A Berkley view of AI systems
[2017-10-26 Thu 10:04:58] speaker: Randy Katz, UC Berkley
Based on a tech report, A Berkley view of sysytem challenges for AI
Real time intelligent secure execution: riselab at UC Berkley
- follow up to AMPLab (2011-2016)
- Algorithms, Machines, People
- making sense of big data
- batch data -> advanced analyticsc
- RISELab: live data -> real time decisions
AI based decisions
- recommendation systems
- building control
- medical diagnosis
- financial decision making
- manufacturing line
- autonomous vehicles
Challenges
- mission critial applications
- handle changing/unpredictable environments
- fraud
- financial market
- learn across multiple organisations
- detect virus outbreak
- fraud detection
Good decisions are
- fast
- uses fresh data
- based on personalized data
- explainable
Task: Shared learning
- without leaking confidential information
- banks cooperate to improve fraud detection
- machine learning as a service on confidential data
- work done on enabling linear models to be shared
Task: Reinforcement learning
- generalization of supervised learning
- allows incremental updates
- policy represented as DNN
- Q: training in simulator
- vs creating an actual robot that runs
RISE stack
- open source platform to develop of RISE like apps
- computational frameworks
- Ray: distributed system for AI
- minimalist execution engine, allows playback
- manage access to data
Smart cities: building, energy, and transportation
- rethinkx report on transportation
- transportation as a service
- EV as part of the grid
- XBOS-DR
- extensible building operation system + electric demand response
- predictive occupancy model using Clipper
- using RL to control building and EV
- EV load is equivalent to a single house
- need to schedule the charging
- Uber surge pricing to manipulate driver/rider behavior
- forecast demand and using RL + rollouts to control surge pricing
- AVs need context and a way to interact with humans
- crossing the street in New York vs India