Restricted Boltzman Machine

What is a Restricted Boltzmann Machine?

A Restricted Boltzmann Machine (RBM) is a generative stochastic artificial neural network that can learn a probability distribution over its set of inputs. RBMs were initially invented under the name Harmonium by Paul Smolensky in 1986, and they gained prominence in the 2000s due to Geoffrey Hinton and collaborators, who developed efficient training methods for them.

RBMs are interesting in that they are capable of learning to represent complex, high-dimensional data in a lower-dimensional space. This makes them valuable for dimensionality reduction, classification, regression, collaborative filtering, feature learning, and topic modeling.

Structure of a Restricted Boltzmann Machine

RBMs are a variant of Boltzmann machines, with the restriction that their neurons must form a bipartite graph: a two-layer structure consisting of visible input units and hidden units. This restriction allows for more efficient training algorithms, particularly the gradient-based contrastive divergence algorithm.

In an RBM, there are connections between the visible and hidden units but no connections within a layer. This is different from other neural networks, where neurons within a layer can have connections. The lack of intra-layer connections in RBMs simplifies the learning process because there are fewer dependencies to consider.

Training an RBM

The training of RBMs involves adjusting the weights and biases of the network to find patterns that allow the network to reconstruct the input data. The learning process is unsupervised, which means that no labeled data is necessary. Instead, RBMs use a cost function that measures the difference between the original data and its reconstruction from the hidden layer.

The most common training algorithm for RBMs is contrastive divergence (CD), which is a form of Markov Chain Monte Carlo (MCMC). CD approximates the gradient of the log-likelihood of the training data with respect to the model's parameters and uses this to update the weights and biases.

Energy-Based Model

RBMs are energy-based models. They assign an energy level to each configuration of the visible and hidden units. The network's goal during training is to adjust the weights and biases to lower the energy of configurations that represent the training data and increase the energy of configurations that do not.

The probability distribution of an RBM is defined by its energy function, and the probability that the network assigns to a particular configuration decreases exponentially with its energy. This means that configurations with lower energy are more probable.

Applications of Restricted Boltzmann Machines

RBMs have been applied in a variety of fields due to their ability to automatically discover and learn the representations needed for pattern recognition, classification, and regression tasks. Here are some of the applications:

  • Dimensionality Reduction: RBMs can be used to reduce the dimensionality of data, which is particularly useful in preprocessing steps for other algorithms that perform poorly with high-dimensional data.
  • Feature Learning:

    RBMs can learn features that can be useful as inputs for other machine learning algorithms, improving their performance on tasks like classification.

  • Collaborative Filtering: RBMs can be used to predict user preferences in recommendation systems, by learning the underlying structure in user-item interaction data.
  • Classification:

    After training, the hidden layer of an RBM can serve as a feature detector that can be used in combination with a classifier to categorize inputs.

  • Deep Belief Networks:

    Stacking RBMs can create deep architectures, known as Deep Belief Networks (DBNs), which can model complex data with multiple levels of abstraction.

Challenges and Considerations

While RBMs have many interesting properties and applications, they also come with challenges. Training RBMs can be tricky due to issues like the difficulty in choosing the right learning rate, the potential for the model to get stuck in poor local optima, and the computational cost associated with training on large datasets.

Furthermore, RBMs have been somewhat overshadowed by other neural network architectures like Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) in many applications, especially in tasks that involve structured data like images and sequences. However, RBMs still hold a place in the toolbox of machine learning techniques due to their unique properties and capabilities.


Restricted Boltzmann Machines are powerful neural networks capable of learning complex distributions and extracting useful features from data. Their structure and training methods make them suitable for a variety of unsupervised and semi-supervised learning tasks. Despite the rise of other neural network models, RBMs continue to be a topic of research and application due to their versatility and the richness of their theoretical foundations.


Smolensky, P. (1986). Information processing in dynamical systems: Foundations of harmony theory. Colorado Univ at Boulder Dept of Computer Science.

Hinton, G. E. (2002). Training products of experts by minimizing contrastive divergence. Neural computation, 14(8), 1771-1800.

Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm for deep belief nets. Neural computation, 18(7), 1527-1554.

Please sign up or login with your details

Forgot password? Click here to reset