Deep Learning for Everybody Elliot English and the MetaMind team - PowerPoint PPT Presentation

Deep Learning for Everybody Elliot English and the MetaMind team elliot@metamind.io Socher, Ng, Manning Socher, Manning, Ng

Unstructured visual and textual data • Enormous growth of images and text • 1.8B images shared / day • 100B business emails sent / day • They span all industries and their analysis is valuable • Advertising, ad-optimization based on content • Medicine, Radiology images: early cancer detection • Insurance, Satellite images: building risk analysis • Finance, Sentiment analysis for trading • Customer Relationship Management, churn prediction • Their analysis requires machine learning Socher, Ng, Manning Socher, Manning, Ng

Machine Learning Used to Require a Ph.D. http://xkcd.com/1425/ Socher, Ng, Manning Socher, Manning, Ng

Why is that? Learning algorithm Describing your data with Learning features a computer can algorithm understand Domain specific, requires Ph.D. Can take largely level talent off the shelf Socher, Ng, Manning Socher, Manning, Ng

Feature Engineering is hard! – Real NLP Example • Task: Predict quality of a radiology report • Features: Parsing Named Entities Char n-grams POS Tags Coreference Taxonomy Socher, Ng, Manning Socher, Manning, Ng

Deep Learning can replace all of these: NLP! • Word vectors: • Recursive structures Socher, Ng, Manning Socher, Manning, Ng

Feature Engineering is hard! – Real Vision Example • Task: Predict class of object in image • Features: Canny Edges Harris Corners http://docs.opencv.org/trunk/doc/py_tutorials/py_feature2d/py_featur http://en.wikipedia.org/wiki/Edge_detection es_harris/py_features_harris.html SIFT SURF http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf http://www.vision.ee.ethz.ch/~surf/eccv06.pdf Socher, Ng, Manning Socher, Manning, Ng

Deep Learning can replace all of these: Vision! • Bottom level features from a convolutional neural network: http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf Socher, Ng, Manning Socher, Manning, Ng

Deep Learning has revolutionized the industry • Speech recognition systems of Google, Microsoft, Baidu all use DL • Google+, Microsoft and others use DL for very accurate image classification, e.g. results for: seat belt, boston rocker, archery, shredder • Let’s take a look at how we’re doing on the latter by examining a popular benchmark. Socher, Ng, Manning Socher, Manning, Ng

ImageNet Large Scale Visual Recognition Challenge 100 95 ILSVRC Classification Task Top-5 Accuracy 90 85 80 75 XRCE (Fisher Features) NEC (SIFT Features) 70 65 Jan-10 Feb-11 Apr-12 May-13 Jun-14 Jul-15 Socher, Ng, Manning Socher, Manning, Ng

State-of-the-art rapidly improving • Convolutional neural networks now the de facto standard for image classification • LeCun, Yann, et al. "Gradient- based learning applied to document recognition .” ( 1998). • Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks.” (2012). • Szegedy, Christian, et al. "Going deeper with convolutions .” ( 2014). Socher, Ng, Manning Socher, Manning, Ng

ImageNet Large Scale Visual Recognition Challenge 100 • We’re now at human accuracy! (not really) Google Microsoft 95 ILSVRC Classification Task Top-5 Accuracy Baidu Google 90 • Deep learning still Clarifai limited to select 85 companies SuperVision 80 75 XRCE (Fisher Features) NEC (SIFT Features) 70 65 Jan-10 Feb-11 Apr-12 May-13 Jun-14 Jul-15 Socher, Ng, Manning Socher, Manning, Ng

ImageNet Large Scale Visual Recognition Challenge 100 • We’re now at human accuracy! (not really) Google Microsoft 95 ILSVRC Classification Task Top-5 Accuracy Baidu Google 90 • Deep learning still Clarifai limited to select 85 companies SuperVision 80 • MetaMind makes 75 XRCE (Fisher state-of-the-art deep Features) learning readily usable NEC (SIFT Features) 70 65 Jan-10 Feb-11 Apr-12 May-13 Jun-14 Jul-15 Socher, Ng, Manning Socher, Manning, Ng

MetaMind: Deep learning for everybody • We take care of the details: – Machine learning algorithm selection 100 – Hyper parameter tuning – Efficient training procedures 80 Accuracy* (%) – Computational resource management 60 • you don’t need to worry about owning 40 your own GPU machines – Scalable inference infrastructure 20 0 0 2 4 6 8 Training time (Days) • We constantly improve your performance Socher, Ng, Manning Socher, Manning, Ng

Demos! Socher, Ng, Manning Socher, Manning, Ng

Language Demo: Twitter Sentiment Socher, Ng, Manning Socher, Manning, Ng

Language Demo: Semantic Similarity Socher, Ng, Manning Socher, Manning, Ng

Language Demo: Train Your Own Classifier Socher, Ng, Manning Socher, Manning, Ng

Vision: General Image Classifier Socher, Ng, Manning Socher, Manning, Ng

Vision Demo: Food Classifier Socher, Ng, Manning Socher, Manning, Ng

Vision Demo: Train your own classifier Socher, Ng, Manning Socher, Manning, Ng

Vision Classifier Use Cases Language Classifiers Vision Classifiers Socher, Ng, Manning Socher, Manning, Ng

API • Pre-trained classifiers: – https://www.metamind.io/api-quick-start • Train your own classifier tutorial: – https://www.metamind.io/api-tutorial-fit Socher, Ng, Manning Socher, Manning, Ng

Also doing research • Developing new models to improve accuracy • Improving both training and inference speed • Addressing new problems involving multimodal systems Socher, Ng, Manning Socher, Manning, Ng

MetaMind’s Vision Breakthrough AI for Everybody Socher, Ng, Manning Socher, Manning, Ng

Socher, Ng, Manning Socher, Manning, Ng

Socher, Ng, Manning Socher, Manning, Ng

Grounded sentence-image search Image-Sentence Demo Socher, Ng, Manning Socher, Manning, Ng

Socher, Ng, Manning Socher, Manning, Ng

Socher, Ng, Manning Socher, Manning, Ng