Arnaud Delorme

Enhanced detection of artifacts in EEG data using higher-order statistics and independent component analysis

Detecting artifacts produced in EEG data by muscle activity, eye blinks and electrical noise is a common and important problem in EEG research. It is now widely accepted that independent component analysis (ICA) may be a useful tool for isolating artifacts and/or cortical processes from electroencephalographic (EEG) data. We present results of simulations demonstrating that ICA decomposition, here tested using three popular ICA algorithms, Infomax, SOBI, and FastICA, can allow more sensitive automated detection of small non-brain artifacts than applying the same detection methods directly to the scalp channel data. We tested the upper bound performance of five methods for detecting various types of artifacts by separately optimizing and then applying them to artifact-free EEG data into which we had added simulated artifacts of several types, ranging in size from thirty times smaller (-50 dB) to the size of the EEG data themselves (0 dB). Of the methods tested, those involving spectral thresholding were most sensitive. Except for muscle artifact detection where we found no gain of using ICA, all methods proved more sensitive when applied to the ICA-decomposed data than applied to the raw scalp data: the mean performance for ICA was higher and situated at about two standard deviations away from the performance distribution obtained on raw data. We note that ICA decomposition also allows simple subtraction of artifacts accounted for by single independent components, and/or separate and direct examination of the decomposed non-artifact processes themselves.

Frontal midline EEG dynamics during working memory

We show that during visual working memory, the electroencephalographic (EEG) process producing 5-7 Hz frontal midline theta (fmtheta) activity exhibits multiple spectral modes involving at least three frequency bands and a wide range of amplitudes. The process accounting for the fmtheta increase during working memory was separated from 71-channel data by clustering on time/frequency transforms of components returned by independent component analysis (ICA). Dipole models of fmtheta component scalp maps were consistent with their generation in or near dorsal anterior cingulate cortex. From trial to trial, theta power of fmtheta components varied widely but correlated moderately with theta power in other frontal and left temporal processes. The weak mean increase in frontal midline theta power with increasing memory load, produced entirely by the fmtheta components, largely reflected progressively stronger theta activity in a relatively small proportion of trials. During presentations of letter series to be memorized or ignored, fmtheta components also exhibited 12-15 Hz low-beta activity that was stronger during memorized than during ignored letter trials, independent of letter duration. The same components produced a brief 3-Hz burst 500 ms after onset of the Probe letter following each letter sequence. A new decomposition method, log spectral ICA, applied to normalized log time/frequency transforms of fmtheta component Memorize-letter trials, showed that their low-beta activity reflected harmonic energy in continuous, sharp-peaked theta wave trains as well as independent low-beta bursts. Possibly, the observed fmtheta process variability may index dynamic adjustments in medial frontal cortex to trial-specific behavioral context and task demands.

Spike-based strategies for rapid processing

Most experimental and theoretical studies of brain function assume that neurons transmit information as a rate code, but recent studies on the speed of visual processing impose temporal constraints that appear incompatible with such a coding scheme. Other coding schemes that use the pattern of spikes across a population a neurons may be much more efficient. For example, since strongly activated neurons tend to fire first, one can use the order of firing as a code. We argue that Rank Order Coding is not only very efficient, but also easy to implement in biological hardware: neurons can be made sensitive to the order of activation of their inputs by including a feed-forward shunting inhibition mechanism that progressively desensitizes the neuronal population during a wave of afferent activity. In such a case, maximum activation will only be produced when the afferent inputs are activated in the order of their synaptic weights.

A Limit to the Speed of Processing in Ultra-Rapid Visual Categorization of Novel Natural Scenes

The processing required to decide whether a briefly flashed natural scene contains an animal can be achieved in 150 msec (Thorpe, Fize, & Marlot, 1996). Here we report that extensive training with a subset of photographs over a 3-week period failed to increase the speed of the processing underlying such Rapid Visual Categorizations: Completely novel scenes could be categorized just as fast as highly familiar ones. Such data imply that the visual system processes new stimuli at a speed and with a number of stages that cannot be compressed. This rapid processing mode was seen with a wide range of visual complex images, challenging the idea that short reaction times can only be seen with simple visual stimuli and implying that highly automatic feed-forward mechanisms underlie a far greater proportion of the sophisticated image analysis needed for everyday vision than is generally assumed.

Electroencephalographic brain dynamics following manually responded visual targets.

Scalp-recorded electroencephalographic (EEG) signals produced by partial synchronization of cortical field activity mix locally synchronous electrical activities of many cortical areas. Analysis of event-related EEG signals typically assumes that poststimulus potentials emerge out of a flat baseline. Signals associated with a particular type of cognitive event are then assessed by averaging data from each scalp channel across trials, producing averaged event-related potentials (ERPs). ERP averaging, however, filters out much of the information about cortical dynamics available in the unaveraged data trials. Here, we studied the dynamics of cortical electrical activity while subjects detected and manually responded to visual targets, viewing signals retained in ERP averages not as responses of an otherwise silent system but as resulting from event-related alterations in ongoing EEG processes. We applied infomax independent component analysis to parse the dynamics of the unaveraged 31-channel EEG signals into maximally independent processes, then clustered the resulting processes across subjects by similarities in their scalp maps and activity power spectra, identifying nine classes of EEG processes with distinct spatial distributions and event-related dynamics. Coupled two-cycle postmotor theta bursts followed button presses in frontal midline and somatomotor clusters, while the broad postmotor “P300” positivity summed distinct contributions from several classes of frontal, parietal, and occipital processes. The observed event-related changes in local field activities, within and between cortical areas, may serve to modulate the strength of spike-based communication between cortical areas to update attention, expectancy, memory, and motor preparation during and after target recognition and speeded responding.

