To provide the most recent news and documentation www.pymvpa.org reflects the
development 2.0 series (renamed 0.6 series) of PyMVPA. If you are interested in the
documentation of the previous stable 0.4 series of PyMVPA, please
The literature concerning the application of multivariate pattern analysis
procedures to neuro-scientific datasets contains a lot of specific terms to
refer to procedures or types of data, that are of particular importance.
Unfortunately, sometimes various terms refer to the same construct and even
worse these terms do not necessarily match the terminology used in the machine
learning literature. The following glossary is an attempt to map the various
terms found in the literature to the terminology used in this manual.
Averaging all samples recorded during a block of continuous stimulation in a
block-design fMRI experiment. The rationale behind this technique is, that
averaging might lead to an improved signal-to-noise ratio. However,
averaging further decreases the number of samples in a dataset, which is
already very low in typical fMRI datasets, especially in comparison to the
number of features/voxels. Block-averaging might nevertheless improve the
classifier performance, if it indeed improves signal-to-noise and the
respective classifier benefits more from few high-quality samples than
from a larger set of lower-quality samples.
A model that maps an arbitrary feature space into a discrete set of
An internal to PyMVPA term to describe a classifier which is usually a
proxy to the main classifier which it wraps to provide additional data
preprocessing (e.g. feature selection) before actually training and/or
testing of the wrapped classifier.
A technique to assess the generalization of the constructed model
by the analysis of accuracy of the model predictions on presumably
A chunk is a group of samples. In PyMVPA chunks define independent groups
of samples (note: the groups are independent from each other, not the
samples in each particular group). This information is important in the
context of a cross-validation procedure, as it is required to measure the
classifier performance on independent test datasets to be able to compute
unbiased generalization estimates. This is of particular importance in the
case of fMRI data, where two successively recorded volumes cannot be
considered as independent measurements. This is due to the significant
temporal forward contamination of the hemodynamic response whose correlate
is measured by the MR scanner.
An attribute of a learner which might be enabled or disabled,
grouped within .ca attributes collection. If enabled, it might cause
additional computation and memory consumption, so the “heaviest”
conditional attributes are disabled by default.
Visualization of the generalization performance of a
classifier. Each row of the matrix represents the instances in a
predicted class, while each column represents the samples in an
actual (target) class. Each cell provides a count of how many
samples of the target class were (mis)classifier into the
corresponding class. In PyMVPA instances of
ConfusionMatrix class provide not only
confusion matrix itself but a bulk of additional statistics.
An arbitrary auxiliary information that is stored in a dataset.
This term is usually used to refer to the application of machine learning or
pattern recognition techniques to brainimaging datasets, and therefore is
another term for MVPA. Sometimes also ‘brain-reading’ is used as
Sometimes used to refer to a group of successively acquired samples, and,
thus, related to a chunk.
A variable that represents a dimension in a dataset. This might be
the output of a single sensor, such as a voxel, or a refined measure
reflecting specific aspect of data, such as a specific spectral
Analogous to a sample attribute, this is a per-feature vector of
auxiliary information that is stored in a dataset.
A technique that targets detection of features relevant to a given
problem, so that their selection improves generalization of the
This acronym stands for functional magnetic resonance imaging.
An ability of a model to perform reliably well on any novel data in
the given domain.
A label is a special case of a target for specifying discrete
categories of samples in a classification analyses.
A model that upon training given some data (samples and may be
targets) develops an ability to map an arbitrary feature
space of samples into another space. If targets were
provided, such learner is called supervised and tries to achieve
mapping into the space of targets. If the target space defined by
a set of discrete set of labels, such learner is called a
A field of Computer Science that aims at constructing methods, such
as classifiers, to integrate available knowledge extracted from
This term originally stems from the authors of the Matlab MVPA toolbox, and
in that context stands for multi-voxel pattern analysis (see Norman
et al., 2006). PyMVPA obviously adopted this acronym. However, as
PyMVPA is explicitly designed to operate on non-fMRI data as well, the
‘voxel’ term is not appropriate and therefore MVPA in this context stands
for the more general term multivariate pattern analysis.
Neural Data Modality
A reflection of neural activity collected using some available
instrumental method (e.g., EEG, fMRI).
Most objects dealing with data are implemented as processing objects. Such
objects are instantiated once, with all appropriate parameters
configured as desired. When created, they can be used multiple times by
simply calling them with new data.
A sample is a vector with observations for all feature variables.
A per-sample vector of auxiliary information that is stored in a
dataset. This could, for example, be a vector identifying specific
chunks of samples.
A sensitivity is a score assigned to each feature with respect to
its impact on the performance of the learner. So, for a classifier,
sensitivity of a feature might describe its influence on generalization
performance of the classifier. In case of linear classifiers, it could
simply be coefficients of separating hyperplane given by weight
vector. There exist additional scores which are similar to sensitivities
in terms of indicating the “importance” of a particular feature –
examples are a univariate anova score or a
A vector of several sensitivity scores – one for each feature in a
A target associates each sample in the dataset with a
certain category, experimental condition or, in case of a regression
problem, with some metric variable. In case of supervised learning
algorithm targets define the model to be trained, and provide the “ground
truth” for assessing the model’s generalization performance.
This usually refers to the block-averaging of samples from a
block-design fMRI dataset.