14:50 | 15:30
Keywords defining the session:
- Machine learning
Takeaway points of the session:
- Machine learning has been made accessible to anyone through the use of automation.
- Multiparadigm data science uses symbolic computation to integrate many daomains of computation and data
The talk will discuss Wolfram’s progress towards an entirely automated workflow for machine learning that makes it possible for anyone who can code to make use of AI methods.
The talk will assume no knowledge of machine learning and will introduce key concepts before showing how to use the Wolfram Language to create advanced predictors from scratch, in seconds. Use will be made of the Wolfram Language functions Predict and Classify which automate the process of data encoding, feature extraction, method selection, training, end decoding. The net result is a single function that produces a ready-to-use classifier or predictor, directly from raw data.
Examples will be shown from computer vision such as content recognition, scene recogniton and segmenting images into components relevant for autonomous driving applications. The speaker will attempt to train a predictor using live camera images to be able to recognize the hand gestures of the game Rock, Paper, Scissors.
Other live image examples will include predicting the geo location that a photograph was taken, learning an artistic style and applying it to a photograph, creating imagined content such as aerial photographs guided only by a map creation, and the addition of color or depth information to a photograph.
Because the pipeline is fully automated, it is possible to use many types of data and other live examples that will also be demonstrated will include recognizing the content of a recorded sound, extracting facts from freeform text, decomposing the semantic content and structure of text, and sentiment analysis.
The talk will explain how symbolic computation is the they key ingredient needed to make such automation possible. Symbolic computation provides a unified representation of structured data such as images, sounds, documents, databases, networks and more. When all data is fundamentally the same, it makes it easy to pass such apparently different data through the same computational pipeline. To the user, creating a predictor based on sound can be exactly the same as one based on text, or images, numerical or categorical data or any combination of those.
Furthermore, the same symbolic representation can be applied to models and code making it possible to create or transform a neural network in the same way that one might process an image, restyle a document or filter a dataset. This abstract representation of neural networks makes an extremely powerful tool for expert users who as well as being able to describe complex neural networks directly with minimal code, can also acquire, transform, retrain and analyze existing networks or programmatically generate networks
The talk will go on to show that symbolic representation also helps in automating the transition from research experiments to the production deployment of AI services. By representing user interfaces and APIs symbolically, it is possible to automate the machinery of deployment to web sites, mobile devices or IoT devices. A key demonstration of this workflow automation will involve the speaker importing data, creating a machine learning classifier, and deploying a web based user interface and RESTful API, all coded live on stage.