Facebook Has increased Its Training For Visual Recognition Models.

Over the past few years, Facebook has gone through many changes. It is focusing on images and auto identification of people. That all could happen with the deep learning models. They are extremely important and will play a very special role in the near future. Due to the growing importance of deep learning, companies want to make sure that they are not lagging behind in the deep learning game. Facebook, is the best social networking site has increased their level by announcing a major breakthrough in its deep learning capabilities. With this Facebook has really scaled things up. Its AI and machine learning teams are finding ways to train data sets. They teach visual recognition models which distinguish between a large number of images.

 
In today’s world, deep learning training has become the need of an hour so to reduce this training time drastically without any loss in accuracy, Facebook has come up with a solution and published it on a paper. They were able to reduce the training time from 29 hours to one of RestNet-50 deep learning model on ImageNet. Facebook managed to reduce the timing by distributing the training in several mini batches across a greater number of Graphical processing units. Facebook has increased the batch size to 8192 images across 256 Graphical processing units which were earlier 256 images to 8 Graphical processing units. Because of the large number of GPUS, Facebook managed to reduce the training time. Also, with the increase in GPUs, Facebook was able to achieve 90% of the scaling efficiency.
 

Also Read: Facebook Introduces M To Pop Up Suggestions

Facebook is now able to train 40,000 images per second per machine which is possible to train on 1k data set. Earlier it used to take days and even months to do so. This is essentially how training takes place. Greater accuracy is achieved by larger data sets which require more training time and even additional resources. It would then be possible to achieve more accurate results.
Now-a -days every company is handling a large amount of data. Then, they depend themselves on AI and machine learning. Larger the data, larger will be the training time. Facebook’s AI Research which is abbreviated as (FAIR) and Applied Machine Learning which is abbreviated as (AML) teams is working on to reduce the minibatch size. Machine learning team have to compromise a lot. 

 
Facebook also announced that it is open-sourcing its hardware stack which has helped them to reduce the training time. It would also help many companies that work with large amount of data