DOCSLIB.ORG

Decoding Dynamic Brain Patterns from Evoked Responses: a Tutorial on Multivariate Pattern Analysis Applied to Time Series Neuroimaging Data

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Decoding Dynamic Brain Patterns from Evoked Responses: a Tutorial on Multivariate Pattern Analysis Applied to Time Series Neuroimaging Data Decoding Dynamic Brain Patterns from Evoked Responses: A Tutorial on Multivariate Pattern Analysis Applied to Time Series Neuroimaging Data Tijl Grootswagers1,2,3, Susan G. Wardle1,2, and Thomas A. Carlson2,3 Abstract ■ Multivariate pattern analysis (MVPA) or brain decoding aim is to “decode” different perceptual stimuli or cognitive states methods have become standard practice in analyzing fMRI data. over time from dynamic brain activation patterns. We show that Although decoding methods have been extensively applied in decisions made at both preprocessing (e.g., dimensionality reduc- brain–computer interfaces, these methods have only recently tion, subsampling, trial averaging) and decoding (e.g., classifier been applied to time series neuroimaging data such as MEG selection, cross-validation design) stages of the analysis can sig- and EEG to address experimental questions in cognitive neuro- nificantly affect the results. In addition to standard decoding, science. In a tutorial style review, we describe a broad set of we describe extensions to MVPA for time-varying neuroimaging options to inform future time series decoding studies from a data including representational similarity analysis, temporal gener- cognitive neuroscience perspective. Using example MEG data, alization, and the interpretation of classifier weight maps. Finally, we illustrate the effects that different options in the decoding we outline important caveats in the design and interpretation of analysis pipeline can have on experimental results where the time series decoding experiments. ■ INTRODUCTION 2011; Lemm, Blankertz, Dickhaus, & Müller, 2011), the The application of “brain decoding” methods to the anal- aims of time series decoding for cognitive neuroscience ysis of fMRI data has been highly influential over the past are distinct from those that drive the application of these 15 years in the field of cognitive neuroscience (Kamitani methods in BCI, thus requiring a targeted introduction. &Tong,2005;Carlson,Schrater,&He,2003;Cox& Although there are many reviews and tutorials for fMRI Savoy, 2003; Haxby et al., 2001; Edelman, Grill-Spector, decoding (Haynes, 2015; Schwarzkopf & Rees, 2011; Kushnir, & Malach, 1998). In addition to their increased Mur, Bandettini, & Kriegeskorte, 2009; Pereira, Mitchell, & sensitivity, the introduction of fMRI decoding methods Botvinick, 2009; Formisano, De Martino, & Valente, 2008; offered the possibility to address questions about infor- Haynes & Rees, 2006; Norman, Polyn, Detre, & Haxby, 2006; mation processing in the human brain, which have comple- Cox & Savoy, 2003), there are no existing tutorial intro- mented traditional univariate analysis techniques. Although ductions to decoding time-varying brain activity. Although decoding methods for time series neuroimaging data the approaches are conceptually similar, there are impor- such as MEG/EEG have been extensively applied in brain– tant distinctions that stem from fundamental differences in computer interfaces (BCI; Müller et al., 2008; Curran & the nature of the neuroimaging data between fMRI and Stokes, 2003; Wolpaw, Birbaumer, McFarland, Pfurtscheller, MEG/EEG. In this article, we provide a tutorial introduction & Vaughan, 2002; Kübler, Kotchoubey, Kaiser, Wolpaw, & using an example MEG data set. Although there are many Birbaumer, 2001; Farwell & Donchin, 1988; Vidal, 1973), possible analyses targeting time series data (e.g., oscillatory they have only recently been applied in cognitive neuro- [Jafarpour, Horner, Fuentemilla, Penny, & Duzel, 2013] or science (Carlson, Hogendoorn, Kanai, Mesik, & Turret, induced responses), we restrict the scope of this article to 2011; Duncan et al., 2010; Schaefer, Farquhar, Blokland, decoding information from evoked responses, with statis- Sadakata, & Desain, 2010). tical inference at the group level on single time points or The goal of this article is to provide a tutorial style guide small time windows. As with most neuroimaging analysis to the analysis of time series