Presenters
- Greg Corrado, Senior Research Scientist, Google
- Vincent Nestler, Professor & Assistant Director of Cybersecurity, CSU San Bernardino
- David Vasilia, Enterprise Network Administrator & Faculty, CSU, San Bernardino
- Internet2 & GCP: internet2.edu/gcp
- CS edu grants: cloud.google.com/edu
Machine Learning 101
- Already in everyday products: photos, inbox, maps
- 2 disciplines: AI and machine learning
- Traditional AI systems are programmed to be clever
- ML-based AI systems are designed to learn to be clever
- Classic AI works on rules and contingencies; ML AI learns from examples and data.
- Machines learn by example: models (which have parameters) feed predictions, which feeds a learner, which in turn feeds the parameters. This is surprisingly simple and generic.
- Need 4 things: computational resources, good tools & algorithms, training examples, creativity and ingenuity of people.
- Effective, but very gradual process that takes millions or billions of examples for it to work. It needs to cycle many many times.
- ML coming of age in this decade because the computational power is exists now and it’s cheap and plentiful enough, i.e. CPU, GPU, Google TPU.
- tensorflow.org a toolkit for machine learning
- Open standard
- Next gen deep learning tools built in
- One system flexible enough for ML research
- Robust enough for use in real products
- Same software Google researchers use
- Deep learning not one function, but a set of composable subfunctions for model building.
- Distributing ML Tech Globally
- Shared Tools: TensorFlow + CloudML
- Ready-made ML systems (Cloud Vision API, Cloud Speech API, Cloud Translate API, etc.)
- Use our tools to build your own system!
- Example: TensorFlow cucumber sorting tool (really!)
- Shared knowledge: open research publication at intl conferences; global direct community education; funding academic research and education.
- Google published 90+ papers in the last 4 years
- Takeaways:
- Differentiation between AI vs. ML vs. Robotics
- It isn’t magic, just a tool
- Machines learn best from examples
- Why now? fast computation
- Make ML work requires creativity/ingenuity, cheap/fast computation/examples to learn from (data), tools & algorithms, TensorFlow makes ML software available for free.
- Google Cloud makes hardware available.
Cloud for Higher Ed
- Programming a campus rover: students are given a sensor, a raspberry pi, and Python. Then, they need to figure out how to integrate it.
- Hacking now means hacking things together. You don’t have to be an engineer and you don’t need to know everything.
- How can I level the playing field for my students? Be able to connect to Chrome and a Google compute engine. Everyone can look at and work with this environment, and they can explore from there.
- A project we worked on in class: Android mapping for WiFi signal strength on campus. War driving took signal strength and using mapping API to literally map it to a real topographical map. Now we can “see our WiFi.”
- We used Intermapper software to map the Internet, specifically the CENIC network from Los Angeles. The students loved this.
Panel
- What is the difference between deep learning and machine learning? ML is the larger field of making machines that learn. DL is a small subset of this.
- How far is Google taking cultural sensitivity into account with ML? Take translate as an example: you can dig into what the algorithm did to come up with its response.
- If we use a Google tool, does this tool report what it learns back to Google? NO. What is pricing model for Google Cloud for Google Apps customers? It is independent of G-Suite.
Next Steps
- Google is now a member of Internet2.
- Will work with universities across the US to explore how Google Cloud Platform can better serve higher education
- Help students build what’s next!
- GCP Education Grants are available to: faculty in US, teaching university courses in CS or related fields in 2016-17 academic year. Examples: general CS, Cybersecurity, systems administration, networking.
