Academic Homepage

Pavan Ramkumar
Ph.D. Student

Brain Research Unit
Low Temperature Laboratory
Aalto University School of Science

AND

Department of Information and Computer Science
Aalto University School of Science

Contact: pavan@neuro.hut.fi
Curriculum Vitæ: [pdf]

For decades, computational modelers, artificial intelligence researchers, theoretical neuroscientists and experimental neurophysiologists (mainly single and multi-unit electrophysiology) have carefully developed rigorous models (and theories) of brain function that corroborate with experimental observations. These models particularly abound in the field of early sensory perception, and even more so, in vision. By contrast, such rigorous validation of theory by experiment and constraint of experimental design by theory, has not penetrated the world of non-invase functional brain imaging (EEG/ MEG/ fMRI), with few exceptions. A lack of theory in the neuroimaging world so far has been a legitimate shortcoming for various reasons, but I believe the time is right to attempt a change. Why do I believe so?

I believe that today we have (1) sufficient understanding of signal generation principles, (2) sufficiently advanced signal analysis methods, and (3) sufficient candidate theories of brain function from computational neuroscience or artificial intelligence that we may subject to rigorous validation. Paralleling these advances in knowledge, we are also experiencing rapid technological advances in (1) development of multimodal imaging methods, (2) fast and distributed computing, and (3) large-scale databasing of anatomical and functional datasets.

An influx of studies which test theories of brain function under the scanner, would simultaneously (1) extend the applicability of non-invasive brain imaging, (2) provide new constraints for comparing brain imaging at various scales, and (3) provide new constraints for theories of brain function. In the next decade of my research career, I wish to take a step in that direction.

However, all of the above is easier said than done and requires the acquisition of new skills and ideas over an extended period of time. Towards this end, I seek post-doctoral training under the guidance of an established scientist who considers my research goals to be in alignment with the larger directives of their lab. My ideal lab would constitute a good mix of computational modelers, cognitive scientists with a well-honed research question in a domain of brain function (I am equally open to sensory perception, action, or higher cognitive function), computer scientists, statisticians or machine-learners. I am equally open to working on existing datasets or acquiring my own datasets while learning a new imaging techique.

It goes without saying that my thesis advisors Prof. Riitta Hari and Prof. Aapo Hyvärinen have been important driving forces behind most of the work showcased here. Dr. Lauri Parkkonen has also played a significant advisory role in most of my projects.

Induced changes in oscillations are time-locked by not necessarily phase-locked to the onset of the stimulus or task. It has been reported that the amplitude envelope of a spontaneous oscillation is suppressed during task performance. With few exceptions, reports of the dynamics of suppression have been qualitative. To take a more quantiative approach, we proposed and developed an oscillatory response function (ORF) [Ramkumar et al., 2010]. Readers who are familiar with modeling the blood oxygenation level dependent (BOLD) response in functional magmetic resonance imaging (fMRI) signal analysis will readily spot the similarity to the hemodynamic response function (HRF). Hence our nomenclature.

The ORF quantitatively maps any arbitrary stimulus or task design to the envelope dynamics of an oscillation of interest, such as the rolandic 10- and 20-Hz mu rhythms. We learned a class of parametric ORF models in the generalized convolution (Volterra kernel) framework using Laguerre polynomials as basis kernels. From the learned parametric models, we predicted on a new MEG dataset, with greater accuracy than a boxcar function, the cortical regions at which the 20-Hz component of the mu rhythm manifests.

[Ramkumar et al., 2010] Ramkumar P, Parkkonen L, Hari R. 2010. Oscillatory Response Function: Towards a parametric model of rhythmic brain activity. Human Brain Mapping, 31:820--834.

Amplitude-modulated oscillatory activity as recorded by magnetoencephalograhy (MEG) encodes various aspects of stimulus and task information. Besides, it is hypothesized that narrowband envelope-corelations could be (a prime candidates for) the electrophysiological equivalent of the resting-state networks observed with funtional magnetic resonance imaging (fMRI). To facilitate data-driven exploration of oscillatory networks active during rest and natural stimulation, we are developing a series of methods to statistically decompose MEG oscillatory activity into functionally segregated space-time-frequency atoms at the cortical surface. To this end, we combine distributed source modeling (minimum norm estimation) and blind source separation (independent component analysis).

