Performance scalability

Performance of perClass 3.0 considering moderately-large problems.

Time of k-NN classifier execution

Up to 2.1x faster than previous generation.
Up to 10.9x faster than Matlab-based implementation.


k-NN classifier (k=5) was trained on moderate data set with 5700 samples, 11 features and two classes. The time of classifier execution (obtaining decisions on new data) is measured on a data set with 10000 samples. Experiments were performed in Matlab 2010b on 2.8GHz Intel Core Duo 2 machine (Mac OS X 10.5.8). Tests were repeated 10 times reporting the best attained performance.

perClass performance is compared with the leading academic toolbox for pattern recognition PRTools to illustrate the performance scalability of perClass C-based solution compared to Matlab-only implementation.

Time of PCA projection execution

Up to 1.5x faster than previous generation.
Up to 49x faster than Matlab-based implementation.


PCA dimensionality reduction to 3D subspace is trained on a two-class data set with 20000 samples and 20 features. The time needed for execution of the projection on new data is measured on a data set with 100000 samples. Experiments were performed in Matlab 2010b on 2.8GHz Intel Core Duo 2 machine (Mac OS X 10.5.8). Tests were repeated 10 times reporting the best attained performance. For PRTools, default wait-bar GUI is switched off using prwaitbar off to reach faster response.

Time of PCA dimensionality reduction training

Up to 1.28x slower than previous generation.


PCA dimensionality reduction to 3D subspace is trained on a two-class data set with 20000 samples and 20 features. The time needed for training the dimensionality-reduction projection is measured. Experiments were performed in Matlab 2010b on 2.8GHz Intel Core Duo 2 machine (Mac OS X 10.5.8). Tests were repeated 10 times reporting the best attained performance. For PRTools, default wait-bar GUI is switched off using prwaitbar off to reach faster response.

Note, that perClass sacrifices some training speed to gain faster execution.

Time of k-means clustering

Up to 45x faster than Matlab-based implementation.


Data set with 5700 samples in 11D was clustered with k-means algorithm to 10 clusters. The number of iterations was fixed to 50 for all algorithms. Experiments were performed in Matlab 2010b on 2.8GHz Intel Core Duo 2 machine (Mac OS X 10.5.8). Tests were repeated 10 times reporting the best attained performance. For PRTools, default wait-bar GUI is switched off using prwaitbar off to reach faster response.