The Google Brain project started in 2011 to explore the use of very-large-scale deep neural networks, both for research and for use in Google’s products. As part of the early work in this project, DistBelief – first-generation scalable distributed training and inference system – was developed. Based on the experience with DistBelief and a more complete understanding of the desirable system properties and requirements for training and using neural networks, TensorFlow, system for the implementation and deployment of largescale machine learning models, was born.
Machine Learning is about finding patterns in data. A person on the same principle performs many tasks as in childhood. Our brain receives information, processes, accumulates, eliminates unnecessary, “repeats” and learns something new.
In fact, TensorFlow is a tool that allows you to implement or simplify the implementation of machine learning for your system or task. For example, to add two numbers, no one needs to write the numbers to the register, call the add operations and read the result from another register. In other words, TensorFlow is Deep Learning, an open source engine and an area of Machine Learning that uses the concept of the functioning of the human brain’s neural connections to solve various problems. such as pattern and speech recognition, i.e. such tasks that are difficult to solve using just computing power.
In contrast to the usual familiar functions and services, the neural network can not only provide the output at the output according to predetermined rules but learns by itself (the principle is based on how our brain behaves), i.e. the network determines itself the rules (weights and offsets) and gives the result.
With the release of TensorFlow version 1.0, the platform has gained such features as:
Speed. TensorFlow 1.0 is incredibly fast! The XLA lays the groundwork for even greater productivity gains in the future, and tensorflow.org already has tips and tricks to achieve maximum speed for your models.
Versatility. TensorFlow works almost everywhere, including Android, iOS, Raspberry PI, laptops and desktops. In addition to the discovery of TensorFlow itself, Google also opened tools for maintaining models in production.
Flexibility. High-level APIs appeared in TensorFlow 1.0, with tf.layers, tf.metrics and tf.losses-modules.
What is important for developer that would like to become acquainted with the TensorFlow?
Knowledge of machine learning will help. To learn more about TensorFlow, you can see the tutorials at www.tensorflow.org. Basic knowledge of neural networks, probability theory and mathematical analysis are also needed. But don`t be so afraid about it – the universality and speed of training and hypothesis testing (or piloting the model) will allow building a TF-based system without deep fundamental pieces of knowledge (but note that the understanding of the general principles is essential).
In conclusion, there are some interesting ways to implement TensorFlow in your industry:
– A tool that allows doctors to diagnose eye diseases.
– An application that creates pictures and music.
– Australian marine biologists use TensorFlow to search for sea cows in tens of thousands of photos.
– The Japanese farmer has trained TensorFlow to sort vegetables.
– Radiologists use TensorFlow to detect signs of Parkinson’s disease on medical scans.
– Stanford doctoral students with TensorFlow detect skin cancer.
– Gulf researchers use TensorFlow and Raspberry Pi to track Caltrain trains.
For now, there are already enough neural networks that can be used to solve their problems, without modifications or with minor modifications. Perhaps this is one of the most popular ways to use Machine Learning among a wide range of developers.