China
Africa
Explore chapters and articles related to this topic
Prescriptive and Predictive Analytics Techniques for Enabling Cybersecurity
Published in Kuan-Ching Li, Beniamino DiMartino, Laurence T. Yang, Qingchen Zhang, Smart Data, 2019
Nitin Sukhija, Sonny Sevin, Elizabeth Bautista, David Dampier
Deep Belief Networks (DBN), a variety of DNN, are made up of layers of hidden and visible layers. The most common way of constructing a DBN is stacking RBMs [15]. DBNs can be used to monitor network and user activity for threats by analyzing and flagging behavior that is out of the norm in real time [15]. The Boltzmann Machines come in several different flavors, but in general, they are a network of stochastic neurons that make decisions based on weighted connections. The RBMs are named so because the connections in a layer are not connected to each other but rather only to the layers before and after. Each layer in a RBM is trained sequentially and then stacked on top of each other using unlabeled training data and a selected training algorithm [16]. This allows for each layer to be finely tuned and increases efficiency over training all the layers at the same time [16–18]. The main challenge in using DBNs is balancing accuracy, training times, and efficiency [18].
Electronic Realization
Published in Bahrain Nabet, Robert B. Pinter, Sensory Neural Networks: Lateral Inhibition, 2017
Bahrain Nabet, Robert B. Pinter
A research group at Bell Communications Research has been among the first to concentrate on implementing learning networks. Alspector and Allen (1987) proposed the implementation of a Boltzmann machine. Boltzmann machines are recurrent networks in which each “neuron” thresholds a linear sum of its inputs but fires according to a probabilistic decision rule. A global minimum, as opposed to local minima of associative memory systems, and its corresponding weights determine a learned category. Boltzmann machines use noise to escape from local minima (incorrect representations), and reach global minima by varying a “temperature” parameter which is initially set high and then reduced as the system approaches global minima; hence the term simulated annealing. In this implementation thermal circuit noise is amplified to produce the annealing schedule. A 7 × 7 mm chip using 2 micron CMOS design rules was projected to require 250,000 transistors to produce a 40 neuron Boltzmann machine.
Symmetric Weights and Deep Belief Networks
Published in Stephen Marsland, Machine Learning, 2014
The Boltzmann machine is a very interesting neural network, particularly because it produces a generative probabilistic model (which is what is sampled from for the weight updates). However, it is computationally very expensive since every learning step involves Monte Carlo sampling from the two distributions. In fact, since it is based on symmetric connections the Boltzmann machine is actually a Markov Random Field (MRF) (see Section 16.2). The computational expense has meant that the normal Boltzmann machine has never been popular in practical use, which is why we haven’t examined an implementation of it. However, a simplification of it has, and we will consider that next.
Hybrid deep learning controller for nonlinear systems based on adaptive learning rates
Published in International Journal of Control, 2023
Ahmad M. El-Nagar, Ahmad M. Zaki, F. A. S. Soliman, Mohammad El-Bardini
In the study by Qiao et al. (2018), the nonlinear system was modelled based on self-organising deep belief network with growing and pruning algorithms. In the study by Qiu et al. (2014; Yu & De la Rosa, 2019), the researchers introduced a deep Boltzmann machine based on the probability distributions of the inputs/outputs for modelling the nonlinear systems and for data regression and time series modelling. In the study by Chen et al. (2018), DL was introduced for prediction wind speed forecasting based on extremely optimisation algorithm, support vector regression machine and long short term memory NNs. At last, the DL was applied in medicine, such as classification for COVID-19 infected patients (Pathak et al., 2020), electrocardiogram (ECG) data (Murat et al., 2020), computed tomography images data for predicting the risk for overall survival (Zhang et al., 2020).
Analysis of early fault vibration detection and analysis of offshore wind power transmission based on deep neural network
Published in Connection Science, 2022
Boyu Yang, Anmin Cai, Weirong Lin
The characteristic of a deep belief network is that each layer of it captures the complex hidden nodes associated with higher commands in the layer below it. The Deep Boltzmann machine is very interesting in many ways. First of all, like deep belief networks, deep Boltzmann machines have the potential to learn more and more complex statements inside the model, which is a very promising way to solve problems such as objective things and speech recognition. Secondly, the in-depth statement can be from a large supply of unrepresented sensory input data and very limited data that has been identified, and these data can only be slightly changed later, a good model is established, and then go immediately Complete some special tasks. The most fundamental feature of a deep neural network is its own topological organisation. The weight update of the neurone subset formed by topological correlation is very similar. In addition, the neurones in the subset derived from the learning process are also different. Figure 5 depicts the signal-to-noise ratio as a function of noise intensity.
Identifying nonlinear variation patterns with deep autoencoders
Published in IISE Transactions, 2018
Phillip Howard, Daniel W. Apley, George Runger
An example of the network architecture learned during the pre-training phase is provided in Figure 3. In this example, three hidden layers with n1, n2, and n3 hidden nodes are fully connected to nodes in the layers above and below, but are not connected to nodes within the same layer. The final encoded layer contains one node for each of the K expected variation patterns in the data (K = 2 in this example). During pre-training, sigmoid activation functions at each of the nodes in the network are used to make a stochastic decision about the state of the hidden node. Each layer in the network and its inputs from the layer below can be viewed as a Restricted Boltzmann Machine, which allows us to greedily pre-train the weights one layer at a time using contrastive divergence (Hinton et al., 2006).