Piotr Mirowski

Classification of patterns of EEG synchronization for seizure prediction

OBJECTIVE Research in seizure prediction from intracranial EEG has highlighted the usefulness of bivariate measures of brainwave synchronization. Spatio-temporal bivariate features are very high-dimensional and cannot be analyzed with conventional statistical methods. Hence, we propose state-of-the-art machine learning methods that handle high-dimensional inputs. METHODS We computed bivariate features of EEG synchronization (cross-correlation, nonlinear interdependence, dynamical entrainment or wavelet synchrony) on the 21-patient Freiburg dataset. Features from all channel pairs and frequencies were aggregated over consecutive time points, to form patterns. Patient-specific machine learning-based classifiers (support vector machines, logistic regression or convolutional neural networks) were trained to discriminate interictal from preictal patterns of features. In this explorative study, we evaluated out-of-sample seizure prediction performance, and compared each combination of feature type and classifier. RESULTS Among the evaluated methods, convolutional networks combined with wavelet coherence successfully predicted all out-of-sample seizures, without false alarms, on 15 patients, yielding 71% sensitivity and 0 false positives. CONCLUSIONS Our best machine learning technique applied to spatio-temporal patterns of EEG synchronization outperformed previous seizure prediction methods on the Freiburg dataset. SIGNIFICANCE By learning spatio-temporal dynamics of EEG synchronization, pattern recognition could capture patient-specific seizure precursors. Further investigation on additional datasets should include the seizure prediction horizon.

Predictive network modeling of the high-resolution dynamic plant transcriptome in response to nitrate

BackgroundNitrate, acting as both a nitrogen source and a signaling molecule, controls many aspects of plant development. However, gene networks involved in plant adaptation to fluctuating nitrate environments have not yet been identified.ResultsHere we use time-series transcriptome data to decipher gene relationships and consequently to build core regulatory networks involved in Arabidopsis root adaptation to nitrate provision. The experimental approach has been to monitor genome-wide responses to nitrate at 3, 6, 9, 12, 15 and 20 minutes using Affymetrix ATH1 gene chips. This high-resolution time course analysis demonstrated that the previously known primary nitrate response is actually preceded by a very fast gene expression modulation, involving genes and functions needed to prepare plants to use or reduce nitrate. A state-space model inferred from this microarray time-series data successfully predicts gene behavior in unlearnt conditions.ConclusionsThe experiments and methods allow us to propose a temporal working model for nitrate-driven gene networks. This network model is tested both in silico and experimentally. For example, the over-expression of a predicted gene hub encoding a transcription factor induced early in the cascade indeed leads to the modification of the kinetic nitrate response of sentinel genes such as NIR, NIA2, and NRT1.1, and several other transcription factors. The potential nitrate/hormone connections implicated by this time-series data are also evaluated.

Comparing SVM and convolutional networks for epileptic seizure prediction from intracranial EEG

Recent research suggests that electrophysiological changes develop minutes to hours before the actual clinical onset in focal epileptic seizures. Seizure prediction is a major field of neurological research, enabled by statistical analysis methods applied to features derived from intracranial Electroencephalographic (EEG) recordings of brain activity. However, no reliable seizure prediction method is ready for clinical applications. In this study, we use modern machine learning techniques to predict seizures from a number of features proposed in the literature. We concentrate on aggregated features that encode the relationship between pairs of EEG channels, such as cross-correlation, nonlinear interdependence, difference of Lyapunov exponents and wavelet analysis-based synchrony such as phase locking. We compare L1-regularized logistic regression, convolutional networks, and support vector machines. Results are reported on the standard Freiburg EEG dataset which contains data from 21 patients suffering from medically intractable focal epilepsy. For each patient, at least one method predicts 100% of the seizures on average 60 minutes before the onset, with no false alarm. Possible future applications include implantable devices capable of warning the patient of an upcoming seizure as well as implanted drug-delivery devices.