First, we proposed that applying complex-valued ICA to short-time Fourier transforms of magnetoencephalography (MEG) signals is likely to reveal physiologically more meaningful components than plain-vanilla temporal ICA [Hyvärinen et al., 2010]. Subsequently, we extended this work from the sensor space to source space and further argued the advantage of imposing spatial and spectral sparseness on the Fourier coefficients (spatial Fourier-ICA or SFICA) in relation to the cortically sparse and spectrally narrowband character of the oscillations [Ramkumar et al., 2012a]. Currently, we are applying real-valued ICA on the absolute values of the Fourier coefficients (Fourier energies) while imposing spatial and spectral sparseness as before [Ramkumar et al., 2012b]. Preliminary results suggest that the method seems to reveal long-range cortico-cortical anticorrelations in oscillatory power across frequencies.

[Hyvärinen et al., 2010] Hyvärinen A, Ramkumar P, Parkkonen L, Hari R. 2010. Independent component analysis of short-time Fourier transforms for spontaneous EEG/MEG analysis. Neuroimage, 49:257--271.
[Ramkumar et al., 2012a] Ramkumar P, Parkkonen L, Hari R, Hyvärinen A. 2011. Characterization of neuromagnetic brain rhythms over time scales of minutes using spatial independent component analysis. Human Brain Mapping, EPub ahead of print.
[Ramkumar et al., 2012b] Ramkumar P, Parkkonen L, Hyvärinen A. Group-level spatial independent component analysis of Fourier-envelopes of resting-state MEG data. Submitted.

It is well known that the early visual cortex is selective to low-level visual features such as orientation and spatial frequency (SF), and further that they are organized in columnar structures, as well as retinotopic maps. However, it is not clear to what extent such selectivity manifests in non-invasive electrophysiological recordings such as electro/ magnetoencephalography (EEG/ MEG). Here we approach the above question by solving a series of decoding problems.

We measured MEG data from 8 healthy adults acquired during the presentation of full-field parafoveal sinusoidal grating stimuli within an annulus spanning 2--10 degrees of the visual field. For static gratings, we attempted to decode from single-trial MEG responses, either orientation (OR), spatial frequency (SF), or both. For dynamic gratings which rotated either clockwise or anticlockwise, we attempted to decode the direction of rotation (RD). Our single-subject decoders (linear support vector machine (SVM) classifiers) were able to identify SF (93.4±3.1%), OR (72.0±6.4%), and RD (86.6±4.9%) well above chance. However, decoders trained by pooling data across all subjects except the target subject, and tested on the target subject data performed poorly as expected, with only SF (72.8±5.0%) performing above chance.

Critically, time-resolved classifiers trained on 20 ms response epochs, shifted by 1 ms, exceeded chance level as early as 33 ms for SFs, 47 ms for ORs and 85 ms for RDs. Accuracies stayed above chance level up to 300 ms for SFs and 200 ms for ORs, approximately towards the end of the evoked response and then returned to chance.

We also performed several control experiments on one healthy adult to show that spatial frequency can be decoded robustly despite variation of contrast, phase, foveal vs. parafoveal stimulation, or orientation type. Finally, we performed a control experiment to systematically study SF tuning by measuring MEG data during the presentation of gratings at 6 different SFs within annuli of 4 different sizes. Channel-level tuning curves revealed differential sensitivity to high and low SFs at low and high eccentricities. In particular, the classification of high spatial frequencies at large eccentricities 9--11 degrees performed below chance. Altogether, our results suggest that several low-level visual features are robustly encoded in single-trial electrophysiological population responses and these can be read out at fairly high temporal resolution.

The above experiments and analysis were performed in collaboration with Mainak Jas and Dr. Sebastian Pannasch.

[Ramkumar, 2012] Ramkumar P, Jas M, Pannasch S, Parkkonen L, Hari R. 2012. Uncovering the dynamics of low-level visual feature processing using time-resolved decoding of single-trial MEG responses. In Preparation.

Independent component analysis (ICA) is a popular technique to statistically factorize the matrix of functional imaging data (mostly fMRI) into spatio-temporal atoms, which can be loosely described as functionally segregated. To perform ICA across subjects (i.e. at the group level), a couple of different approaches exist. The first approach concatenates data across subjects after suitable spatial normalization and performs ICA on the concatenated data matrix. The second approach performs ICA on each subject separately and matches the functionally corresponding components across subjects. The latter approach makes fewer assumptions regarding inter-subject similarity, and is therefore preferred, provided that a foolproof method exists to solve the problem of identifying corresponding components across subjects. Let us refer to this problem as the the correspondence problem. Why is solving the correspondence problem important? Once functional correspondence is established across subjects, some second-level analysis can be done on the ICs: such as interrogating their timecourses for modulation to external stimuli, measuring inter-subject variability of the timecourses, or interrogating them for functional connectivity.