neuroimaging data for cog- techniques, the number of possible permutations for a nitive neuroscience experiments. Although introductions given set of analysis decisions is very large, and the par- to BCI exist (Blankertz, Lemm, Treder, Haufe, & Müller, ticular choice of analysis pipeline is guided by the experi- mental question at hand. Here we aim to provide a broad 1Macquarie University, Sydney, Australia, 2ARC Centre of Excel- demonstration of how the analysis may be approached, lence in Cognition and its Disorders, 3University of Sydney rather than prescribing a particular analysis pipeline. ©2017MassachusettsInstituteofTechnology JournalofCognitiveNeuroscience29:4,pp.677–697 doi:10.1162/jocn_a_01068 Early studies using time-resolved decoding methods brain activity was recorded.Thegoalofthedecoding have revealed significant potential for experimental in- analysis is to test whether we can predict if the partici- vestigation using this approach with MEG/EEG (see pant was viewing a blue circle or a red square based on MVPA for MEG/EEG section). However, compared with their patterns of brain activation. If the experimental the popularity of decoding methods in fMRI, to date only stimuli can be successfully “decoded” from the partici- a small number of studies have applied multivariate pat- pant’spatternsofbrainactivation,wecanconcludethat tern analysis (MVPA) techniques to EEG or MEG. Accord- some information relevant to the experimental manipula- ingly, the aims of this article are to (a) introduce the critical tion exists in the neuroimaging data. First, brain activation differences between decoding time series (e.g., MEG/EEG) patterns in response to the different stimuli (or experi- versus spatial (e.g., fMRI) neuroimaging data, (b) illustrate mental conditions) are recorded using standard neuro- the time series decoding approach using a practical tutorial imaging (MEG, fMRI, etc.) techniques (Figure 1A). The with example MEG data, (c) demonstrate the effect that activation levels of the variables (e.g., voxels in fMRI, selecting different analysis parameters has on the results, channels in MEG/EEG) in different experimental conditions and (d) outline important caveats in the interpretation of are represented as complex patterns in high-dimensional time series decoding studies. In summary, this article will space (each voxel, channel, or principal component is one provide a broad overview of available methods to inform dimension). For simplicity, in Figure 1B, these patterns are future time-resolved decoding studies. This tutorial is pre- shown in two-dimensional space. Each point in the plot sented in the context of MEG; however, the methods and represents an experimental observation corresponding to analysis principles generalize to other time-varying brain the simultaneous activation level in two example voxels/ recording techniques (e.g., ECoG, EEG, electrophysiol- channels in response to one of the experimental condi- ogical recordings.). As this review is targeted at providing tions (blue circles or red squares). a broad overview to a general audience, we avoid formal The first step in a decoding analysis involves training mathematical definitions and implementation details of aclassifiertoassociatebrainactivationpatternswith the methods and instead focus on the rationale behind the experimental conditions using a subset of the data the decoding approach as applied to time series data. (Figure 1C). In effect, during training the classifier finds the decision boundary in higher-dimensional space that best separates the patterns of brain activation correspond- MVPA for MEG/EEG ing to the two experimental categories into two distinct The term “multivariate pattern analysis” (or MVPA) en- groups. As neuroimaging data are inherently noisy, this compasses a diverse set of methods for analyzing neuro- separation is not necessarily perfect (note the red square imaging data. The common element that unites these on the wrong side of the decision boundary in Figure 1C). approaches is that they take into account the relation- Next, the trained classifier is used to predict the condition ships between multiple variables (e.g., voxels in fMRI or labels for new data that were not used for training the channels in MEG/EEG), instead of treating them as inde- classifier (Figure 1D). The classifier predicts whether pendent and