Fuzzy Restricted Boltzmann Machine to Enhance Deep Learning

Student’s Name
Professor’s Name
Course
Date
Fuzzy Restricted Boltzmann Machine to Enhance Deep Learning
Lately, Deep learning has been an area of interest in machine learning mainly because of its effectiveness such as its capability in the pattern, speech, and image recognition as well as video processing. Restricted Boltzmann Machine (RBM) is a stochastic graph model which can learn a combined probability distribution over its n visible units x = [x1, · · ·, xn] and m hidden feature units h = [h1, · · ·, hm]. The model is overseen by parameter θ denoting joint weights and biases between cross-layer units, as illustrated in Figure 1 below.

RBM was developed in 1986, though, it was recently transitioned when G. Hinton and his colleagues proposed numerous deep network and corresponding fast learning algorithms, comprising deep autoencoder, deep belief networks, and deep Boltzmann machine. Dimensionality reduction, classification, collaborative filtering, feature learning as well as topic modeling are the main RBM’s applications and its deep architectures. Regular RBM has connections between the observable units and the concealed units, which are restricted to be constants.
The restriction will undoubtedly downgrade the demonstration capability of the RBM. To evade this flaw and improve deep learning capability, Chen et al. (2015) propose the FRBM and learning algorithm associated with the model. FRBM is proved to have a better generative and discriminative capabilities than the traditional RBM.

The parameters controlling the model are substituted with fuzzy numbers. Hence, the traditional RBM is an exception in the FRBM, when fuzziness is not present in the FRBM model.
During Fuzzy Restricted Boltzmann Machine learning, and prior to defining the probability, the fuzzy free energy function is usually defuzzified. According to the experimental results, the Bar-And-Stripe benchmark in-painting, as well as MNIST handwritten digits classification problems, indicate that the representation capability of Fuzzy Restricted Boltzmann Machine version is considerably better than the original Restricted Boltzmann Machine. Furthermore, the Fuzzy Restricted Boltzmann Machine showed enhanced robustness features as compared to RBM for cases where the training data are polluted by noises.

Work Cited
Chen, CL Philip, et al. “Fuzzy restricted Boltzmann machine for the enhancement of deep learning.” IEEE Transactions on Fuzzy Systems 23.6 (2015): 2163-2173.