Uniform feedforward 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

The simplest strategy for training a generic “vector-in, vector-out” neural network is to set up a number of fully connected layers of the same size as the input vector, resulting in the topology shown in the top diagram of Fig. 3. The performance metrics for a family of such networks are given in Table 2 and illustrated graphically in Figs. 4 and 5. The “relative error” metric is defined as the 2-norm of the difference between the network output and the true answer divided by the 2-norm of the true answer.

A schematic of the network topology is given in the top diagram of Fig. 3.

All inner layers have hyperbolic tangent transfer functions; the last layer has the strictly positive logistic sigmoid transfer function.

All layers have hyperbolic tangent transfer functions.

It is clear from the performance statistics that, for a single neural network, the average norm of the deviation drops below 10% of the total signal norm and stops improving once the network is five to six layers deep. Training iteration time depends linearly on the depth of the network.

The data for the visual performance illustrations (Figs. 4 and 5) were selected from the training database in the following way: the “easy case” was sampled from the relative error histogram region located between 0 and 1 SD; the “tough” case was sampled from the region between 1 and 2 SDs; the “bad case” was sampled from 100 worst fits in the entire 100,000-trace training database. Performance illustrations for the rest of the networks reported in Table 2 are given in figs. S1 to S3. Given that the bad cases are the worst 0.1% of the training data set, the performance is rather impressive. Similar sequential improvements are observed for the networks tasked with the recovery of the DEER form factor (Fig. 5).

For the vast majority of DEER traces in the training database, the recovery of the form factor is close to perfect. Performance illustrations for the rest of the form factor recovery networks reported in Table 2 are given in figs. S4 to S6.

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

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.