measuring relative activation strengths. The the new (unlabeled) data are more similar to the pattern term “decoding” refers to the prediction of a model from of activation evoked by viewing a blue circle or a red the data (“encoding” approaches do the reverse, predict- square. If the classifier performs higher than that expected ing the data from the model, reviewed in Naselaris, Kay, by chance (in this case 50% is the guessing rate as there Nishimoto, & Gallant, 2011; see also, e.g., Ding & Simon, are two stimuli), it provides evidence that the classifier 2012, for an example of encoding models for MEG). The can successfully generalize the learned associations to most common application of decoding in cognitive neu- labeling new brain response patterns. Consequently, it is roscience is the use of machine learning classifiers (e.g., assumed that the patterns of brain activation contain infor- correlation classifiers (Haxby et al., 2001) or discriminant mation that distinguishes between the experimental con- classifiers (Carlson et al., 2003; Cox & Savoy, 2003) to ditions (i.e., the conditions blue circle/red square can be identify patterns
Load more

Recommended publications
  • Local Field Potential Decoding of the Onset and Intensity of Acute Pain In
    www.nature.com/scientificreports OPEN Local feld potential decoding of the onset and intensity of acute pain in rats Received: 25 January 2018 Qiaosheng Zhang1, Zhengdong Xiao2,3, Conan Huang1, Sile Hu2,3, Prathamesh Kulkarni1,3, Accepted: 8 May 2018 Erik Martinez1, Ai Phuong Tong1, Arpan Garg1, Haocheng Zhou1, Zhe Chen 3,4 & Jing Wang1,4 Published: xx xx xxxx Pain is a complex sensory and afective experience. The current defnition for pain relies on verbal reports in clinical settings and behavioral assays in animal models. These defnitions can be subjective and do not take into consideration signals in the neural system. Local feld potentials (LFPs) represent summed electrical currents from multiple neurons in a defned brain area. Although single neuronal spike activity has been shown to modulate the acute pain, it is not yet clear how ensemble activities in the form of LFPs can be used to decode the precise timing and intensity of pain. The anterior cingulate cortex (ACC) is known to play a role in the afective-aversive component of pain in human and animal studies. Few studies, however, have examined how neural activities in the ACC can be used to interpret or predict acute noxious inputs. Here, we recorded in vivo extracellular activity in the ACC from freely behaving rats after stimulus with non-noxious, low-intensity noxious, and high-intensity noxious stimuli, both in the absence and chronic pain. Using a supervised machine learning classifer with selected LFP features, we predicted the intensity and the onset of acute nociceptive signals with high degree of precision.
    [Show full text]
  • Improving Whole-Brain Neural Decoding of Fmri with Domain Adaptation
    This is a repository copy of Improving whole-brain neural decoding of fMRI with domain adaptation. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/136411/ Version: Submitted Version Article: Zhou, S., Cox, C. and Lu, H. orcid.org/0000-0002-0349-2181 (Submitted: 2018) Improving whole-brain neural decoding of fMRI with domain adaptation. bioRxiv. (Submitted) https://doi.org/10.1101/375030 © 2018 The Author(s). For reuse permissions, please contact the Author(s). Reuse Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of the full text version. This is indicated by the licence information on the White Rose Research Online record for the item. Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing [email protected] including the URL of the record and the reason for the withdrawal request. [email protected] https://eprints.whiterose.ac.uk/ Improving Whole-Brain Neural Decoding of fMRI with Domain Adaptation Shuo Zhoua, Christopher R. Coxb,c, Haiping Lua,d,∗ aDepartment of Computer Science, the University of Sheffield, Sheffield, UK bSchool of Biological Sciences, the University of Manchester, Manchester, UK cDepartment of Psychology, Louisiana State University, Baton Rouge, Louisiana, USA dSheffield Institute for Translational Neuroscience, Sheffield, UK Abstract In neural decoding, there has been a growing interest in machine learning on whole-brain functional magnetic resonance imaging (fMRI).
