![MQL5 - Language of trade strategies built-in the MetaTrader 5 client terminal](https://c.mql5.com/i/registerlandings/logo-2.png)
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Check out the new article: MQL5 Wizard Techniques you should know (Part 10). The Unconventional RBM.
Restrictive Boltzmann Machines are at the basic level, a two-layer neural network that is proficient at unsupervised classification through dimensionality reduction. We take its basic principles and examine if we were to re-design and train it unorthodoxly, we could get a useful signal filter.
Restrictive Boltzmann Machines (RBMs) are a form of neural network that are quite simple in their structure but are none the less revered, in certain circles, for what they can accomplish when it comes to revealing hidden properties and features in data-sets. This they accomplish by learning the weights in a smaller dimension from a larger dimensioned input data, with these weights often referred to as probability distributions. As always more reading can be done here, but usually their structure can be illustrated by the image below:
Typically, RBMs consist of 2 layers, (I say typically because there are some networks that stack them into transformers) a visible layer and a hidden layer with the visible layer being larger (having more neurons) than the hidden layer. Every neuron in the visible layer connects to each neuron in the hidden layer during what is referred to as the positive phase such that during this phase as is common with most neural networks, the input values at the visible layer gets multiplied by weight values at connecting neurons and the sum of these products gets added to a bias to determine the values at the respective hidden neurons. The negative phase which is the reverse of this is then what follows and through different neuron connections, it aims to restore the input data to its original state starting with the computed values in the hidden layer.
So, with the early cycles as one would expect the reconstructed input data is bound to mismatch the initial input because often the RBM is initialized with random weights. This implies the weights need to be adjusted to bring the reconstructed output closer to the input data and this is the additional phase that would follow each cycle. The end-result and objective of this cycle positive phase followed by a negative phase and weights adjustment, is to arrive at connecting neuron weights which when applied to input data, can give us ‘intuitive’ neuron values in the hidden layer. These neuron values in the hidden layer are what is referred to as the probability distribution of the input data across the hidden neurons.
Author: Stephen Njuki