Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage News Deep Learning at Gilt

Deep Learning at Gilt

This item in japanese

Deep Learning is a rapidly evolving subfield of Machine Learning originating from Neural Networks. Recent algorithmic advances and utilization of GPU parallelization have resulted in Deep Learning based algorithms mastering the game of Go as well as several practical applications.

The fashion industry is one of the target sectors for Deep Learning. Gilt has been using deep learning for product recommendations and clothing properties categorization. Dress faceting, automatically identifying occasion, silhouette, neckline and sleeve type is based on Facebook’s Torch library. Torch uses models trained by ImageNet to utilize the tags that each picture already has and enhances them by Gilt specific features. The system uses Amazon’s cloud infrastructure based on EBS and P2 instances that provides up to 16 GPUs per server. To test the classification quality, the system uses F1 Score that can weigh in both false negatives and false positives. Gilt tested SaaS alternatives which were not satisfying either in terms of accuracy or tags provided.

Product recommendations for dress similarity on the other hand is based on TiefVision, a deep learning image-similarity search engine. TiefVision is also based on classification of ImageNet data that swaps the last layers of the Neural Network with a new specialized network in a technique known as transfer learning. The first step is locating the dress in an image using Yann LeCunn’s overfeat method. After locating the dress, the algorithm uses a siamese network and Hinge loss function for training.

With Intel open sourcing BigDL, the distributed Deep Learning Library for Apache Spark, Amazon promoting MXNet as Deep Learning framework of choice for AWS and Deep Learning being used for anomaly detection among other use cases, it seems like not only software is eating the world, but also deep learning is eating the machine learning world.

Rate this Article