TensorFlow - Ep. 22 (Deep Learning SIMPLIFIED)

Google's TensorFlow is currently the most popular Deep Learning library on GitHub. This video will provide an overview of the library's strengths, weaknesses, and numerous features.

Deep Learning TV on
Facebook: https://www.facebook.com/DeepLearningTV/
Twitter: https://twitter.com/deeplearningtv

TensorFlow grew out of a proprietary deep net library called DistBelief, which was built by Google as part of the Google Brain project. The goal of this project is to build a system for large scale machine learning models so that they can be deployed on a variety of platforms – from smart phones to clusters consisting of 100s of nodes and 1000s of GPUs. Another goal is to increase the portability of machine learning and to simplify the process of transferring research models to commercial-grade applications.

Like Theano, TensorFlow is based on a computational graph – a directed graph where nodes represent mathematical operations and edges represent the flow of data between nodes. The data set carried by an edge from one node to another is called a tensor, a type of multi-dimensional array. The library's name stems from the way these tensors flow across the network. Due to its structure, TensorFlow is not limited to neural net applications. Any domain where computation can be modelled as a data flow graph can benefit from the TensorFlow library. TensorFlow also shares several important features with Theano such as auto differentiation, shared and symbolic variables, and common sub-expression elimination.

Different types of deep nets can be built using TensorFlow, although TensorFlow does not allow for hyper-parameter configuration of deep nets. The TensorFlow Roadmap suggests a possible layer configuration feature, but there is currently no mention of when this would be implemented. For now, the library Keras provides TensorFlow with a layer configuration option. TensorFlow also provides a “no-nonsense” interface for C++. In addition, TensorFlow has comprehensive and informative documentation, including the March 2016 release of a free Massive Open Online Course (MOOC) on Udacity.

TensorFlow and Theano are very similar, but TensorFlow was several orders of magnitude slower than Theano from its release until March 2016. The TensorFlow community has since worked to combat performance and other issues. An April 2016 update from Soumith Chintala of Facebook shows that TensorFlow performed well in the ImageNet category, with no Theano-based libraries in the rankings.

An important feature of TensorFlow v0.8 is the implementation of data parallelism, which is similar to the Iterative Map-Reduce from Deeplearning4j. This version of TensorFlow also implements model parallelism, where different portions of the graph can be trained on multiple devices in parallel. The TensorBoard feature allows users to visualize both performance and the different levels of the network architecture. Due to community requests, TensorFlow's Roadmap also includes support for OpenCL, a fast-rising standard for GPU computing.

URLs
Soumith Chintala benchmarks - https://github.com/soumith/convnet-benchmarks
TensorFlow road map - https://www.tensorflow.org/ver....sions/r0.8/resources
Github ML Showcase - https://github.com/showcases/machine-learning
Somatic Ruby post - http://www.somatic.io/blog/ten....sorflow-is-coming-to
TensorFlow MOOC - https://www.udacity.com/course..../deep-learning--ud73

Credits
Nickey Pickorita (YouTube art) -
https://www.upwork.com/freelan....cers/~0147b8991909b2
Isabel Descutner (Voice) -
https://www.youtube.com/user/IsabelDescutner
Dan Partynski (Copy Editing) -
https://www.linkedin.com/in/danielpartynski
Marek Scibior (Prezi creator, Illustrator) -
http://brawuroweprezentacje.pl/
Jagannath Rajagopal (Creator, Producer and Director) -
https://ca.linkedin.com/in/jagannathrajagopal