Sequence learning in a single trial

Also consistent with the present results, it has been suggested that the PFC may participate in the extraction of regularities based on internal representations, so as to improve behavioral control [59] , [60]. Finally, in support of the role of hippocampal-prefrontal interactions in memory processes, Peyrache and al. While this could only be observed in one patient, the present finding of a contribution of both the hippocampus and PFC activity to the classification of learning stages would be consistent with these findings.

The present study demonstrates that the proposed multivariate decoding approach to single-trial iEEG datasets from individual patients offers an important methodological alternative to group studies of averaged data.

Indeed, grouping iEEG data from different patients can be extremely challenging because the configuration of electrode placement varies across individuals and pathological conditions e. The restricted availability of such iEEG recordings is an additional factor motivating the development of methods that fully exploit individual patient datasets. Following a classification scheme, we made use of non-overlapping splits of the available data to validate differences in the temporal and spatial configuration of the signal.

Such classification analysis exploits stimulus-related information extracted from distributed activity patterns i. Pattern information analysis has predominantly been developed for neuroimaging studies based on fMRI [3] — [6].

Classification analysis of electrophysiological data has been predominantly developed in the context of Brain Computer Interface [71] — [74] studies while less effort has been devoted to its application in neuroimaging studies [7].

As illustrated by the present study, one main advantage of this methodological framework is that it does not require any a priori selection of the relevant neural sites or time-window of activity measurements. We complemented our approach by a localization procedure and showed that we can identify the regions contributing most to the performance of the classifier.

Our results demonstrate that this methodological strategy is particularly adapted to test the effect of learning-related changes capturing task-related effects at the single-trial level, that are not time-locked to the visual cue or motor response, and do not require a fixed trial length. The present findings could benefit from the simultaneous recording of scalp EEG which could allow a useful comparison with existing literature on the same topic [24] , [75].

This will be the focus of future studies including a larger set of patients. Intracranial EEG recordings in humans provide a unique opportunity to investigate spatially localized neural activity with a high temporal resolution, in particular for deep regions that cannot be easily accessible to surface EEG recordings. These data can thus provide an intermediate level of observation linking animal cell-recording data and human neuroimaging findings.

However, much like cell-recordings in animals, iEEG studies have a sparse distribution of recordings sites. In this study, we propose a multivariate decoding strategy to optimize the use of such distributed neural signals by allowing the full analysis of datasets from individual patients at the single-trial level and by offering an unbiased test for the contribution of individual electrodes to the observed effects.

We also thank Laurent Spinelli for technical support and the patients for participating in the study. Sleep parameters. Description of the sleep parameters of the night between the two recordings for each of the two patients.

Talairach coordinates of the implanted electrodes. List of the normalized Talaraich coordinates of the implanted electrodes for each of the two patients.

We report here all those coordinates that were included in the multivariate decoding analysis. Competing Interests: The authors have declared that no competing interests exist.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. National Center for Biotechnology Information , U.

PLoS One. Published online Dec 9. Nicole Wenderoth, Editor. Author information Article notes Copyright and License information Disclaimer.

Received Feb 23; Accepted Nov Copyright De Lucia et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are properly credited. This article has been cited by other articles in PMC. Table S1: Talairach coordinates of the implanted electrodes. Abstract We propose and validate a multivariate classification algorithm for characterizing changes in human intracranial electroencephalographic data iEEG after learning motor sequences.

Introduction In functional neuroimaging studies, the application of machine learning techniques has recently become a popular method for decoding stimulus-related information at the level of the single response to external stimuli [1] , [2]. Materials and Methods Experimental Paradigm Ethics Statement Patients provided written informed consent to participate in this study, which was approved by the ethical committee of the Geneva University Hospitals.

Patients description We tested epileptic patients who had depth-electrodes implanted in several brain regions for presurgical evaluation purposes. Open in a separate window. Figure 1. Illustration of the experimental procedure.

Behavioral task and experimental procedure The patients were tested on a serial reaction time task SRTT [24] , [25]. Figure 2. Classification procedure. Data analysis We first analyzed iEEG data from all implanted regions using a classification method based on single-trial responses.