To measure functional correspondence, a heuristic approach is used, such as a similarity measure between spatial maps of the ICs, followed by a reasonable threshold. Now if each subject's brain activity as observed by a functional imaging method (say fMRI) resulted in a data matrix with identical higher order statistics, and further, if ICA factorized the data matrix into identical spatio-temporal atoms, our correspondence problem would be easy to solve. Specifically, a randomly chosen pair of components (one from each subject) would either be highly similar, or highly dissimilar, making the issue of assigning correspondence trivial. Thus under these ideal circumstances, a heuristic method would be sufficient.

As we know however, real data comes with inter-subject differences in anatomy, physiology and cognition. Further, ICA is a stochastic algorithm with algorithmic randomness, modeling inaccuracies and observation noise. Hence, instead of a heuristic approach, Hyvärinen [2011] suggested a principled approach, which starts by defining a null hypothesis and then constructing a statistical test around the estimated probability that two components (one from each of two subjects) are corresponding under the null hypothesis. Hyvärinen's null hypothesis deals with algorithmic randomness and this results in a good statistical test for temporal ICA. However for spatial ICA, it seems to be necessary to deal with individual differences as well. I am currently working on constructing a novel statistical test based on a new null hypothesis which takes individual differences into account. A working draft is available here.

[Hyvärinen, 2011] Hyvärinen A. 2011. Testing the ICA mixing matrix based on inter-subject or inter-session consistency. Neuroimage, 58:122--136.

The success of the Gallant group to decode natural images from fMRI data, has suggested that non-invasive brain imaging data contain important and reliable decodable information about the tuning properties of the primary visual cortex. Inspired by their success, we ask: is it possible to learn an encoding transformation from an analytic decomposition of natural images (such as a Gabor wavelet pyramid), to model single-trial MEG responses from a selection of occipital channels? To this end, we fit a support-vector regression model (enabling sparse nonlinear regression) between Gabor wavelet coefficients of natural images and their corresponding single-trial MEG responses. We then applied this regression model to predict the MEG responses to novel natural images, with the intention of identifying which image in a closed set evoked the measured MEG signal. Preliminary results suggest that image identification performs above chance, but much poorer than the results reported for fMRI by Kay et al. (2008).

This analysis evolved out of an approach envisioned for building an encoding model for movie data, outlined here. The stimuli for this study were designed and acquired by Dr. Sebastian Pannasch and collaborators.

In collaboration with a machine-learning group, we hosted an MEG data analysis contest for the International Conference on Artificial Neural Networks (ICANN) 2011. Participants were faced with the challenge of decoding the semantic category of one among five natural movie clips from 1-s blocks of MEG data during the free viewing of these clips. The results were presented by the top 3 participant teams at a workshop and a report summarizing the challenge and contributions is available here.

Along with Dr. Sebastian Pannasch, I serve as an instructor to Kranthi Kumar Nallamothu. With simultaneous MEG and eye-tracking data during the free viewing of natural movie clips, we are validating the recent observation by Nakano et al. [2009] that eye-blinks synchronize across subjects during the free viewing of silent films. Further, our goal is to study what brain processes precede synchronized blinks. Specificially, we ask: can MEG signals averaged with respect to synchronized blinks as triggers explain the cause of the synchrony?

Ramkumar P, Jas M, Pannasch S, Parkkonen L, Hari R. 2012. Uncovering the dynamics of low-level visual feature processing using time-resolved decoding of single-trial MEG responses. In Preparation.

Ramkumar P, Parkkonen L, Hyvärinen A. 2011. Group-level spatial independent component analysis of Fourier envelopes of resting-state MEG data. Under Revision.

Ramkumar P, Pannasch S, Hansen BC, Larson AM, Loschky LC. 2011. How does the brain represent visual scenes? A neuromagnetic scene categorization study. 2011. Neural Information Processing Systems (NIPS) --- Workshop on Machine Learning and Interpretation in Neuroimaging, to appear

Ramkumar P, Parkkonen L, Hari R, Hyvärinen A. 2011. Characterization of neuromagnetic brain rhythms over time scales of minutes using spatial independent component analysis. Hum Brain Mapp, EPub Ahead of Print.

Ramkumar P, Hyvärinen A, Parkkonen L, Hari R. Characterization of spontaneous neuromagnetic brain rhythms using independent component analysis of short-time Fourier transforms. Proceedings of the 17th International Conference on Biomagnetism, April 2010, Dubrovnik.

Hyvärinen A, Ramkumar P, Parkkonen L, Hari R. 2010. Independent component analysis of short-time Fourier transforms for spontaneous EEG/MEG analysis. Neuroimage, 49:257-271.

Ramkumar P, Parkkonen L, Hari R. 2010. Oscillatory Response Function: Towards a parametric model of rhythmic brain activity. Hum Brain Mapp, 31: 820-834.

