Look Into the LITE in Deep Learning for Time Series Classification

TL;DR

This work introduces LITE, a lightweight Time Series Classification architecture that leverages DepthWise Separable Convolutions and boosting techniques (multiplexing, dilation, and custom filters) to achieve a fraction of the parameter count of state-of-the-art models while maintaining competitive accuracy. It extends to LITEMV for multivariate time series and employs ensemble methods (LITETime, LITEMVTime) to rival larger architectures with far fewer parameters, yielding faster training and lower energy consumption on the UCR archive. The paper also demonstrates practical applicability in human rehabilitation data and provides CAM-based interpretability analyses to illuminate decision-making. Ablation studies dissect the contributions of each boosting technique, and limitations related to receptive field and scalability are discussed, outlining paths for future work.

Abstract

Deep learning models have been shown to be a powerful solution for Time Series Classification (TSC). State-of-the-art architectures, while producing promising results on the UCR and the UEA archives , present a high number of trainable parameters. This can lead to long training with high CO2 emission, power consumption and possible increase in the number of FLoating-point Operation Per Second (FLOPS). In this paper, we present a new architecture for TSC, the Light Inception with boosTing tEchnique (LITE) with only 2.34% of the number of parameters of the state-of-the-art InceptionTime model, while preserving performance. This architecture, with only 9, 814 trainable parameters due to the usage of DepthWise Separable Convolutions (DWSC), is boosted by three techniques: multiplexing, custom filters, and dilated convolution. The LITE architecture, trained on the UCR, is 2.78 times faster than InceptionTime and consumes 2.79 times less CO2 and power. To evaluate the performance of the proposed architecture on multivariate time series data, we adapt LITE to handle multivariate time series, we call this version LITEMV. To bring theory into application, we also conducted experiments using LITEMV on multivariate time series representing human rehabilitation movements, showing that LITEMV not only is the most efficient model but also the best performing for this application on the Kimore dataset, a skeleton based human rehabilitation exercises dataset. Moreover, to address the interpretability of LITEMV, we present a study using Class Activation Maps to understand the classification decision taken by the model during evaluation.

Figures (22)

• Figure 1: For each model, the accuracy on the FreezerSmallTrain dataset is presented on the $y$-axis and the number of FLoat-point Operations Per Second (FLOPS) is presented on the $x$-axis in $\log_{10}$ scale. The diameter of the circles represents the number of trainable parameters of the model. The smallest model is LITE (ours) with only $~10k$ trainable parameters and the lowest number of FLOPS ($~4$ in $\log_{10}$ scale); it also presents the highest accuracy score on the test set in this comparison.
• Figure 2: DWSC for time series represented in its two phases: (1) DepthWise convolution (purple block), and (2) PointWise convolution (blue block).
• Figure 3: Multiplexing one dimensional convolution on the input time series (in blue) using filters (in green) with three different kernel sizes, respectively, 3, 5, and 7. The output of the convolutions (in red) is different for each filter.
• Figure 4: One dimensional convolution performed w/o dilation on the left (rate=1), and with dilation on the right (rate=2).
• Figure 5: The proposed LITE architecture that uses multiplexing convolutions in the first layer (three convolution blocks in orange) with custom filters (in green). The second and third layer are composed of DWSC (in beige). The last layer is followed by a Global Average Pooling (GAP) on the time axis and finished by a classification Fully Connected layer to approximate the class distribution.