Hidden Markov Model of single-trial iEEG To model the neural response in the S-sequence, we used a selection of contacts for each of the two patients, choosing the most distant contacts on each electrode array.

Accuracy estimation To select which HMM provided the best discrimination power between day 1 and day 2, we tested the models ten times. Estimating contacts with higher classification power We aimed at testing whether the classification performance was equally driven by all the recording sites across the available electrode arrays or whether some sites might contribute more to the classification.

Relation between classification accuracy and parameters of the HMM models One important step of multivariate decoding in neuroimaging studies is the identification of the features that the classifier exploits for achieving a good classification performance. Results Behavior For each patient, we first averaged the reaction times RT over the 8 key presses of each sequence in each block i.

Figure 3. Classification results. Figure 4. Localization results. Figure 5. Visualization of the areas most contributing to the classification. Parameters of the HMM model underlying classification accuracy We found that the values of the covariances were the most informative for discriminating day 1 and day 2 and that none of the other parameters did significantly impact the classification accuracy.

Figure 6. Discussion Intracranial EEG recordings in humans provide spatially-localized measurements of brain activity with a high temporal resolution. Learning-related changes the stability of momentary neural states The present study provides evidence of learning-related changes of intracranial electrical activity corresponding to single events individual keypresses belonging to a visuomotor sequence. Distributed versus localized neural correlates underlying sequence learning Our results show that the underlying neural correlates of sequence learning is distributed within a large network, involving activities from multiple locations.

Advantages and limitations of multivariate decoding approach to single-trial iEEG The present study demonstrates that the proposed multivariate decoding approach to single-trial iEEG datasets from individual patients offers an important methodological alternative to group studies of averaged data. Conclusions Intracranial EEG recordings in humans provide a unique opportunity to investigate spatially localized neural activity with a high temporal resolution, in particular for deep regions that cannot be easily accessible to surface EEG recordings.

Supporting Information Material S1 Sleep parameters. DOC Click here for additional data file. Table S1 Talairach coordinates of the implanted electrodes. Footnotes Competing Interests: The authors have declared that no competing interests exist.

References 1. Haynes JD. Multivariate decoding and brain reading: introduction to the special issue. Brain decoding: Opportunities and challenges for pattern recognition.

Pattern Recognition. Beyond mind-reading: multi-voxel pattern analysis of fMRI data. Trends Cogn Sci. Theoretical, statistical, and practical perspectives on pattern-based classification approaches to the analysis of functional neuroimaging data.

J Cogn Neurosci. Machine learning classifiers and fMRI: a tutorial overview. Introduction to machine learning for brain imaging.

Single-trial analysis and classification of ERP components - A tutorial. Brain-based decoding of human voice and speech. Predicting human brain activity associated with the meanings of nouns. Sparse estimation automatically selects voxels relevant for the decoding of fMRI activity patterns. Decoding of emotional information in voice-sensitive cortices.

Curr Biol. Sound categories are represented as distributed patterns in the human auditory cortex. Unraveling the distributed neural code of facial identity through spatiotemporal pattern analysis. Predicting the recognition of natural scenes from single trial MEG recordings of brain activity.

Quality of evidence for perceptual decision making is indexed by trial-to-trial variability of the EEG. Identifying object categories from event-related EEG: toward decoding of conceptual representations.

Decoding stimulus-related information from single-trial EEG responses based on voltage topographies. Cross-frequency coupling supports multi-item working memory in the human hippocampus. Modulation of face processing by emotional expression and gaze direction during intracranial recordings in right fusiform cortex.

Differential electrophysiological response during rest, self-referential, and non—self-referential tasks in human posteromedial cortex. Single subject EEG analysis based on topographic information. International Journal of Bioelectromagnetism. Principles of Topographic Analyses for Electrical Neuroimaging. In: Handy TC, editor. Andres FG, Gerloff C. Coherence of sequential movements and motor learning. J Clin Neurophysiol.

Neuronal activity in the primate supplementary, pre-supplementary and premotor cortex during externally and internally instructed sequential movements. Neurosci Res.

