Restricted Boltzmann Machine
Robert H. Chen, Chelsea Chen · 2022
A Restricted Boltzmann Machine (RBM) is an early artificial neural network with only an input layer composed of vectors v, one hidden layer composed of vectors h, and no output layer. An RBM, starting from a random initial distribution, compares its distribution with the distribution of the input data; the difference is just the free energy of the candidate distribution, so just as in other artificial neural networks, minimizing that free energy by backpropagation will cause the RBM-generated probability distribution to converge to the probability distribution of the input data, and thus reveal the ground truth probabilities and the latent inferences hidden therein. An RBM is typically used in unsupervised learning initially to first model the unknown input distribution, and RBMs can be stacked to deep learn. An RBM can serve as a preprocessor for convolutional neural networks performing image and text recognition, and in conjunction with recurrent neural networks, also do speech recognition.