Machine Learning APIs Comm mmon n appli pplications cations - PowerPoint PPT Presentation

Machine Learning APIs

Comm mmon n appli pplications cations Autonomous vehicles Optical character recognition Automatic image tagging Language translation Speech to text Video recommendations Portland State University CS 430P/530 Internet, Web & Cloud Systems

AI/ML: /ML: an en enormou rmous s sp space ce  Many different models and approaches  Expert systems / decision trees / knowledge engineering (medical diagnosis)  Supervised learning (Bayesian filters for spam detection)  Unsupervised learning (Clustering algorithms for Google News)  Combinatorial search (Chess)  Reinforcement learning (NPC in games)  Evolutionary/genetic algorithms (Smart fuzzing)  Often combined with each other  But recently…neural networks with supervised learning  Convolutional neural networks (CNNs) Portland State University CS 430P/530 Internet, Web & Cloud Systems

Neu eural ral netw etwor orks ks app pproach oach  Model the way the brain works via large collections of simulated neurons  Formation and pruning of selected connections encodes information Portland State University CS 430P/530 Internet, Web & Cloud Systems

Simu mulat lated ed in comput puter ers  Selected weights within a simulated network of neurons encodes information Portland State University CS 430P/530 Internet, Web & Cloud Systems

Everything old…  Neural Networks circa 1950s-1960s (perceptrons)  Multi-layer neural networks 1970s-1980s  Ran on original Macintosh!  Why the renaissance?  Video cards with massive numbers of processing units (thanks to gaming)  Massive storage capacities for data  Crowd-sourced platforms to provide labeling (to learn by example) Portland State University CS 430P/530 Internet, Web & Cloud Systems

Ima mageNe geNet t (2009) 09)  Pioneered "Deep Learning"  Use convolutional neural networks with massive sets of labeled data to solve the computer vision problem  Approach  Ignore academic skeptics and take risk on ancient algorithm  Collect image data from the Internet  Hire humans to label it via Mechanical Turk  Send through enormous neural network  Profit?  Fei Fei Li https://youtu.be/40riCqvRoMs?t=2m46s Portland State University CS 430P/530 Internet, Web & Cloud Systems

Now…“The of x” SpaceNet MusicNet Medical ImageNet DigitalGlobe, CosmiQ Works, NVIDIA J. Thickstun et al, 2017 Stanford Radiology, 2017 ShapeNet EventNet ActivityNet A.Chang et al, 2015 G. Ye et al, 2015 F. Heilbron et al, 2015  But, not all sunshine and rainbows… Portland State University CS 430P/530 Internet, Web & Cloud Systems

Corpus rpus can n have e iss ssues ues  US vs. Russian tanks in 1980s with early NN  US images crisp marketing shots on sunny days  Russian images grainy and on cloudy days  ML trains on wrong feature (crisp/sunny vs. grainy/cloudy) Portland State University CS 430P/530 Internet, Web & Cloud Systems

 SE Asian workers used to label ideal vacation photos  Label conference reception photos in hotels as the ideal vacation!  Perhaps the beach is hard work if they fish for a living?  Labeling is a relative task!  What happens when one uses ImageNet photos taken by humans on high-quality cameras for drone applications with poor imaging hardware and no human? Portland State University CS 430P/530 Internet, Web & Cloud Systems

Bi Bias as in labeli eling ng  Institutionalize bias behind facade of an "objective" algorithm  Racial, socio-economic bias in data being labeled and used  TED talk  GCP podcast #114 (2/2018)  FAT* conference  Example: ML for domestic violence  Most common way is via neighbor complaint  ML incorrectly learns that only those who live in row-homes and apartments commit domestic violence!  Hard problem  "It’s not always so obvious ahead of time what the bad outcomes might be" Portland State University CS 430P/530 Internet, Web & Cloud Systems

 MM's NY Times article (11/2018) and book… Portland State University CS 430P/530 Internet, Web & Cloud Systems