    [Show full text]
  • Deep Learning Approaches for Neural Decoding: from Cnns to Lstms and Spikes to Fmri
    Deep learning approaches for neural decoding: from CNNs to LSTMs and spikes to fMRI Jesse A. Livezey1,2,* and Joshua I. Glaser3,4,5,* [email protected], [email protected] *equal contribution 1Biological Systems and Engineering Division, Lawrence Berkeley National Laboratory, Berkeley, California, United States 2Redwood Center for Theoretical Neuroscience, University of California, Berkeley, Berkeley, California, United States 3Department of Statistics, Columbia University, New York, United States 4Zuckerman Mind Brain Behavior Institute, Columbia University, New York, United States 5Center for Theoretical Neuroscience, Columbia University, New York, United States May 21, 2020 Abstract Decoding behavior, perception, or cognitive state directly from neural signals has applications in brain-computer interface research as well as implications for systems neuroscience. In the last decade, deep learning has become the state-of-the-art method in many machine learning tasks ranging from speech recognition to image segmentation. The success of deep networks in other domains has led to a new wave of applications in neuroscience. In this article, we review deep learning approaches to neural decoding. We describe the architectures used for extracting useful features from neural recording modalities ranging from spikes to EEG. Furthermore, we explore how deep learning has been leveraged to predict common outputs including movement, speech, and vision, with a focus on how pretrained deep networks can be incorporated as priors for complex decoding targets like acoustic speech or images. Deep learning has been shown to be a useful tool for improving the accuracy and flexibility of neural decoding across a wide range of tasks, and we point out areas for future scientific development.
    [Show full text]
  • Neural Decoding of Collective Wisdom with Multi-Brain Computing
    NeuroImage 59 (2012) 94–108 Contents lists available at ScienceDirect NeuroImage journal homepage: www.elsevier.com/locate/ynimg Review Neural decoding of collective wisdom with multi-brain computing Miguel P. Eckstein a,b,⁎, Koel Das a,b, Binh T. Pham a, Matthew F. Peterson a, Craig K. Abbey a, Jocelyn L. Sy a, Barry Giesbrecht a,b a Department of Psychological and Brain Sciences, University of California, Santa Barbara, Santa Barbara, CA, 93101, USA b Institute for Collaborative Biotechnologies, University of California, Santa Barbara, Santa Barbara, CA, 93101, USA article info abstract Article history: Group decisions and even aggregation of multiple opinions lead to greater decision accuracy, a phenomenon Received 7 April 2011 known as collective wisdom. Little is known about the neural basis of collective wisdom and whether its Revised 27 June 2011 benefits arise in late decision stages or in early sensory coding. Here, we use electroencephalography and Accepted 4 July 2011 multi-brain computing with twenty humans making perceptual decisions to show that combining neural Available online 14 July 2011 activity across brains increases decision accuracy paralleling the improvements shown by aggregating the observers' opinions. Although the largest gains result from an optimal linear combination of neural decision Keywords: Multi-brain computing variables across brains, a simpler neural majority decision rule, ubiquitous in human behavior, results in Collective wisdom substantial benefits. In contrast, an extreme neural response rule, akin to a group following the most extreme Neural decoding opinion, results in the least improvement with group size. Analyses controlling for number of electrodes and Perceptual decisions time-points while increasing number of brains demonstrate unique benefits arising from integrating neural Group decisions activity across different brains.
