

Cited by 1,200 publications

(938 citation statements)

References 22 publications

Supporting

Mentioning

928

Contrasting

Unclassified

Order By: Relevance

“…This is relevant since the process of feature engineering is time-consuming and laborious. Moreover, unsupervised feature extraction from such methods has been proven to work well [17,30]. Sparse encoding via linear-algebra methods for dictionary learning is also a possibility.…”

Section: Technical Approachmentioning

confidence: 99%

“…The age distribution consisted of children (7-10) and adults (26)(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(40). Recorded matches had a minimum duration of about 40 s and a maximum of about 1 min and 10 s. The collected data correspond to acceleration values along the x-, y-and z-axis with a sampling frequency of 50 Hz, which is five-times larger than the frequency considered to be sufficient for detecting daily activities from accelerometer data (10 Hz) [32][33][34].…”

Section: Game Scenariomentioning

confidence: 99%

See 1 more Smart Citation

Learning and Mining Player Motion Profiles in Physically Interactive Robogames

et al. 2018

View full text Add to dashboard Cite

Physically-Interactive RoboGames (PIRG) are an emerging application whose aim is to develop robotic agents able to interact and engage humans in a game situation. In this framework, learning a model of players' activity is relevant both to understand their engagement, as well as to understand specific strategies they adopted, which in turn can foster game adaptation. Following such directions and given the lack of quantitative methods for player modeling in PIRG, we propose a methodology for representing players as a mixture of existing player's types uncovered from data. This is done by dealing both with the intrinsic uncertainty associated with the setting and with the agent necessity to act in real time to support the game interaction. Our methodology first focuses on encoding time series data generated from player-robot interaction into images, in particular Gramian angular field images, to represent continuous data. To these, we apply latent Dirichlet allocation to summarize the player's motion style as a probabilistic mixture of different styles discovered from data. This approach has been tested in a dataset collected from a real, physical robot game, where activity patterns are extracted by using a custom three-axis accelerometer sensor module. The obtained results suggest that the proposed system is able to provide a robust description for the player interaction.

show abstract

Section: Technical Approachmentioning

confidence: 99%

Section: Game Scenariomentioning

confidence: 99%

Learning and Mining Player Motion Profiles in Physically Interactive Robogames

et al. 2018

View full text Add to dashboard Cite

show abstract

“…x, y) and the spatial axis represents time. In this formulation, the value of the t, c element in the (Time, Coordinate) matrix represents the position value for a tracked cell at time t in coordinate c. Multiple problem domains have shown success in applying CNNs to multi-channel time series data in this manner [37], [38], [39], [40]. Similarly, convolutional layers with one spatial dimension and one channel dimension may allow for motility behavior classification and unsupervised feature learning without a priori definition of handcrafted features.…”

Section: A Related Workmentioning

confidence: 99%

Deep convolutional and recurrent neural networks for cell motility discrimination and prediction

Kimmel

Brack

Marshall

2017

Preprint

View full text Add to dashboard Cite

Cells in culture display diverse motility behaviors that may reflect differences in cell state and function, providing motivation to discriminate between different motility behaviors. Current methods to do so rely upon manual feature engineering. However, the types of features necessary to distinguish between motility behaviors can vary greatly depending on the biological context, and it is not always clear which features may be most predictive in each setting for distinguishing particular cell types or disease states. Convolutional neural networks (CNNs) are machine learning models allowing for relevant features to be learned directly from spatial data. Similarly, recurrent neural networks (RNNs) are a class of models capable of learning long term temporal dependencies. Given that cell motility is inherently spacio-temporal data, we present an approach utilizing both convolutional and long-short-term memory (LSTM) recurrent neural network units to analyze cell motility data. These RNN models provide accurate classification of simulated motility and experimentally measured motility from multiple cell types, comparable to results achieved with hand-engineered features. The variety of cell motility differences we can detect suggests that the algorithm is generally applicable to additional cell types not analyzed here. RNN autoencoders based on the same architecture are capable of learning motility features in an unsupervised manner and capturing variation between myogenic cells in the latent space. Adapting these RNN models to motility prediction, RNNs are capable of predicting muscle stem cell motility from past tracking data with performance superior to standard motion prediction models. This advance in cell motility prediction may be of practical utility in cell tracking applications.

show abstract

“…In literature, time series classification is tackled by a range of traditional Machine Learning (ML) algorithms such as Hidden Markov Models, Neural Networks or Linear Dynamic Systems [2]. Wang et al [3] are the first who introduced Convolutional Neural Networks (CNN) for the classification of univariate time series where no local pooling layers are included and thus, the length of time series kept the same for all convolutions. They also applied a deep Residual Network (ResNet) for TSC.…”

Section: Introductionmentioning

confidence: 99%

Prediction of MRI Hardware Failures based on Image Features Using Time Series Classification

2020

View full text Add to dashboard Cite

Already before systems malfunction one has to know if hardware components will fail in near future in order to counteract in time. Thus, unplanned downtime is ought to be avoided. In medical imaging, maximizing the system's uptime is crucial for patients' health and healthcare provider's daily business. We aim to predict failures of Head/Neck coils used in Magnetic Resonance Imaging (MRI) by training a statistical model on sequential data collected over time. As image features depend on the coil's condition, their deviations from the normal range already hint to future failure. Thus, we used image features and their variation over time to predict coil damage. After comparison of different time series classification methods we found Long Short Term Memorys (LSTMs) to achieve the highest F-score of 86.43% and to tell with 98.33% accuracy if hardware should be replaced.

show abstract

Time series classification from scratch with deep neural networks: A strong baseline

Cited by 1,200 publications

References 22 publications

Learning and Mining Player Motion Profiles in Physically Interactive Robogames

Learning and Mining Player Motion Profiles in Physically Interactive Robogames

Deep convolutional and recurrent neural networks for cell motility discrimination and prediction

Prediction of MRI Hardware Failures based on Image Features Using Time Series Classification

Contact Info

Product

Resources

About