[pymvpa] [mvpa-toolbox] Fast ER design (Dr. Aguirre please help ; -) )

Yaroslav Halchenko yarikoptic at gmail.com
Tue Jul 12 16:41:19 UTC 2011

disclaimer: I've not played/used debruijn sequences myself yet.  I am
   CCing Dr. Aguirre seeking for the ultimate support/disprove of my
   claim (feel free to reply to me directly -- I will channel your 
   response to the lists which might reject post without subscription).
   Question was about the best choice of timing for trials for MVPA on fast
   ER design.  I forwarded to your paper, and now claiming...

if you look at that page into some paragraphs, they suggest that for the
best detection power trade-off between number of trials, SOA etc would
be experiment/design specific, thus could/should be evaluated by trying
different values for those given a specific experiment in mind.  Here
are the relevant excerpts from wiki. I believe the paper has more:

Guide function

Any valid de Bruijn cycle will provide the specified level of
counter-balance amongst the labels, and thus stimuli, of the experiment. Not
all orderings of stimuli, however, are equally useful for neuroimaging
experiments. Because of the signal and noise properties of fMRI, some temporal
frequencies of neural modulation are preferentially detected by the method. The
path-guided approach to de Bruijn cycle generation encodes in the ordering of
the stimuli a hypothesized neural modulation at these preferred frequencies.

To do so, we define a guide function in one of three ways:

    Enter the word HRF. A guide function will be internally generated
    that has the same power spectrum as the BOLD hemodynamic response function,
    although with no power below 0.01 Hz. This will generally produce a sequence
    with good detection power and a stochastic variation in stimulus transitions,
    although further improvements can be obtained by using a guide function that is
    positioned solely at a few or a single frequency.


    Enter a range of periods (in units of labels), as described by
    [Tmin,Tmax]. A guide function will be internally generated as the sum of
    sinusoids of random phase, each having a period equal to an integer between
    Tmin and Tmax. Enter the same value for Tmin and Tmax to guide the modulation
    at a single frequency. Generally, when (1e5 / SOA in ms) > T > (1e4 / SOA in
    ms) encoded modulations will be within a detectable range for the BOLD system.


    Provide the path to an external file, which contains kn 
    space-separated floating point values. Each element will correspond to a
    relative transition in the output sequence. The elements in the guide function
    may be either positive or negative; the vector will be normalized.

For fMRI, the optimal range of temporal frequencies is ~0.01-0.1 Hz2).
Higher frequencies are attenuated by the dispersed hemodynamic response, while
lower frequencies are lost within the pink (1/f) noise of the system. The
interpretation of the path-guided sequence in Hz requires a specification of
the stimulus-onset asynchrony, described next.

Evaluation mode and SOA

Given a particular hemodynamic response function, and a model of the 1/f
noise of the BOLD fMRI system, detection power3) is the proportion of neural
variance which appears in the imaging signal. It is calculated as the variance
of the hypothesized neural modulation proportional to sequential stimulus
distance as the denominator, and the numerator as the variance of that
modulation after passing through the BOLD fMRI system. In our implementation,
the BOLD system is modeled using a standard, population averaged hemodynamic
response4) and the elevated 1/f noise range by a 0.01 Hz, high-pass notch

When the -eval flag is set, the stimulus-onset asynchrony (SOA) parameter
is specified in units of milliseconds (i.e., the time that elapses between the
start of one stimulus and the start of the next stimulus). Along with the
path-guided de Bruijn sequence, the routine then returns both the detection
power, and the correlation coefficient of the guide function (input) with the
sequence of distances between stimuli generated (output).

On Tue, 12 Jul 2011, Jane Klemen wrote:

> Hi Yaroslav,
> Thanks a lot, that looks interesting.  From just a quick look at the
> abstract it looks like it's about stimulus order.  Any idea whether
> there is also anything out there on stimulus timing?
> Cheers,
> Jane

> On 12/07/2011 17:20, Yaroslav Halchenko wrote:
> >for fast ER MVPA it is crucial to have proper trials order, so have a
> >look at recent

> >GK Aguirre, MG Mattar, L Magis-Weinberg. (2011) de Bruijn cycles for neural decoding. NeuroImage 56: 1293-1300
> >http://www.ncbi.nlm.nih.gov/pubmed/21315160

> >with software
> >http://www.cfn.upenn.edu/aguirre/wiki/public:de_bruijn_software
> >readily available from Debian (and NeuroDebian):

> >sudo apt-get install debruijn

> >On Tue, 12 Jul 2011, Jane Klemen wrote:

> >>Hi all,
> >>We're wanting to run an experiment that, for psych reasons, needs to
> >>have a fast ER design.  Piloting shows that our sanity tests work,
> >>but, ideally, we'd like to optimise the design timing in order to have
> >>some chance of our conditions of interst also being classified
> >>correctly.  I've done a literature search and can't find anything for
> >>MVPA that is similar to the design optimimisation papers we all know
> >>for univariate analysis (Birn et al. 2002; Burock et al. 1998 etc.).
> >>Looking through the mailing list I also get the sense from some posts
> >>that there aren't any studies that address stimulus timing for fast ER
> >>designs for MVPA.  Is this correct, or have I just completely
> >>overlooked something?
> >>Thanks in advance for your feedback,
> >>Jane
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic

More information about the Pkg-ExpPsy-PyMVPA mailing list