    [Show full text]
  • Provably Optimal Design of a Brain-Computer Interface Yin Zhang
    Provably Optimal Design of a Brain-Computer Interface Yin Zhang CMU-RI-TR-18-63 August 2018 The Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 Thesis Committee: Steven M. Chase, Carnegie Mellon University, Chair Robert E. Kass, Carnegie Mellon University J. Andrew Bagnell, Carnegie Mellon University Patrick J. Loughlin, University of Pittsburgh Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Robotics. Copyright c 2018 Yin Zhang Keywords: Brain-Computer Interface, Closed-Loop System, Optimal Control Abstract Brain-computer interfaces are in the process of moving from the laboratory to the clinic. These devices act by reading neural activity and using it to directly control a device, such as a cursor on a computer screen. Over the past two decades, much attention has been devoted to the decoding problem: how should recorded neural activity be translated into movement of the device in order to achieve the most pro- ficient control? This question is complicated by the fact that learning, especially the long-term skill learning that accompanies weeks of practice, can allow subjects to improve performance over time. Typical approaches to this problem attempt to maximize the biomimetic properties of the device, to limit the need for extensive training. However, it is unclear if this approach would ultimately be superior to per- formance that might be achieved with a non-biomimetic device, once the subject has engaged in extended practice and learned how to use it. In this thesis, I first recast the decoder design problem from a physical control system perspective, and investigate how various classes of decoders lead to different types of physical systems for the subject to control.
    [Show full text]
  • Neural Networks for Efficient Bayesian Decoding of Natural Images From
    Neural Networks for Efficient Bayesian Decoding of Natural Images from Retinal Neurons Nikhil Parthasarathy∗ Eleanor Batty∗ William Falcon Stanford University Columbia University Columbia University [email protected] [email protected] [email protected] Thomas Rutten Mohit Rajpal E.J. Chichilniskyy Columbia University Columbia University Stanford University [email protected] [email protected] [email protected] Liam Paninskiy Columbia University [email protected] Abstract Decoding sensory stimuli from neural signals can be used to reveal how we sense our physical environment, and is valuable for the design of brain-machine interfaces. However, existing linear techniques for neural decoding may not fully reveal or ex- ploit the fidelity of the neural signal. Here we develop a new approximate Bayesian method for decoding natural images from the spiking activity of populations of retinal ganglion cells (RGCs). We sidestep known computational challenges with Bayesian inference by exploiting artificial neural networks developed for computer vision, enabling fast nonlinear decoding that incorporates natural scene statistics implicitly. We use a decoder architecture that first linearly reconstructs an image from RGC spikes, then applies a convolutional autoencoder to enhance the image. The resulting decoder, trained on natural images and simulated neural responses, significantly outperforms linear decoding, as well as simple point-wise nonlinear decoding. These results provide a tool for the assessment and optimization of reti- nal prosthesis technologies, and reveal that the retina may provide a more accurate representation of the visual scene than previously appreciated. 1 Introduction Neural coding in sensory systems is often studied by developing and testing encoding models that capture how sensory inputs are represented in neural signals.
    [Show full text]
  • Using Deep Learning for Visual Decoding and Reconstruction from Brain Activity: a Review
    Using Deep Learning for Visual Decoding and Reconstruction from Brain Activity: A Review Madison Van Horn The University of Edinburgh Informatics Forum, Edinburgh, UK, EH8 9AB [email protected] Abstract This literature review will discuss the use of deep learning methods for image recon- struction using fMRI data. More specifically, the quality of image reconstruction will be determined by the choice in decoding and reconstruction architectures. I will show that these structures can struggle with adaptability to various input stimuli due to complicated objects in images. Also, the significance of feature representation will be evaluated. This paper will conclude the use of deep learning within visual decoding and reconstruction is highly optimal when using variations of deep neural networks and will provide details of potential future work. 1 Introduction The study of the human brain is a continuous growing interest especially for the fields of neuroscience and artificial intelligence. However, there are still many unknowns to the human brain. Recent advancements of the combination of these fields allow for better understanding between the brain and visual perception. This notion of visual perception uses neural data in order to make predictions of stimuli [1]. The predictions of stimuli will allow researchers to not only see if we are capable of reconstructing visual thoughts but consider the accuracy with these predictions. This reconstruction is achievable due to the ability of using neural data from tools such as functional magnetic resonance imaging (fMRI) to assist in recording brain activity to later use for decoding and reconstructing a stimulus.
