I am setting up a Radial basis Neural Network (RBNN). Most applications have used gaussian radial basis function. However I plotted the distribution of my output and noticed it to be bimodal. Will this have an effect?
Hello Kwadwo,
What do you mean with "output data"?
In case you mean that the input to the NN is multimodal, this should be unproblematic for RBNNs. The idea is that the weighted sum of a high number of RBF kernels can resemble any (multimodal or not) distribution function. Illustratively, this is similar to resembling a unknown distribution function as a sum of e.g. Gaussians in Kernel density estimation (the samples in KDE would correspond to neurons in RBNN, if there was no weighting).
Michael
Hello Kwadwo,
Switching from RBF to GPs (Gaussian Processes) point of view, follow:
- Svensson et al., " Marginalizing Gaussian Process Hyperparameters Using Sequential Monte Carlo ", 2016 - https://arxiv.org/pdf/1502.01908.pdf
- Zand et al., " MiDGaP: Mixture Density Gaussian Processes ", 2017 - http://www.robots.ox.ac.uk/~sjrob/Pubs/nips_2017.pdf
- Alvarez et al, " Computationally Efficient Convolved Multiple Output Gaussian Processes ", 2011 - http://www.jmlr.org/papers/volume12/alvarez11a/alvarez11a.pdf
- Skolidis et al., " Bayesian Multitask Classification with Gaussian Process Priors ", 2011 - http://booksc.org/book/31435865/4185b9
Regards
- Badges
- Science topic