ICompELM - Independent Component Analysis Based Extreme Learning Machine
Single Layer Feed-forward Neural networks (SLFNs) have
many applications in various fields of statistical modelling,
especially for time-series forecasting. However, there are some
major disadvantages of training such networks via the widely
accepted 'gradient-based backpropagation' algorithm, such as
convergence to local minima, dependencies on learning rate and
large training time. These concerns were addressed by Huang et
al. (2006) <doi:10.1016/j.neucom.2005.12.126>, wherein they
introduced the Extreme Learning Machine (ELM), an extremely
fast learning algorithm for SLFNs which randomly chooses the
weights connecting input and hidden nodes and analytically
determines the output weights of SLFNs. It shows good
generalized performance, but is still subject to a high degree
of randomness. To mitigate this issue, this package uses a
dimensionality reduction technique given in Hyvarinen (1999)
<doi:10.1109/72.761722>, namely, the Independent Component
Analysis (ICA) to determine the input-hidden connections and
thus, remove any sort of randomness from the algorithm. This
leads to a robust, fast and stable ELM model. Using functions
within this package, the proposed model can also be compared
with an existing alternative based on the Principal Component
Analysis (PCA) algorithm given by Pearson (1901)
<doi:10.1080/14786440109462720>, i.e., the PCA based ELM model
given by Castano et al. (2013) <doi:10.1007/s11063-012-9253-x>,
from which the implemented ICA based algorithm is greatly
inspired.
