11. Reinforcement Learning
Learning to take actions to maximize reward
• Agents
• Games
• Policies
Google’s
Alpha GO
http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html
13. • Data come in all shapes and sizes
o Text
o images
o audio
o video
o graphs (aka networks)
o gene sequences
o gravitational waves
• In order for a machine to learn from these data,
we first need to represent them.
Features
(not bugs)
14. Typically, we need a vector representation
(aka a bunch o’ numbers)
Features
(not bugs)
15. Deep neural nets learn hierarchical levels of representation
Features
(not bugs)
http://www.datarobot.com/blog/a-primer-on-deep-learning/
16. Models
In order to find (approximate) the mapping between inputs
and outputs, we need a model
22. Summary
• We are trying to learn about the world, not about the
data
• ML is about finding mappings between inputs and
outputs that generalize to new inputs
• This is done by representing data as features, defining a
model and using optimization to find the best
parameters using data.
23. Data Science @
• Develop data products and predictive applications
• Apply cutting edge machine learning and computational
statistics.
• Collaborate with top medical professionals
• Revolutionize Health care delivery
Contact:
corey.chivers@uphs.upenn.edu @cjbayesian
Editor's Notes
You are blindfolded, you only have an altimeter, and _maybe_ (if you’re lucky) you can tell which way the ground is sloping below you.