NORDICS20 - Assets

Deep Learning on AWS

Amazon Web Services Resources EMEA

Issue link:

Contents of this Issue


Page 8 of 50

Amazon Web Services Deep Learning on AWS Page 4 6. Patterns for Deep Learning at Scale 7. AWS Guidance If you are not familiar with the deep learning process and the deep learning stack, read this guide in its entirety, in sequence. If you are familiar with AWS Deep Learning building blocks, deep learning challenges, and deep learning process, you can skip to sections 4, 5, 6, and 7. Deep Learning Process for Build, Train, and Deploy The following image shows the six steps of the deep learning process. In the following sections, we provide more information on each step of the deep learning process, explain challenges in terms of infrastructure performance, bottlenecks, scalability, reliability, and ease of use. Figure 2: Six steps of deep learning process Step 1. Collect Data Deep learning is different from traditional machine learning with regard to data collection and preparation steps. Although feature engineering tends to be the bottleneck in traditional machine learning implementations, in deep learning (specifically in image recognition and natural language processing [NLP] use cases), features can be generated automatically by the neural network as it learns. The features are extracted by having each node layer in a deep network learn features by repeatedly attempting to reconstruct the input from which it draws its samples, allowing it to minimize the delta between the network's guesses and the probability distribution of the input data itself. However, when training from scratch, large amounts of training data are still necessary to develop a well-performing model, and this necessitates substantial amounts of labeled data. There may not be enough labeled data available upfront especially when dealing with new applications or new use cases for a deep learning implementation.

Articles in this issue

view archives of NORDICS20 - Assets - Deep Learning on AWS