Malinen S, Vartiainen N, Hlushchuk Y, Koskinen M, Ramkumar P, Forss N, Kalso E, Hari R. 2010. Aberrant spatiotemporal resting-state brain activation in patients with chronic pain. Proc Natl Acad Sci USA, 107: 6493-6497.

Ramkumar P, Parkkonen L, Hyvärinen A. Independent Component Analysis of Fourier Energies: characterizing long-range cortico-cortical interactions in magnetoencephalography (MEG) data. Society for Neuroscience, November 2011, Washington DC.

Hyvärinen A, Ramkumar P, Hari R. Advances in analysis of spontaneous EEG/MEG activity by independent component analysis. 29th International Congress on Clinical Neurophysiology, October 2010, Kobe.

Ramkumar P, Hyvärinen A, Parkkonen L, Hari R. Characterization of spontaneous neuromagnetic brain rhythms using spatial independent component analysis of short-time Fourier transforms. International Congress on Default Mode Network, June 2010, Barcelona.

Yokosawa K, Pamilo S, Hirvenkari L, Ramkumar P, Pihko E, Hari R. Activation of auditory cortex by anticipating and hearing emotional sounds: an MEG study. 16th Annual Meeting of the Organization for Human Brain Mapping, June 2010, Barcelona.
Nangini C, Ramkumar P, Hari R. SII neurons can phase-lock to trains of bilateral 4-Hz tactile stimuli. 16th Annual Meeting of the Organization for Human Brain Mapping, June 2010, Barcelona.
Mudigonda M, Ramkumar P, Zhu D, Stockman G, Jin R. Multivoxel pattern analysis identifies brain regions that discriminate indoor and outdoor scenes. 16th Annual Meeting of the Organization for Human Brain Mapping, June 2010, Barcelona.
Ramkumar P, Malinen S, Vartiainen N, Hlushchuk Y, Forss N, Kalso E, Hari R. Hub maps reveal reduced resting-state connectivity of insular cortex in patients with chronic pain. 16th Annual Meeting of the Organization for Human Brain Mapping, June 2010, Barcelona.
Hyvärinen A, Ramkumar P, Hari R. Selecting independent components by testing inter-subject reproducibility. 16th Annual Meeting of the Organization for Human Brain Mapping, June 2010, Barcelona.

Hyvärinen A, Zhang K, Ramkumar P, Hari R. Analyzing statistical dependencies of MEG source envelopes. 17th International Conference on Biomagnetism, April 2010, Dubrovnik.
Ramkumar P, Hyvärinen A, Parkkonen L, Hari R. Separating independent components of neuromagnetic brain rhythms by combining spatial and spectral sparseness. 17th International Conference on Biomagnetism, April 2010, Dubrovnik.

Ramkumar P, Parkkonen L. Characterization of the temporal structure of neuromagnetic rhythms using clustering and self-organizing maps. 2nd INCF Congress on Neuroinformatics, September 2009, Pilzen.

Parkkonen L, Ramkumar P, Hari R. A descriptive model of the dynamics of rhythmic brain activity. 1st INCF Congress on Neuroinformatics, September 2008, Stockholm.
Hyvärinen A, Parkkonen L, Ramkumar P, Hari R. A new method for unsupervised analysis of spontaneous MEG/EEG data: Combination of projection pursuit and parallel factor analysis. 1st INCF Congress on Neuroinformatics, September 2008, Stockholm.

Hyvärinen A, Parkkonen L, Ramkumar P, Hari R. Finding 'interesting' frequency bands in MEG using an unsupervised learning approach. 16th International Conference on Biomagnetism, August 2008, Sapporo.

Ramkumar P, Parkkonen L, Hari R. Oscillatory Response Functions: Towards a parametric model of rhythmic activity. PENS Spring School: Models in Neuroscience, April 2008, St. Petersburg.

Ramkumar P, Parkkonen L, He B, Raichle M, Hämäläinen M, Hari R. Identification of stimulus related and intrinsic networks by spatial independent component analysis of MEG signals. Society for Neuroscience, November 2007, San Diego.

Ramkumar P, Parkkonen L, Hari R. Independent component analysis of neuromagnetic data reveals extrinsic and intrinsic cortical networks during natural stimulation. Nordic Neuroinformatics meeting, October 2007, Helsinki.

Singhal GK, Ramkumar P. Person Identification using evoked potentials and peak matching. Biometrics Symposium, September 2007, Baltimore.

Ramkumar P, Singhal GK, Dandapat S. EEG Correlates of highly cognitive mental tasks for closed-set biometric authentication. International Biometric Conference, July 2006, Montreal.

An essay on neurovascular coupling and the blood oxygenation level dependent (BOLD) signal

A brief survey of the nascent fMRI decoding field

An essay reflecting on the possibility of a neuroscience theory in the distant future