Deep Learning at Flickr, Pierre Garrigues

Home / Data Science & Machine Learning / Deep Learning at Flickr, Pierre Garrigues

Pierre Garrigues is a Researcher in Machine Perception and Learning at Flickr and also spoke at the Deep Learning Summit at the end of January to give an insight into how Flickr are automating the labelling of their image libraries using Deep Learning techniques as well as the 10 million uploads which they receive each day.

Flickr are part of Yahoo which means that they have access to some immense Hadoop grids so are able to transfer photos onto HDFS (Hadoop Distributed File System) and bring the computation to the data in a matter of days.

Garrigues describes how Flickr have seen an immense improvement in the accuracy of photo labelling for users in search results as well as applying the same labelling to private photo search. Flickr are acutely aware that mitigation of mistakes is extremely important as users would normally be able to label photos by themselves with 100% accuracy and deep learning for sites like Flickr is a new phenomenon carrying inherent risk for the business as a whole if it doesn’t work accurately enough.

Garrigues’ stewardship of Deep Learning at Flickr means that they have gone from zero to full scale image recognition in 18 months using a pipeline from feature recognition on a GPU through training and calibrating classifiers through to product integration.

 

Related Posts