TL;DR
This paper introduces an asymmetrical sparse autoencoder with a Discrete Cosine Transform (DCT) layer for efficient Electroencephalogram (EEG) data compression. By combining a lightweight encoder with a more complex decoder and incorporating trainable hard-thresholding and scaling within the DCT layer, the method achieves superior compression ratios and reconstruction quality compared to existing techniques, making it suitable for real-time sensor applications.
What Problem Does This Paper Address?
Electroencephalogram (EEG) signals are rich in information but also generate large amounts of data, posing a significant challenge for wireless transmission, storage, and real-time analysis. Existing compression methods, while improving, often struggle to balance compression efficiency with high reconstruction accuracy. Traditional signal transform methods can lack accuracy, while neural network-based methods can be computationally intensive, especially at the sensor side. This paper aims to bridge this gap by developing a highly efficient and accurate EEG compression technique suitable for resource-constrained sensor devices.
Methods
The core innovation is an asymmetrical sparse autoencoder with a Discrete Cosine Transform (DCT) layer.
- Encoder: It's designed for low complexity. It starts with a fully connected linear layer, followed by a DCT layer. Within the DCT layer, the authors introduce:
- Trainable Hard-Thresholding: This nonlinearity, unlike soft-thresholding, is found to perform better for EEG data, helping to denoise and compress by zeroing out less significant DCT coefficients. The threshold parameters are trainable.
- Trainable Scaling Layers: These layers apply weights to DCT coefficients, akin to filtering in the transform domain, allowing emphasis or de-emphasis on specific coefficients. This is an alternative to hand-crafted quantization matrices used in standards like JPEG.
- One-by-One Convolution: This layer reduces the number of channels and creates the compressed latent space.
- Sparsity Penalty: A cost function incorporating a Kullback-Leibler Divergence (KLD) penalty is used to encourage sparsity in the latent space, further improving compression efficiency. A threshold parameter controls when training terminates based on the proportion of zeros in the latent representation.
- Decoder: Designed for higher complexity to enhance reconstruction accuracy, it uses the inverse DCT (IDCT) followed by two fully connected linear layers. This asymmetry allows for efficient encoding at the sensor and detailed reconstruction on a more powerful host.
- Data Encoding: For transmission, the compressed data is converted to integers and then encoded using a combination of Run-Length Encoding (RLE) and Lempel-Ziv-Markov chain algorithm (LZMA).
Datasets
The proposed model was evaluated on two publicly available EEG datasets:
- BCI2 dataset: Collected from a normal subject with 28 EEG channels, sampled at 100Hz. It includes 316 training and 100 testing sets.
- Bonn University dataset: This dataset includes five records (F, N, O, Z, S). For evaluating generalization, record S from an epileptic subject was used.
All data was partitioned into blocks of size N=64 for processing.
Key Insights and Ideas
- Asymmetrical Design is Key: The paper highlights that an asymmetrical autoencoder, with a lightweight encoder and a more powerful decoder, is crucial for efficient sensor-based compression and accurate reconstruction.
- Hard-Thresholding for EEG: Contrary to common practice, hard-thresholding in the DCT domain proved more effective than soft-thresholding for EEG compression, suggesting it better preserves important signal components while removing noise.
- Trainable DCT Parameters: Making thresholding and scaling parameters trainable within the DCT layer, rather than fixed, allows the model to adapt to the specific characteristics of EEG data, leading to better compression.
- Sparsity for Efficiency: The use of a sparsity penalty in the cost function is a direct mechanism to reduce the number of non-zero coefficients in the latent space, directly translating to higher compression ratios.
- DCT as a Neural Layer: Integrating the DCT as a trainable layer within a neural network framework allows for end-to-end learning, optimizing the transform and its subsequent processing for the specific task.
Key Results
- The proposed ASAEDCT model significantly improved compression performance on both the BCI2 and Bonn University datasets compared to state-of-the-art methods like ANN+DCT, Sparse AE, and DCT-DOST.
- On the BCI2 dataset, ASAEDCT achieved a compression ratio (CR) of 25.66 and a percent root-mean-square difference (PRD) of 5.09, outperforming others in terms of quality score (QS).
- The model demonstrated good generalization ability, performing well on the Bonn University dataset even when trained on BCI2.
- The computational cost for compression was very low, with an input duration of 6.4s taking only 0.016s to compress, making it suitable for real-time applications.
- An ablation study showed that removing the scaling layer, sparsity penalty, or nonlinearity degraded performance, emphasizing the importance of each component. The three-channel (C=3) configuration was found to be optimal.
Limitations and Future Work
- Limitations: The paper focuses on a specific type of EEG signal compression. While it shows good generalization, further testing on a wider variety of EEG data types and conditions might be beneficial. The paper mentions that the decoder can be implemented on a powerful host computer, implying that the computational cost of the decoder itself might still be a consideration for fully embedded systems.
- Future Work: The authors suggest that the proposed asymmetrical autoencoder with a DCT layer can be implemented in sensors for EEG data compression. Further research could explore extending this architecture to other biosignal compression tasks or investigating more advanced hybrid coding schemes for even higher compression ratios.
Summary updated