Rare e pr problematic blematic cases ses  Example: Self-driving cars  Kangaroos, white trucks against a white sky  Need millions of miles trained to get enough anomalous conditions  Must also train in winter/spring, in snow/rain/clear conditions Portland State University CS 430P/530 Internet, Web & Cloud Systems

Res esistance stance to adver ersaries saries  Both in training and in inference  Small changes fool classifier Portland State University CS 430P/530 Internet, Web & Cloud Systems

Requi uires res la large-scale scale particip icipat ation/da ion/data ta  In thousands of entries 0.28 0.66 172 157 123 81 0.03 0.23 35 29 2010 2011 2012 2013 2014 2015 2016 Average Precision Number of Entries Classification Errors (top-5) For Object Detection Portland State University CS 430P/530 Internet, Web & Cloud Systems

“Datasets— not algorithms — might be the key limiting factor to development of human- level artificial intelligence.” A L E X A N D E R W I S S N E R - G R O S S Edge.org, 2016 Portland State University CS 430P/530 Internet, Web & Cloud Systems

Towar ards ds crowd wd-sour sourced ced ML We’re passing the Why? democratizing image-net.org baton to Kaggle : a data is vital to remains live at community of more democratizing AI. Stanford. than 1M data scientists. Portland State University CS 430P/530 Internet, Web & Cloud Systems

An Ex Expl plosion osion of Da Datase tasets ts 1627 276 1919 1MM 4MM Hosted Datasets Commercial Student Data Scientists ML Models Competitions Competitions Submitted Portland State University CS 430P/530 Internet, Web & Cloud Systems

Qu Ques estio tion  Who on this planet has the largest, most interesting, data-sets?  Monetize data via ML models Portland State University CS 430P/530 Internet, Web & Cloud Systems

Softw tware's are's new w ap applic lication ation bu buildin lding bl blocks cks  Don't build your own when you can use pre-trained models done by experts on massive datasets  Part of an emerging API development platform  Accessed via REST APIs with results sent back in JSON  Abstraction raised, “Hello world” will never be the same Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Visi sion on API  Image recognition  Image labeling on thousands of labels  Face detection  Sentiment analysis (Emotobooth)  Text detection (optical character recognition)  SafeSearch content identification (adult/violent content)  Logo identification  Landmark identification Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Spe peech ech APIs Is  Speech-to-Text  Word recognition  Context-aware transcription  Automated punctuation  Offensive content detection  Text-to-Speech (speech synthesis)  Both in 120 languages Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Translation anslation API  Language detection and translation  100+ languages  Python, Java, Ruby, Objective-C bindings via Google API client libraries Cloud ud Natural tural Lang nguage uage API  Language analysis  Syntax analysis  Semantic analysis, entity recognition  Sentiment analysis  Common e-mail responses Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Video eo Intelligence elligence API  Summary and information extraction from video  Autotag objects in video to enable searching  Scene detection for thumbnail generation  Automated highlights (RedZone!)  Trained on massive set of labeled YouTube videos Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Video eo Intelligence elligence API  Demo code similar to image blurring lab  Flow chart triggered when new video placed in bucket Portland State University CS 430P/530 Internet, Web & Cloud Systems

ML as s a se service ice  Use ML to generate custom ML models so regular users can build their own ML models  AmazonML  http://cloudacademy.com/blog/aws-machine-learning/  AzureML  http://cloudacademy.com/blog/azure-machine-learning/  Watson Analytics  http://www.ibm.com/analytics/watson-analytics/  Google AutoML Portland State University CS 430P/530 Internet, Web & Cloud Systems

Cloud ud Aut utoML ML for Visi sion on (2018) 8)  ML services trained on general datasets, but custom images and domain labels often needed  AutoML  Apply transfer learning to re-use trained models to quickly learn new domains  Custom models via  Labeled data (done by user)  Unlabeled data (done by humans at Google)  Evaluate many different ML models and pick best one (> 13)  Example: Custom models to recognize machine parts  Democratizes ML for the masses  All that is required is for you to upload your data with labels or exemplars for Google to label Portland State University CS 430P/530 Internet, Web & Cloud Systems