Today, the new release of an academic report, Facebook introduces new AI open source software Torchnet, to simplify the depth of learning.
The depth of the current study is a fashionable way of learning involves training artificial neural network processing large amounts of data, such as pictures, and then let the neural network to predict new data. With the establishment of a new deep learning framework, Facebook chose the Torch compile an open source library, Facebook has previously committed to this.
& Ldquo; This makes it easy to implement completely hidden I / O (input / output) cost function, and if you want to create a practical large-scale deep learning system, you may well need this feature. & Rdquo; Facebook Institute of Artificial Intelligence (FAIR) is a scientist Laurens van der Maaten says in an interview, & ldquo; we do not want the operation Torch faster or slower, or other event, these are not the focus frame lies. & Rdquo; van der Maaten said, Torchnet is written in Lua scripting language that can run on standard x86 chip or graphics processing units (GPUs), but also allows programmers to reuse some code. In this way it is possible to reduce the workload, but can also reduce the chances of bug appears.
Facebook is not the only company to Torch compilation tools, Twitter also done so, be sure that the same van der Maaten talked to some Twitter employees also seem to Torchnet interested. Amazon, Google and Microsoft in recent months, also announced a new deep learning framework, has been committed to the development of open source projects before native Facebook, such as React Native and Presto, so now Facebook to make some changes seem quite interesting.
van der Maaten said: & ldquo; this way with Theano frame Blocks and Fuel library is somewhat similar. & Rdquo;
In this report, van der Maaten wrote: Torchnet may not have been confined to the Torch use, Torchnet is abstract, can be easily applied to other frameworks, such as Caffe well as Google's TensorFlow framework.
In six or seven months ago, Facebook released the first version of Torchnet. van der Maaten said: & ldquo; now there are many teams use this version in different applications. & Rdquo; he did not say specifically what kind of work dependent on Facebook in Torchnet, but it can be applied to aspects of the image recognition and natural language processing, so that they can come in handy on a number of things, such as find relevant Instagram pictures to choose the best Facebook post as News Feed. Facebook hopes that its content will be better than other content on the network more attractive, both to retain & ldquo; repeat & rdquo; but also to attract new users, so this is a very important thing.