    [Show full text]
  • Neural Networks for Efficient Bayesian Decoding Of
    Neural Networks for Efficient Bayesian Decoding of Natural Images from Retinal Neurons Nikhil Parthasarathy∗ Eleanor Batty∗ William Falcon Stanford University Columbia University Columbia University [email protected] [email protected] [email protected] Thomas Rutten Mohit Rajpal E.J. Chichilniskyy Columbia University Columbia University Stanford University [email protected] [email protected] [email protected] Liam Paninskiy Columbia University [email protected] Abstract Decoding sensory stimuli from neural signals can be used to reveal how we sense our physical environment, and is valuable for the design of brain-machine interfaces. However, existing linear techniques for neural decoding may not fully reveal or ex- ploit the fidelity of the neural signal. Here we develop a new approximate Bayesian method for decoding natural images from the spiking activity of populations of retinal ganglion cells (RGCs). We sidestep known computational challenges with Bayesian inference by exploiting artificial neural networks developed for computer vision, enabling fast nonlinear decoding that incorporates natural scene statistics implicitly. We use a decoder architecture that first linearly reconstructs an image from RGC spikes, then applies a convolutional autoencoder to enhance the image. The resulting decoder, trained on natural images and simulated neural responses, significantly outperforms linear decoding, as well as simple point-wise nonlinear decoding. These results provide a tool for the assessment and optimization of reti- nal prosthesis technologies, and reveal that the retina may provide a more accurate representation of the visual scene than previously appreciated. 1 Introduction Neural coding in sensory systems is often studied by developing and testing encoding models that capture how sensory inputs are represented in neural signals.
    [Show full text]
  • Neural Decoding for Motor and Communication Prostheses 7
    CHAPTER Neural Decoding for Motor and Communication Prostheses 7 Byron M. Yu∗†∗∗, Gopal Santhanam∗, Maneesh Sahani∗∗, Krishna V. Shenoy∗†∗∗∗ ∗Department of Electrical Engineering, †Neurosciences Program, Stanford University, Stanford, California, ∗∗Gatsby Computational Neuroscience Unit, UCL, London, ∗∗∗Department of Bioengineering, Stanford University 7.1 INTRODUCTION Neural prostheses, which are also termed brain–machine and brain–computer interfaces, offer the potential to substantially increase the quality of life for people suffering from motor disorders, including paralysis and amputation. Such devices translate electrical neural activity from the brain into control signals for guiding paralyzed upper limbs, prosthetic arms, and computer cursors. Several research groups have now demonstrated that monkeys (e.g., Serruya et al., 2002; Taylor et al., 2002; Carmena et al., 2003; Musallam et al., 2004; Santhanam et al., 2006; Mulliken et al., 2008; Velliste et al., 2008) and humans (e.g., Kennedy et al., 2000a; Leuthardt et al., 2004; Wolpaw and McFarland, 2004b; Hochberg et al., 2006; Kim et al., 2008; Schalk et al., 2008) can learn to move computer cursors and robotic arms to various target locations simply by activating neural populations that participate in natural arm movements. Figure 7.1 illustrates the basic operating principle of neural prosthetic systems. Neural activity from various arm movement–related brain regions (e.g., Wise et al., 1997) is electronically processed to create control signals for enacting the desired movement. Non invasive sensors can collect neural signals representing the average activity of many neurons. When invasive permanently implanted arrays of electrodes are employed, as depicted, it is possible to use waveform shape differences to discriminateindividual neurons (e.g., Lewicki, 1998; Santhanam et al., 2004).
