Machine Learning Discussion Dave Draffin 04/24/ 2 018 After this - - PowerPoint PPT Presentation

Machine Learning Discussion

Dave Draffin 04/24/2018

Goals

After this discussion you should:

• Know why Machine Learning is needed
• Understand the definition of Machine

Learning

• Understand the types of problems that

benefit from Machine Learning

• Be aware of some typical uses
• Understand the concept of teaching

Machines to Learn

• Be capable of further study on your own

This discussion will NOT teach you how to implement Machine Learning

www.hcltech.com

The Key to Understanding Machine Learning

www.hcltech.com

Hand Drop

What Happens when I drop the tennis ball by hand – about 4 feet?

• ? Rise ?
• ? Go sideways?
• ? Fall ?

How do you know?

• Experience
• Observation
• Equations

www.hcltech.com

Hand Drop - Observations

Observations

• Behaves the same each time (pretty much)
• Observed results match equation predictions (Prettyclose)

Conclusions

• Equations MODEL behavior very well
• Equations can be used to PREDICT behavior
• ➔ CLOSED FORM Solution

www.hcltech.com

Balloon Drop Thought Experiment

What Happens when I drop the tennis ball from a balloon 1,000ft up?

• No wind

What do you observe?

• Where does the ball land?
• Do the hand drop equations still work?

www.hcltech.com

Balloon Drop Thought Experiment - Results

What Happens when I drop the tennis ball from a balloon 1,000ft up?

• Ball lands directly below the balloon

Observations

• The ball lands later than prior equations predict.
• The ball is moving slower than prior equations predict

What is happening?

www.hcltech.com

Balloon Drop Thought Experiment - Conclusions

What happened?

• Air resistance became SIGNIFICANT
• It was present in hand drop, but small effect.
• With longer drop it became a SIGNIFICANT factor.

Can we write equations for this case?

• Yes!
• A little more complicated
• Can still create a CLOSED FORM solution
• Actual results will closely match PREDICTIONs

www.hcltech.com

Tennis Cannon Shot Thought Experiment

Point a perfect cannon at Anaheim Stadium near Los Angeles.

• No wind
• About 1200 miles

What do you observe?

• Where does the ball land?
• Can we write equations to predict the landing spot?
• Why would this be so difficult?

Does allowing the existence of wind change your answers?

• If you shot it over and over again would the ball land in the same spot?

www.hcltech.com

Tennis Cannon Shot Thought Experiment - Conclusions

No wind

• Ball would land about 1,000 miles off the coast of California (Coriolis effect)
• Air density and therefore air resistance would change with temperature.
• Is gravity constant?

Wind

• Winds may not be known at launch
• Winds may vary after launch

Conclusion

• COMPLEX PROBLEM (a lot of elements interacting with each other)
• Unknown initial conditions
• Unknown change in conditions
• Can not create a equation that consistently calculates landing location.
• ➔ as the problem gets bigger, longer, further – more factors SIGNIFICANT

www.hcltech.com

Tennis Cannon Shot Thought Experiment – Coriolis Effect

http://abyss.uoregon.edu/~js/glossary/coriolis_effect.html

www.hcltech.com

Machine Learning - Uses

Problems without a CLOSED FORM solution

• Or where the CLOSED FORM solution is unaffordable

Problems with lots of elements that interact Problems where you don’t know the way and degree that elements interact Problems where it is not clear which inputs are SIGNIFICANT

www.hcltech.com

Machine Learning –What is it?

Machine Learning is a STATISTICAL process

• Results are probabilities – not assurances

Strives to improve performance to a GOAL

• The more learning, the better performance to the goal (in general)

Uses FEEDBACK(Learning) to improve GOAL seeking performance Machine Learning builds a MODEL to predict outcomesbased on inputs.

• This MODEL may be modified by further learning.

Two broad categories

• Supervised learning
• Unsupervised learning

www.hcltech.com

Machine Learning – Supervised Learning

Card identification

• Color of Letter/symbol
• Letter
• Symbol
• Location

Learning

• Preset possibilities (red/black)
• Present examples (good and bad)
• Feedback correct answer
• Used to improve model

www.hcltech.com

Machine Learning – Unsupervised Learning

Goal Seeking

• Maximize, Minimize, or both
• Some parameter or small set
• Either end state or continuous
• Classifications not predefined

Learning

• Totally by experience
• May involve training data set
• Machine/person interactions
• Machine/machine interactions

GO! Example

• Trained by playing People and
• ther programs, or itself
• Plays vastly different than human
• Google AlphaGO! Beat some of

the best players in the world.

www.hcltech.com

Machine Learning – Key “take aways”

Machine Learning is a STATISTICAL process Strives to improve performance to a GOAL (or goals) Uses FEEDBACK(Learning) to improve GOAL seeking performance Machine Learning builds a MODEL to predict outcomesbased on inputs.

• This MODEL may be modified by further learning.

Requires TRAINING or datasets to improve the model. Useful for large and/or complex problems that don’t have an affordable CLOSED FORM solution. Can identify previously unrecognized relationships (see Analytics)

www.hcltech.com

Machine Learning – Questions

Questions?