SignalSLAM: Simultaneous localization and mapping with mixed WiFi, Bluetooth, LTE and magnetic signals

Indoor localization typically relies on measuring a collection of RF signals, such as Received Signal Strength (RSS) from WiFi, in conjunction with spatial maps of signal fingerprints. A new technology for localization could arise with the use of 4G LTE telephony small cells, with limited range but with rich signal strength information, namely Reference Signal Received Power (RSRP). In this paper, we propose to combine an ensemble of available sources of RF signals to build multi-modal signal maps that can be used for localization or for network deployment optimization. We primarily rely on Simultaneous Localization and Mapping (SLAM), which provides a solution to the challenge of building a map of observations without knowing the location of the observer. SLAM has recently been extended to incorporate signal strength from WiFi in the so-called WiFi-SLAM. In parallel to WiFi-SLAM, other localization algorithms have been developed that exploit the inertial motion sensors and a known map of either WiFi RSS or of magnetic field magnitude. In our study, we use all the measurements that can be acquired by an off-the-shelf smartphone and crowd-source the data collection from several experimenters walking freely through a building, collecting time-stamped WiFi and Bluetooth RSS, 4G LTE RSRP, magnetic field magnitude, GPS reference points when outdoors, Near-Field Communication (NFC) readings at specific landmarks and pedestrian dead reckoning based on inertial data. We resolve the location of all the users using a modified version of Graph-SLAM optimization of the users poses with a collection of absolute location and pairwise constraints that incorporates multi-modal signal similarity. We demonstrate that we can recover the user positions and thus simultaneously generate dense signal maps for each WiFi access point and 4G LTE small cell, “from the pocket”. Finally, we demonstrate the localization performance using selected single modalities, such as only WiFi and the WiFi signal maps that we generated.

KL-divergence kernel regression for non-Gaussian fingerprint based localization

Various methods have been developed for indoor localization using WLAN signals. Algorithms that fingerprint the Received Signal Strength Indication (RSSI) of WiFi for different locations can achieve tracking accuracies of the order of a few meters. RSSI fingerprinting suffers though from two main limitations: first, as the signal environment changes, so does the fingerprint database, which requires regular updates; second, it has been reported that, in practice, certain devices record more complex (e.g bimodal) distributions of WiFi signals, precluding algorithms based on the mean RSSI. In this article, we propose a simple methodology that takes into account the full distribution for computing similarities among fingerprints using Kullback-Leibler divergence, and that performs localization through kernel regression. Our method provides a natural way of smoothing over time and trajectories. Moreover, we propose unsupervised KL-divergence-based recalibration of the training fingerprints. Finally, we apply our method to work with histograms of WiFi connections to access points, ignoring RSSI distributions, and thus removing the need for recalibration. We demonstrate that our results outperform nearest neighbors or Kalman and Particle Filters, achieving up to 1m accuracy in office environments. We also show that our method generalizes to non-Gaussian RSSI distributions.

Demand forecasting in smart grids

Data analytics in smart grids can be leveraged to channel the data downpour from individual meters into knowledge valuable to electric power utilities and end-consumers. Short-term load forecasting (STLF) can address issues vital to a utility but it has traditionally been done mostly at system (city or country) level. In this case study, we exploit rich, multi-year, and high-frequency annotated data collected via a metering infrastructure to perform STLF on aggregates of power meters in a mid-sized city. For smart meter aggregates complemented with geo-specific weather data, we benchmark several state-of-the-art forecasting algorithms, including kernel methods for nonlinear regression, seasonal and temperature-adjusted auto-regressive models, exponential smoothing and state-space models. We show how STLF accuracy improves at larger meter aggregation (at feeder, substation, and system-wide level). We provide an overview of our algorithms for load prediction and discuss system performance issues that impact real time STLF. ® 2014 Alcatel-Lucent.