    [Show full text]
  • Decoding the Neural Correlates of Consciousness Rimona S
    Decoding the neural correlates of consciousness Rimona S. Weila,b and Geraint Reesa,b aWellcome Trust Centre for Neuroimaging at UCL, Purpose of review Institute of Neurology, University College London and bUCL Institute of Cognitive Neuroscience, University Multivariate pattern analysis (MVPA) is an emerging technique for analysing functional College London, London, UK imaging data that is capable of a much closer approximation of neuronal activity than Correspondence to Dr Rimona Weil, Wellcome Trust conventional methods. This review will outline the advantages, applications and Centre for Neuroimaging at UCL, 12 Queen Square, limitations of MVPA in understanding the neural correlates of consciousness. London WC1N 3BG, UK Tel: +44 20 7833 7472; fax: +44 20 7813 1420; Recent findings e-mail: r.weil@fil.ion.ucl.ac.uk MVPA has provided important insights into the processing of perceptual information by Current Opinion in Neurology 2010, 23:649–655 revealing content-specific information at early stages of perceptual processing. It has also shed light on the processing of memories and decisions. In combination with techniques to reconstruct viewed images, MVPA can also be used to reveal the contents of consciousness. Summary The development of multivariate pattern analysis techniques allows content-specific and detailed information to be extracted from functional MRI data. This may lead to new therapeutic applications but also raises important ethical considerations. Keywords consciousness, decoding, fMRI, multivariate pattern analysis Curr Opin Neurol 23:649–655 ß 2010 Wolters Kluwer Health | Lippincott Williams & Wilkins 1350-7540 way is not always possible when communication or Introduction insight is limited. Therefore techniques allowing direct Consciousness is the rich, constantly changing internal determination of mental content through measuring experience which makes us who we are.
    [Show full text]
  • Visual Encoding and Decoding of the Human Brain Based on Shared Features
    Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI-20) Visual Encoding and Decoding of the Human Brain Based on Shared Features Chao Li*1, Baolin Liu*†2 and Jianguo Wei1 1 College of Intelligence and Computing, Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin University, Tianjin, China 2 School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, China Abstract The framework proposed in this paper has two parts (see Figure 1), the first of which is “encoding.” In this part, we Using a convolutional neural network to build visual assume that only some of the features of the neural network encoding and decoding models of the human brain have a clear relationship with human brain signals, which are is a good starting point for the study on relationship called shared features. They are obtained by projecting the between deep learning and human visual cognitive features of low and high layers of the neural network into mechanism. However, related studies have not fully their respective shared spaces. Then a non-negative sparse considered their differences. In this paper, we as- map is established between the shared features and brain ac- sume that only a portion of neural network features tivity to achieve the encoding process. The second part of the is directly related to human brain signals, which we framework is “decoding,” in which visual stimuli are recon- call shared features. In the encoding process, we ex- structed through neural network visualization technology, tract shared features from the lower and higher lay- meanwhile, dictionary learning and sparse representation are ers of the neural network, and then build a non-neg- introduced to improve the robustness of the algorithm.
    [Show full text]
  • Neural Decoding of Cursor Motion Using a Kalman Filter
    Neural Decoding of Cursor Motion using a Kalman Filter W. Wu∗ M. J. Blacky Y. Gao∗ E. Bienenstock∗x M. Serruyax{ A. Shaikhounix{ J. P. Donoghuex ∗Division of Applied Mathematics, yDept. of Computer Science, xDept. of Neuroscience, {Division of Biology and Medicine, Brown University, Providence, RI 02912 [email protected], [email protected], [email protected], [email protected], Mijail [email protected], Ammar [email protected], john [email protected] Abstract The direct neural control of external devices such as computer displays or prosthetic limbs requires the accurate decoding of neural activity rep- resenting continuous movement. We develop a real-time control system using the spiking activity of approximately 40 neurons recorded with an electrode array implanted in the arm area of primary motor cortex. In contrast to previous work, we develop a control-theoretic approach that explicitly models the motion of the hand and the probabilistic re- lationship between this motion and the mean firing rates of the cells in 70ms bins. We focus on a realistic cursor control task in which the sub- ject must move a cursor to “hit” randomly placed targets on a computer monitor. Encoding and decoding of the neural data is achieved with a Kalman filter which has a number of advantages over previous linear filtering techniques. In particular, the Kalman filter reconstructions of hand trajectories in off-line experiments are more accurate than previ- ously reported results and the model provides insights into the nature of the neural coding of movement. 1 Introduction Recent results have demonstrated the feasibility of direct neural control of devices such as computer cursors using implanted electrodes [5, 9, 11, 14].
    [Show full text]