Two Algorithms for Weakly Supervised Denoising of EEG Data

Editor's Notes

• #6: Technique for monitoring brain activity
• #7: Wu, Dongrui, Vernon J. Lawhern, Stephen Gordon, Brent J. Lance, and Chin-Teng Lin. "Offline EEG-based driver drowsiness estimation using enhanced batch-mode active learning (EBMAL) for regression." In Systems, Man, and Cybernetics (SMC), 2016 IEEE International Conference on, pp. 000730-000736. IEEE, 2016.Harvard Lance, Brent J., W. David Hairston, Greg Apker, Keith W. Whitaker, Geoff Slipher, Randy Mrozek, Scott E. Kerick, Jason Metcalfe, Christopher Manteuffel, and Matthew Jaswa. 2012 year-end report on neurotechnologies for in-vehicle applications. No. ARL-SR-267. ARMY RESEARCH LAB ABERDEEN PROVING GROUND MD, 2013. Offline EEG-Based Driver Drowsiness Estimation Using Enhanced Batch-Mode Active Learning (EBMAL) for Regression assess the state of human vehicle operators and integrate that state information into future vehicle system
• #8: Wu, Dongrui, Vernon J. Lawhern, Stephen Gordon, Brent J. Lance, and Chin-Teng Lin. "Offline EEG-based driver drowsiness estimation using enhanced batch-mode active learning (EBMAL) for regression." In Systems, Man, and Cybernetics (SMC), 2016 IEEE International Conference on, pp. 000730-000736. IEEE, 2016.Harvard Lance, Brent J., W. David Hairston, Greg Apker, Keith W. Whitaker, Geoff Slipher, Randy Mrozek, Scott E. Kerick, Jason Metcalfe, Christopher Manteuffel, and Matthew Jaswa. 2012 year-end report on neurotechnologies for in-vehicle applications. No. ARL-SR-267. ARMY RESEARCH LAB ABERDEEN PROVING GROUND MD, 2013. Offline EEG-Based Driver Drowsiness Estimation Using Enhanced Batch-Mode Active Learning (EBMAL) for Regression assess the state of human vehicle operators and integrate that state information into future vehicle system
• #9: Page, Adam, Chris Sagedy, Emily Smith, Nasrin Attaran, Tim Oates, and Tinoosh Mohsenin. "A flexible multichannel EEG feature extractor and classifier for seizure detection." IEEE Transactions on Circuits and Systems II: Express Briefs 62, no. 2 (2015): 109-113. Wu, Dongrui, Vernon J. Lawhern, Stephen Gordon, Brent J. Lance, and Chin-Teng Lin. "Offline EEG-based driver drowsiness estimation using enhanced batch-mode active learning (EBMAL) for regression." In Systems, Man, and Cybernetics (SMC), 2016 IEEE International Conference on, pp. 000730-000736. IEEE, 2016.Harvard Lance, Brent J., W. David Hairston, Greg Apker, Keith W. Whitaker, Geoff Slipher, Randy Mrozek, Scott E. Kerick, Jason Metcalfe, Christopher Manteuffel, and Matthew Jaswa. 2012 year-end report on neurotechnologies for in-vehicle applications. No. ARL-SR-267. ARMY RESEARCH LAB ABERDEEN PROVING GROUND MD, 2013. Offline EEG-Based Driver Drowsiness Estimation Using Enhanced Batch-Mode Active Learning (EBMAL) for Regression assess the state of human vehicle operators and integrate that state information into future vehicle system
• #10: Wu, Dongrui, Vernon J. Lawhern, Stephen Gordon, Brent J. Lance, and Chin-Teng Lin. "Offline EEG-based driver drowsiness estimation using enhanced batch-mode active learning (EBMAL) for regression." In Systems, Man, and Cybernetics (SMC), 2016 IEEE International Conference on, pp. 000730-000736. IEEE, 2016.Harvard assess the state of human vehicle operators and integrate that state information into future vehicle system There are bunch of other applications that I won’t go into.