Independent EEG sources are dipolar

Independent component analysis (ICA) and blind source separation (BSS) methods are increasingly used to separate individual brain and non-brain source signals mixed by volume conduction in electroencephalographic (EEG) and other electrophysiological recordings. We compared results of decomposing thirteen 71-channel human scalp EEG datasets by 22 ICA and BSS algorithms, assessing the pairwise mutual information (PMI) in scalp channel pairs, the remaining PMI in component pairs, the overall mutual information reduction (MIR) effected by each decomposition, and decomposition ‘dipolarity’ defined as the number of component scalp maps matching the projection of a single equivalent dipole with less than a given residual variance. The least well-performing algorithm was principal component analysis (PCA); best performing were AMICA and other likelihood/mutual information based ICA methods. Though these and other commonly-used decomposition methods returned many similar components, across 18 ICA/BSS algorithms mean dipolarity varied linearly with both MIR and with PMI remaining between the resulting component time courses, a result compatible with an interpretation of many maximally independent EEG components as being volume-conducted projections of partially-synchronous local cortical field activity within single compact cortical domains. To encourage further method comparisons, the data and software used to prepare the results have been made available (http://sccn.ucsd.edu/wiki/BSSComparison).

EEGLAB, SIFT, NFT, BCILAB, and ERICA: new tools for advanced EEG processing

We describe a set of complementary EEG data collection and processing tools recently developed at the Swartz Center for Computational Neuroscience (SCCN) that connect to and extend the EEGLAB software environment, a freely available and readily extensible processing environment running under Matlab. The new tools include (1) a new and flexible EEGLAB STUDY design facility for framing and performing statistical analyses on data from multiple subjects; (2) a neuroelectromagnetic forward head modeling toolbox (NFT) for building realistic electrical head models from available data; (3) a source information flow toolbox (SIFT) for modeling ongoing or event-related effective connectivity between cortical areas; (4) a BCILAB toolbox for building online brain-computer interface (BCI) models from available data, and (5) an experimental real-time interactive control and analysis (ERICA) environment for real-time production and coordination of interactive, multimodal experiments.

Face processing using one spike per neurone.

The speed with which neurones in the monkey temporal lobe can respond selectively to the presence of a face implies that processing may be possible using only one spike per neurone, a finding that is problematic for conventional rate coding models that need at least two spikes to estimate interspike interval. One way of avoiding this problem uses the fact that integrate-and-fire neurones will tend to fire at different times, with the most strongly activated neurones firing first (Thorpe, 1990, Parallel Processing in Neural Systems). Under such conditions, processing can be performed by using the order in which cells in a particular layer fire as a code. To test this idea, we have explored a range of architectures using SpikeNET (Thorpe and Gautrais, 1997, Neural Information Processing Systems, 9), a simulator designed for modelling large populations of integrate-and-fire neurones. One such network used a simple four-layer feed-forward architecture to detect and localise the presence of human faces in natural images. Performance of the model was tested with a large range of grey-scale images of faces and other objects and was found to be remarkably good by comparison with more classic image processing techniques. The most remarkable feature of these results is that they were obtained using a purely feed-forward neural network in which none of the neurones fired more than one spike (thus ruling out conventional rate coding mechanisms). It thus appears that the combination of asynchronous spike propagation and rank order coding may provide an important key to understanding how the nervous system can achieve such a huge amount of processing in so little time.

Face identification using one spike per neuron: resistance to image degradations.

The short response latencies of face selective neurons in the inferotemporal cortex impose major constraints on models of visual processing. It appears that visual information must essentially propagate in a feed-forward fashion with most neurons only having time to fire one spike. We hypothesize that flashed stimuli can be encoded by the order of firing of ganglion cells in the retina and propose a neuronal mechanism, that could be related to fast shunting inhibition, to decode such information. Based on these assumptions, we built a three-layered neural network of retino-topically organized neuronal maps. We showed, by using a learning rule involving spike timing dependant plasticity, that neuronal maps in the output layer can be trained to recognize natural photographs of faces. Not only was the model able to generalize to novel views of the same faces, it was also remarkably resistant to image noise and reductions in contrast.

SPIKENET : A SIMULATOR FOR MODELING LARGE NETWORKS OF INTEGRATE AND FIRE NEURONS

SpikeNET is a simulator for modeling large networks of asynchronously spiking neurons. It uses simple integrate-and-fire neurons which undergo step-like changes in membrane potential when synaptic inputs arrive. If a threshold is exceeded, the potential is reset and the neuron added to a list to be propagated on the next time step. Using such spike lists greatly reduces the computations associated with large networks, and simplifies implementations using parallel hardware since inter-processor communication can be limited to sending lists of the neurons which just fired. We have used it to model complex multi-layer architectures based on the primate visual system that involve millions of neurons and billions of synaptic connections. Such models are not only biological but also efficient, robust and very fast, qualities which they share with the human visual system.