Must Know Tips/Tricks in Deep Neural Networks

Posted on Posted in News

This article was posted by Xiu-Shen Wei.  Xiu-Shen Wei is a 2nd-year Ph.D. candidate of Department of Computer Science and Technology in Nanjing University and a member of LAMDA Group.

Deep Neural Networks, especially Convolutional Neural Networks (CNN), allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction. These methods have dramatically improved the state-of-the-arts in visual object recognition, object detection, text recognition and many other domains such as drug discovery and genomics.

In addition, many solid papers have been published in this topic, and some high quality open source CNN software packages have been made available. There are also well-written CNN tutorials or CNN software manuals. However, it might lack a recent and comprehensive summary about the details of how to implement an excellent deep convolutional neural networks from scratch. Thus, they collected and concluded many implementation details for DCNNs.

Here they will introduce these extensive implementation details, i.e., tricks or tips, for building and training your own deep networks.

Table of Contents:

Introduction

Sec. 1: Data Augmentation

Sec. 2: Pre-Processing

Sec. 3: Initializations

  • All Zero Initialization
  • Initialization with Small Random Numbers
  • Calibrating the Variances
  • Current Recommendation

Sec. 4: During Training

  • Filters and pooling size.
  • Learning rate.
  • Fine-tune on pre-trained models.

Sec. 5: Activation Functions

  • Sigmoid
  • tanh(x)
  • Rectified Linear Unit
  • Leaky ReLU
  • Parametric ReLU
  • Randomized ReLU

Sec. 6: Regularizations

  • L2 regularization
  • L1 regularization
  • Max norm constraints
  • Dropout

Sec. 7: Insights from Figures

  • As we have known, the learning rate is very sensitive.
  • Now let’s zoom in the loss curve.
  • Another tip comes from the accuracy curve.

Sec. 8: Ensemble

  • Same model, different initialization.
  • Top models discovered during cross-validation.
  • Different checkpoints of a single model.
  • Some practical examples.

Miscellaneous

References & Source Links

To check out all this information, click here. For more articles about Neural Networks, click here.

DSC Resources

Additional Reading

Follow us on Twitter: @DataScienceCtrl | @AnalyticBridge

Stratelligence

via Featured Blog Posts http://ift.tt/1gdozRJ

July 12, 2016 at 04:49AM

Leave a Reply