• #15: Some electrodes
• #17: There is a strong need for method for denoising the EEG data
• #20: We saw few slides back that it is important problem and large number of methods have been proposed
• #21: Users with relatively short data files, for whom segment rejection would lead to an unacceptably low remaining number of segments for analysis, may choose an optional post-segmentation step involving the interpolation of data for channels determined to be artifact-contaminated within each individual segment, as implemented by FASTER software (Nolanet al., 2010). Each channel in each segment is evaluated onthe four FASTER criteria (variance, median gradient, amplitude range, and deviation from mean amplitude), and the Z score(a measure of standard deviation from the mean) for each channel in that segment is generated for each of the fourmetrics. Any channels with one or more Z scores that are greater than 3 standard deviations from the mean for an individual segment are marked bad for that segment. These criteria may identify segments with residual high-amplitude artifacts(e.g., eye artifacts), electrode discontinuity (e.g., electrode has lost contact with the scalp temporarily), and muscle artifact. Subsequently, for each segment, the bad channels have their data interpolated with spherical splines, as in FASTER. This allows users to maintain the maximum number of available segments, while still maximizing artifact rejection within individual segments.
• #22: captures both frequency and location information Artifacts should be decomposable in a wavelet basis. Mother wavelet, the shrinkage rule, and the threshold are important to the design of the noise removal method (Daly et al. 2013).
• #25: This problem is widely studied and large number of
• #26: Annotation or reference channel
• #27: Annotation or reference channel
• #29: In this paper, we present a novel software-hardware system that uses a weak supervisory signal to indicate that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. The slide shows our proposed systtem architecture , First First 64 EEG data are decomposed into independent components using ICA, and these components form bags that are labeled 1/0 where 1 indicates that there is a noise. Each bag is then classified by a multi-instance learning algorithm to identify if it has noise component or not. We use Symoblic Aggregate Approximation (SAX) for featurizing each component. We go through each of these subsystems in next slides
• #30: https://www.slideshare.net/yokotatsuya/independent-component-analysis-11359849/3
• #31: This problem is widely studied and large number of
• #32: https://www.slideshare.net/yokotatsuya/independent-component-analysis-11359849/3
• #33: https://www.slideshare.net/yokotatsuya/independent-component-analysis-11359849/3
• #34: In this paper, we present a novel software-hardware system that uses a weak supervisory signal to indicate that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. The slide shows our proposed systtem architecture , First First 64 EEG data are decomposed into independent components using ICA, and these components form bags that are labeled 1/0 where 1 indicates that there is a noise. Each bag is then classified by a multi-instance learning algorithm to identify if it has noise component or not. We use Symoblic Aggregate Approximation (SAX) for featurizing each component. We go through each of these subsystems in next slides
• #35: We featurize each component using SAX before using multi instance learning. This slide demonstrates how the SAX works on a simple time series signal. We partition the signal into different levels where each level manifests a letter. For example, this window of the signal is partitioned into three levels which represents the symbols of “baabccbc” word. We use bag of words generated from SAX to classify the time series. For featurizing EEG, we convert time series to symbols In this case time series shown in first figure is converted to symbols “baabccbc”.
• #36: In this paper, we present a novel software-hardware system that uses a weak supervisory signal to indicate that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. The slide shows our proposed systtem architecture , First First 64 EEG data are decomposed into independent components using ICA, and these components form bags that are labeled 1/0 where 1 indicates that there is a noise. Each bag is then classified by a multi-instance learning algorithm to identify if it has noise component or not. We use Symoblic Aggregate Approximation (SAX) for featurizing each component. We go through each of these subsystems in next slides
• #38: Our goal is given a new bag predict if it is positive or negative.
• #39: Multi instance learning algorithms also allows us to predict if the instance is positive or negative. In our case positive instance corresponds to noisy component in EEG data.
• #40: In this paper, we present a novel software-hardware system that uses a weak supervisory signal to indicate that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. The slide shows our proposed systtem architecture , First First 64 EEG data are decomposed into independent components using ICA, and these components form bags that are labeled 1/0 where 1 indicates that there is a noise. Each bag is then classified by a multi-instance learning algorithm to identify if it has noise component or not. We use Symoblic Aggregate Approximation (SAX) for featurizing each component. We go through each of these subsystems in next slides
• #44: Although few reasons these numbers cannot be directly compared: We do 3-Fold cross validation, they randomly generate 20 train test split They are using 7 patients from same dataset. I am not sure which patients they are using The goals are different they are doing classification, we are doing classification as well as artifact component removal. Also, I know there are problems with evaluation, I will talk about evaluation at the end
• #53: Another approach we are working on to remove artifacts from noisy EEG data is using generative adversarial networks. This approach has been used recently in computer vision for converting images from one domain to another. For example, in this paper, unpaired image-to-image translation using cycle-consistent adversarial networks (CycleGAN) convert paintings of Monet to photos. They also convert photos of zebras to horses, winter to summer and reverse. We want to use this deep neural network to remove artifacts from EEG data. This architecture is a natural fit for this problem as it does not require any paired data. I will go over how cycleGAN work in the next few slides.
