Saturday Keynote at #ODSCWest Carrie Grimes Bostock - Data Science & Automated Decision Making. Automation has inputs, model(s) and potential outputs for action.
What is the goal of your models? Interpretability is important, but not the only thing. Other measures exist, but understanding your model is important to prevent outcomes, override recommendations, and to connect models to other systems - or stacks of models & systems. #ODSCWest
Example: placing racks in the Santa Clara data centre. Considerations, restrictions on space, power, future use of space, etc. Also: People install racks. They need to be able to intervene based on info that the placement model doesn't have. #ODSCWest 
So, what do we want out of the system? Future planning, large scale, ability to move, change or adapt. And test. This is HARD to do in a non-interpretive model. #ODSCWest
Example: human-guided experiments in fusion at @GoogleAI ai.googleblog.com/2017/07/so-the…
Optometrist model: human-labelled interference. Repeat recent best shots, pick a new point, expert marks interest or risk at point. Then test and take steps in parameter space to keep moving forward - but cautiously. Now develop a model to help humans in the space. #ODSCWest
Importance of interpretability in these systems
- need to be able to test and integrate
- need to design solutions in context
- models will (and need to) change & evolve
- consider changes in the system.
#ODSCWest
- need to be able to test and integrate
- need to design solutions in context
- models will (and need to) change & evolve
- consider changes in the system.
#ODSCWest
In a nutshell: #Interpretability is super important if you're doing #DataScience at scale. Would also apply to policy, legislation, regulations. (And then all these systems actually interact in the real world.) #ODSCWest
Next up: Derek Haoyang Li - Developed AI tutor - @YixueSquirrelAi - on the future of AI:
- Know-all & remember all - ability to absorb and recall perfectly
- Debater: predict positions of opponent debaters and adjust position
(con't)
- Know-all & remember all - ability to absorb and recall perfectly
- Debater: predict positions of opponent debaters and adjust position
(con't)
(.. cont)
- Understanding: focus on where students need knowledge
Trying to understand everything in every student's mind and design knowledge paths based on student knowledge & activity.
- Understanding: focus on where students need knowledge
Trying to understand everything in every student's mind and design knowledge paths based on student knowledge & activity.
Analogy: a GPS for the mind. Map all the streets, know the buildings, know the traffic patterns.
Example: TouTiao [Tik Tok] - #AI News org. Applies tags to users instead of catagorizing them. Downloads surpassed FB, Twitter, Instagram in Sept 2018.
1M students already. Map knowledge points. How long does it take a student to absorb a knowledge point? Always different - so build this into the plan. Each student needs their own classroom. Models for #Student #Content #Teaching #Prediction #ODSCWest
#AI ability to evaluate massive potential futures in instants.
- Break knowledge points down into "pixels".
- Evaluate video capture, behaviour, time
AI Tutor learned from expert teachers for 3 years. It is now better.
& Scariest part of AI == self-evolving.
- Break knowledge points down into "pixels".
- Evaluate video capture, behaviour, time
AI Tutor learned from expert teachers for 3 years. It is now better.
& Scariest part of AI == self-evolving.
Previous example: SRI Motobot #ODSCWest
Squirrel AI: also breaking down and decomposing the skills of learning - info filtering, pattern identifying, backward reason, creating general models, etc. Helps students to explore new knowledge points. #ODSCWest
Thought: Apply to VR environment? Reminds me of the Vulcan school in the Star Trek reboot. #ODSCWest
Next up: Daphne Koller - CEO and founder of Insitro @DaphneKoller
- Standford Academics
- Saw huge need for high quality online courses
- Founded Coursera in 2012
- Left Aug 2016 (Coursera in a good place)
- Now focusing on AI in human health with Insitro
#ODSCWest
- Standford Academics
- Saw huge need for high quality online courses
- Founded Coursera in 2012
- Left Aug 2016 (Coursera in a good place)
- Now focusing on AI in human health with Insitro
#ODSCWest
Vision for Coursera: anyone anywhere can transform their lives through high-quality learning experiences. #ODSCWest
Advice for entrepreneurs: difference between success and failure can be sticking with it and showing grit. Also: don't become an entrepreneur because you want to be one - find the problem you are burning to solve. Good blog on subject here: #ODSCWest medium.com/@benwerd/found…
Today in #AI we have bilingual people
- get machine learning
- also have domain knowledge
Need both to see opportunities and see where problems lie. Insitro is bringing #MachineLearning as a core skill in field of drug development
#ODSCWest
- get machine learning
- also have domain knowledge
Need both to see opportunities and see where problems lie. Insitro is bringing #MachineLearning as a core skill in field of drug development
#ODSCWest
Problem: cost of successful drugs need to cover cost of unsuccessful drugs. Looking to apply #AI to this and other problems. #data in this field is incredible. Now also have #CRISPR #ODSCWestt insitro.com
Example: do a google image search for #CEO - how many women do you see? #ODSCWest google.com/search?q=CEO+i…
Also points to the need for interpretable models (which has been a big theme here this week). distill.pub/2018/building-… #ODSCWest
Great morning keynotes with a surprising focus on education & learning. #ODSC @threadreaderapp unroll