Westchester, Illinois: American Academy of Sleep Medicine. Neural Syst Rehabil Eng. Hidden Markov models for online classification of single trial EEG data.

Pattern Recognition Letters. Comparison of different feature classifiers for brain computer interfaces Proc. Rabiner LR. A tutorial on Hidden Markov Models and selected applications in speech recognition. Proceedings of the IEEE. Bishop CM. Neural Networks for Pattern Recognition. Maximum likelihood from incomplete data via the EM algorithm.

A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains. Ann Math Statist. Assessment of diagnostic technologies. The neural correlates of implicit and explicit sequence learning: Interacting networks revealed by the process dissociation procedure.

Learn Mem. Implicit oculomotor sequence learning in humans: Time course of offline processing. Brain Res. Practice with sleep makes perfect: sleep-dependent motor skill learning. Experience-dependent changes in cerebral activation during human REM sleep. Nat Neurosci. Sleep-related consolidation of a visuomotor skill: brain mechanisms as assessed by functional magnetic resonance imaging.

J Neurosci. Lehmann D. Principles of spatial analysis. Handbook of electroencephalography and clinical neurophysiology.

Amsterdam: Elsevier; Electric source imaging of human brain functions. Brain Res Brain Res Rev. EEG source imaging. Clin Neurophysiol. Suppose the association to be learned is from pattern A to pattern B , but when A is activated from the outside, a small part, C 1, of a third pattern, C , is also activated. When this state is fol- lowed by the activation of B , both rules strengthen the synapses from A to B and from C 1 to B. Then, when C is activated from the outside, it elicits a portion of B , which inter- feres with the learning of the correct C to D association.

The reason why the pre-synaptic rule does better at high K is that false activations such as C 1 are suppressed to begin with. Of course, the post- synaptic rule also enhances some false associations, and to see why these matter less, consider the situation when the activation of A is followed by the activation of the correct pattern, B , and a part, C 2, of pattern C.

However, this hardly matters, since A will never be re-activated from the outside and its false association does not interfere with learning. Thus, one main conclusion from our simulations is that the postsynaptic rule is better for one-step learning in noisy situa- tions.

This rule might be even better when learning sequences of patterns with overlap, or when learning multiple sequences which share pat- terns. How- ever, very high K works against such spontaneous replay by squelching activity. Thus, the post-synaptic rule, which learns well at lower K , also has the best impression performance.

Conclusion The results of this brief comparative study indicate that a learning rule with post-synaptic gating can be used to rapidly embed pattern sequences in the dynamics of relatively sparse recurrent networks. Future stimulus sequences can then overwrite earlier ones, thus requiring little capacity from the small network. References: Hopfield, J. Neural networks and physical systems with emergent collective computational abilities, Proc. Levy, W. Synapses as associative memory elements in the hippocampal formation, Brain Research, Associative encoding at synapses, Proc.

Cognitive Sci. Minai, A. The dynamics of sparse random networks, In review. Setting the activity level in sparse random networks, In review. Mulkey, R. Mechanisms underlying induction of homosynaptic long-term depression in area CA1 of the hippocampus, Neuron, 9: Sompolinsky, H. Temporal association in asymmetric neural networks, Physical Rev.

Squire, L. The medial temporal lobe memory system, Science, The learning rates are: 0. Graph a shows the situation before train- ing and graph b after training. Peaks indicate spontaneous emergence of sequence fragments.

Latent attractors: a model for context-dependent place representations in the hippocampus By Ali Minai. Latent attractors: a general paradigm for context-dependent neural computation By Ali Minai. Unstable dynamics, nonequilibrium phases, and criticality in networked excitable media By Joaquin Marro. Covariance learning of correlated patterns in competitive networks By Ali Minai. Download PDF. Authors: Advanced Search Include Citations.

Minai , William B. Abstract While recurrent neural networks can store pattern sequences through incremental learning, there could be a trade-off between network capacity and the speed of learning. Keyphrases single trial sequence learning pattern sequence long-term storage rapid temporary storage single exposure recurrent neural network learning rule key factor two-stage system low-capacity subsystem network capacity high learning rate rapid learning ability incremental learning recurrent inhibition post-synaptic gating.