Structured networks
This protocol is extracted from research article:
Deep neural network processing of DEER data
Sci Adv, Aug 24, 2018; DOI: 10.1126/sciadv.aat5218

Table 2 indicates that plain feedforward networks with more than six layers do not produce any further improvements in the performance. If those improvements are even possible, then more sophisticated topologies must be used. One possibility is shown in the bottom diagram of Fig. 3—the first group of layers was trained against the form factor and therefore eliminated noise and background. That form factor was then fed into the second group of layers, making the probability density extraction easier for those layers. In principle, structured networks may be assembled from pretrained pieces. In the case of the bottom diagram of Fig. 3, the pieces would come from one of the form factor extraction networks in Table 1 and a separate network trained to interpret background-free form factors. Performance figures for networks of this type are given in Table 4.

A schematic of the network topology is given in the bottom diagram of Fig. 3. FF, form factor; Int, interpretation.

Unfortunately, it does not appear that tailoring carries any advantages relative to the data reported for the simple feedforward networks in Table 2. Training a 12-layer network against two sets of outputs is also exceedingly expensive. We therefore used uniform feedforward networks (Fig. 3, top) for all production calculations discussed below. The networks were trained on a data set where raw experimental data without any preprocessing go in, and the distance distribution is expected at the output.

Still, the networks evaluated in Table 4 could potentially be beneficial as a safety catch: Humans can easily recognize incorrect form factors visually and thus detect cases of neural networks failing, for example, if they encounter a situation not covered by the training set.

Note: The content above has been extracted from a research article, so it may not display correctly.



Q&A
Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.



We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.