[ad_1]
April 24, 2019 –
The implementation of Morphnet by Tensorflow has been published by Google, open source. This automatically develops and optimizes machine learning algorithms.
Google announced in a blog post the provision of Morphhnet's Tensorflow implementation source code. This approach developed by Google helps to automate and improve artificial intelligence systems by simplifying their architecture. In concrete terms, this makes it possible to further develop the already existing deep neural network (DNN) structure.
The result is smaller, faster, and more powerful models that can be applied to Google-wide problems, "said Andrew Poon, Senior Software Engineer, and Dhyanesh Narayananan, Product Manager at Google's AI Perception. Morphnet introduces machine learning algorithms through a cycle of contraction and expansion phases. In the first case, inefficient neurons, basic elements of many modern machine learning models, are filtered and adapted to each task using a mechanism to calculate the cost of each neuron (and therefore its effectiveness) .
Poon and Narayanan report that Morphnet, applied to the popular InceptionV2 computer vision algorithm, which was trained on the Imagenet open source dataset, could reduce FLOP costs by 11-15% without compromising accuracy. The repetition of the contraction / expansion cycle resulted in a gain in accuracy of 1.1%.
More details on Morphnet can be found in this blog post. The Tensorflow implementation can be downloaded from the Github project website.
(Swe)
[ad_2]
Source link
