DictionaryLearning¶
Find a sparse representation of the data using Dictionary Learning.
This method will learn components, that is linear combinations of input features, such that the transformed data has feature whose activations are maximally sparse. This is also known as sparse coding, and is closely related to Independent Component Analysis (ICA). As such, it can be used for identifying statistically independent or otherwise meaningful and/or interpretable features. Therefore, these are features that can be highly useful in subsequent processing stages, for instance non-linear feature extraction or sparse machine learning techniques. In contrast to PCA or ICA, this method can easily learn more features than there were data dimensions in the input data. Also, unlike most other feature extractors, this method will attempt to estimate the the transformed data in accordance with the sparse modeling assumption, instead of simply linearly transforming it (at some extra computational cost). This node offers multiple algorithms both for estimating the model, and then for transforming (reconstructing) the output data given the model. Important:This node is adaptive to the data, that is, it will learn a transformation of the data that depends on the input data. In order to learn this transformation, the node requires a reasonable amount of input data for calibration or "training" (otherwise it will yield an ill-fitting or noisy model). Since this feature extraction method is not capable of being trained incrementally on streaming data, the method requires a data packet that contains the entire training data; this training data packet can either be accumulated online and then released in one shot using the Accumulate Calibration Data node, or it can be imported from a separate calibration recording and then spliced into the processing pipeline using the Inject Calibration Data, where it passes through the same nodes as the regular data until it reaches the machine learning node, where it is used for calibration. Once this node is calibrated, the trainable state of this node can be saved to a model file and later loaded for continued use. Like most other feature extraction nodes, this node can compute features between elements of an axis of your choice while treating elements of another axis as the observations, trials, or samples. It can also optionally compute multiple separate models on different slices of the data along some axis of choice. It is also possible to pool multiple axes for any of these roles. More Info... Version 1.0.0
Ports/Properties¶
domain_axes¶
Axes which form the input domain of the transformation. Features are computed between elements along these axes (or in other words, elements along these axes will be combined with each other to yield features). This is a comma-separated list of axis names (for example: "space, frequency"), possibly empty, or the special string "(all others)", which stands for all axes that are not listed in the other two lists of axes. For time-series data, this is usually the space axis, and if features have already been extracted from the data through some other method, it would be the features axis. In rare cases it can also include other axes, such as frequency, lag, and time. This axis drops out of the data and is replaced by a feature axis that has a number of elements that corresponds to the output feature dimension of the model.
aggregate_axes¶
Axes that have the statistical observations in them. The elements along these axes are treated as the "trials", "samples", or, equivalently, "observations". Adaptive feature extractors will aggregate statistics along the elements of these axes during training. When the transformer applies to data, this axis remains in the data (if multiple, they will be flattened into one axis, however). This is a comma-separated list of axis names (for example: "time, instance"), possibly empty, or the special string "(all others)", which stands for all axes that are not listed in the other two lists of axes. This is almost always the instance axis (especially if the data has already been segmented, i.e., if the Segmentation node was used), but in some cases it may also be the time axis, or occasionally other axes.
separate_axes¶
Axes along which to learn separate models. It is possible to use multiple separate feature-extraction models, each of which operates on a different slice of the data. This node does not combine data between elements along these axes in any way (although features between these elements may of course be combined in later stages, for instance in a classifier node). This is a comma-separated list of axis names (for example: "time, frequency"), possibly empty, or the special string "(all others)", which stands for all axes that are not listed in the other two lists of axes. This axis, if any, will also persist in the data.
data¶
Data to process.
num_components¶
Number of components to keep. If left unspecified, all components are kept, that is, the number of output features will correspond to the number of input dimensions.
alpha¶
Degree of sparsity. Higher values yield components with more sparse activity.
transform_alpha¶
Degree of sparsity during transformation. Does not apply to 'lars' case.
max_iter¶
Maximum number of iterations. This is one of the stopping criteria to limit the compute time. The default is usually fine, and gains from increasing the number of iterations will be minimal (it can be worth experimenting with lower iteration numbers if the algorithm must finish in a fixed time budget, at a cost of potentially less accurate solutions).
num_jobs¶
Number of parallel compute jobs. This value only affects the running time and not the results. Values between 1 and twice the number of CPU cores make sense to expedite computation, but may temporarily reduce the responsiveness of the machine. The value of -1 stands for all available CPU cores.
verbosity¶
Verbosity level. Higher numbers will produce more extensive diagnostic output.
transform_nonzeroes¶
Targeted number of non-zeroes per column in the solution. Only used by the lars method.
tolerance¶
Convergence tolerance. This is the desired errors tolerance or acceptable inaccuracy in the solution. Using larger values gives less accurate results, but will lead to faster compute times. Note that, for biosignal-driven machine learning systems, one often does not need very small tolerances.
fit_algorithm¶
Method used during fitting. The lars method is faster than coordinate descent if the components are sparse.
transform_algorithm¶
Method used during transform. lasso_lars is fastest when components are sparse.
split_sign¶
Split the sparse feature vector into the concatenation of its negative and positive part.
random_seed¶
Random seed. Different values may yield slightly different results.
set_breakpoint¶
Set a breakpoint on this node. If this is enabled, your debugger (if one is attached) will trigger a breakpoint.
metadata¶
User-definable meta-data associated with the node. Usually reserved for technical purposes.