• #54: For explaining cycleGAN, I will use example of converting a horse image to zebra image as it is easier to explain. In this example, X is images of all horses and Y is images of all zebras. G is convolutional autoencoder which converts horse image to corresponding zebra image and F is another convolutional autoencoder which converts zebra image to horse image. Dx and Dy are convolutional neural networks that discriminate between generated and real images. For training this neural network, we give a horse and zebra image as input and calculate the loss. After calculating the loss, we use standard gradient descent for backpropagation and train the network. We describe how we calculate the loss function in next slide.
• #56: Another approach we are working on to remove artifacts from noisy EEG data is using generative adversarial networks. This approach has been used recently in computer vision for converting images from one domain to another. For example, in this paper, unpaired image-to-image translation using cycle-consistent adversarial networks (CycleGAN) convert paintings of Monet to photos. They also convert photos of zebras to horses, winter to summer and reverse. We want to use this deep neural network to remove artifacts from EEG data. This architecture is a natural fit for this problem as it does not require any paired data. I will go over how cycleGAN work in the next few slides.
• #58: Give Example Talk about how network cheats
• #59: Visual Analogy of the problem
• #60: Give Example Talk about how network cheats
• #61: Give Example Talk about how network cheats
• #62: Color the arrows
• #64: To train the asymmetricGAN and reduce model oscillation, we use the strategy of (Shrivastava et al. 2016) of saving the history of generated time series. We update the discriminators using a history of generated signals rather than the ones produced by the latest generative networks. We keep a buffer that stores the 50 previously generated signals for training. In the next section, we show the effectiveness of our method on synthetic and EEG datasets. Explain the architecture visually
• #65: Artifact removal from EEG data is a problem where ground truth does not exist be- cause we do not have a clean version of the noisy signal. This makes the evaluation of artifact removal methods difficult. Also, visualizing and understanding EEG data is a time consuming task. We solve this problem by creating a simpler synthetic dataset.
• #67: Talk about the fact that this is qualitative example how mse just gives consistency across the dataset MSE(a,b) = 0.5
• #68: Talk about the fact that this is qualitative example how mse just gives consistency across the dataset MSE(a,b) = 0.5
• #69: Give Example
• #71: More Examples in Thesis
• #81: Qualitative Evaluation Subjective Useful for validating the algorithm.
• #82: Most used Evaluation using detection
• #83: Most used Methods generally differ in how they model characteristics of artifact signals. (Daly et al. 2012) proposes to use mean power and its standard deviation, the maximum amplitude and its standard deviation, kurtosis, and skewness of the amplitude of resting state background rhythms to characterize the clean and noisy EEG signal. (Mognon et al. 2010) learns features such as kur- tosis, spatial average difference, and maximum epoch variance for modeling ocular artifacts. (Delorme, Sejnowski, and Makeig 2007b) learns to characterize a larger variety of artifacts using extreme values, linear trends, data improbability, kurtosis and spectral patterns. We use an artifact detector that uses a convolutional neural network to evaluate artifact removal algorithms in Chapter 4.
• #84: Most used Methods generally differ in how they model characteristics of artifact signals. (Daly et al. 2012) proposes to use mean power and its standard deviation, the maximum amplitude and its standard deviation, kurtosis, and skewness of the amplitude of resting state background rhythms to characterize the clean and noisy EEG signal. (Mognon et al. 2010) learns features such as kur- tosis, spatial average difference, and maximum epoch variance for modeling ocular artifacts. (Delorme, Sejnowski, and Makeig 2007b) learns to characterize a larger variety of artifacts using extreme values, linear trends, data improbability, kurtosis and spectral patterns. We use an artifact detector that uses a convolutional neural network to evaluate artifact removal algorithms in Chapter 4.
• #85: Most used Multiple evaluation methods for artifact removal algorithms have been proposed in the literature. Each of these methods has their strengths and limitations. Thus, it is important to use as many evaluation mechanisms as possible. But, most existing papers use one or two of these methods to evaluate their algorithm. This makes performance comparison of these methods hard. This problem is exacerbated because several variants of each of these methods exist in literature and they are tested on different datasets. Our goal in this thesis is to standardize the evaluation mechanism and create a tool for evaluating artifact removal algorithms using all of the above approaches. We also intend to use this tool for comparing our approach with existing methods discussed above.
• #91: where speech can be denoised before amplification.
• #92: TODO : Reduce content on slide 65