Deep learning is an aspect of machine learning that attempts, through algorithms, to model complex abstractions in data. A research goal is to create architectures to achieve this with masses of uncharacterized data. This research paper, for advanced scholars, moves this process forward, in the context of multimedia, into a readily usable and scalable model for nonexperts. As such, it is a must-read for researchers and scholars of deep learning.

The researchers have designed “a distributed deep learning platform called SINGA, which has an intuitive programming model and good scalability.” With SINGA, the user creates a neural network model that may be run with hybrid parallelism. The paper is well organized into readable sections introducing the topic, the programming model, its implementation and architecture, and the experimental study and related work. A number of figures illustrate the described work.

The researchers use a memory-level parallelism (MLP) model for image classification. SINGA provides users with a variety of layers from which to build models. Additionally, users may define their own layers.

The researchers discuss the training for three different applications. They are “a multi-modal deep neural network (MDNN) for multi-model retrieval, an RBM for dimensionality reduction, and a recurrent neural network (RNN) for language modeling.” They provide the logical architecture for SINGA and how SINGA partitions the neural net for various aspects of parallelism. Additionally, details of the experimental study are provided. The research paper concludes with a list of references.

With an introductory abstract, categories, and keywords, as well as excellent organization, the researchers present a significant amount of detail in ten pages. This is recommended reading.