Probability kernel regression for WiFi localisation

Various methods have been developed for indoor localisation using WLAN signals. Algorithms that fingerprint the received signal strength indicators (RSSI) of WiFi for different locations can achieve tracking accuracies of the order of a few metres. RSSI fingerprinting suffers from two main limitations: first, as the signal environment changes, so does the fingerprint database, which requires regular updates; second, it has been reported that, in practice, certain devices record more complex (e.g bimodal) distributions of WiFi signals, precluding algorithms based on the mean RSSI. Mirowski et al. [2011. KL-divergence kernel regression for non-Gaussian fingerprint based localization. In: International conference on indoor positioning and indoor navigation, Guimaraes, Portugal] have recently introduced a simple methodology that takes into account the full distribution for computing similarities among fingerprints using the Kullback–Leibler (KL) divergence, and then performs localisation through kernel regression. Their algorithm provides a natural way of smoothing over time and motion trajectories and can be applied directly to histograms of WiFi connections to access points, ignoring RSSI distributions, hence removing the need for fingerprint recalibration. It has been shown to outperform nearest neighbours or Kalman and particle filtres, achieving up to 1 m accuracy in office environments. In this article, we focus on the relevance of Gaussian or non-Gaussian distributions for modelling RSSI distributions by considering additional probabilistic kernels for comparing Gaussian distributions and by evaluating them on three contrasting datasets. We discuss their limitations and formulate how the KL-divergence kernel regression algorithm bridges the gap with other WiFi localisation algorithms, notably Bayesian networks, support vector machines and K nearest neighbours. Finally, we revisit the assumptions on the fingerprint maps and overview practical WiFi localisation software implementation.

Probabilistic radio-frequency fingerprinting and localization on the run

Indoor localization is a key enabler for pervasive computing and network optimization. Wireless local area network (WLAN) positioning systems typically rely on fingerprints of received signal strength (RSS) measures from access points. In this paper, we review approaches for modeling full distributions of Wi-Fi signals, including Bayesian graphical models, smoothing, compressive sensing, and random field differentiation and concentrate on the Kullback-Leibler divergence metric that compares multivariate RSS distributions. We provide theoretical insights on the required spatial density of fingerprints and on the number of samples necessary, during tracking or during signal map building, to differentiate among signal distributions and to provide accurate location estimates. We validate our methods on contrasting datasets where we obtain state-of-the-art localization results. Finally, we exploit datasets collected by a self-localizing mobile robot that continuously records Wi-Fi along with ground truth position, where we define increasingly denser fingerprint grids and study asymptotic localization accuracy. ® 2014 Alcatel-Lucent.

Statistical Machine Learning and Dissolved Gas Analysis: A Review

Dissolved gas analysis (DGA) of the insulation oil of power transformers is an investigative tool to monitor their health and to detect impending failures by recognizing anomalous patterns of DGA concentrations. We handle the failure prediction problem as a simple data-mining task on DGA samples, optionally exploiting the transformers age, nominal power and voltage, and consider two approaches: 1) binary classification and 2) regression of the time to failure. We propose a simple logarithmic transform to preprocess DGA data in order to deal with long-tail distributions of concentrations. We have reviewed and evaluated 15 standard statistical machine-learning algorithms on that task, and reported quantitative results on a small but published set of power transformers and on proprietary data from thousands of network transformers of a utility company. Our results confirm that nonlinear decision functions, such as neural networks, support vector machines with Gaussian kernels, or local linear regression can theoretically provide slightly better performance than linear classifiers or regressors. Software and part of the data are available at http://www.mirowski.info/pub/dga.

A Sparse Coding Approach to Household Electricity Demand Forecasting in Smart Grids

With the gradual deployment of smart meters in many cities around the world, new opportunities arise in reducing energy usage and improving consumers’ information and control on their electricity consumption. Central to the provision of these newer services is the ability to accurately forecast the electricity demand of individual households. Compared with load forecasting at the city level and larger system aggregates, load forecasting for individual households is a much harder problem as the loads are much more stochastic and volatile. In this paper, we study the use of sparse coding for modeling and forecasting these individual household electricity loads. The proposed methods are tested on a data set of 5000 households in a joint project with electric power board of Chattanooga, for the period from September 2011 to August 2013. We obtain 10% improvements in the accuracy of forecasting next-day total load and next-week total load when we add sparse code features in ridge regression in this difficult problem. We also evaluate more classical forecasting methods on this forecasting problem, including autoregressive integrated moving average and Holt-Winters smoothing.