Enterprise Using automation to extract meaning from data Michael - PowerPoint PPT Presentation

Future directions of AI in the Enterprise Using automation to extract meaning from data Michael Schmidt, Ph.D.

About me • Cornell University, Ph.D. – CCSL Lab • Founded Nutonian in 2011 • Eureqa = AI Software, >50,000 users • Cited in > 500 medical, scientific and research advances

“Computer Program Discovers Laws of Physics” – New York Times Nature News Schmidt M. , Lipson H. (2009) "Distilling Free-Form Natural Laws from Experimental Data," Science, Vol. 324, no. 5923, pp. 81 - 85.

y = x 2 25 25 20 20 15 15 y 10 10 5 5 0 0 -5 -5 -4 -4 -3 -3 -2 -2 -1 -1 0 0 1 1 2 2 3 3 4 4 5 5 x

y = 0.02 x cos(4 x ) + 1/(1 + exp(-4 x )) 1 1 0.8 0.8 0.6 0.6 y 0.4 0.4 0.2 0.2 0 0 -6 -6 -4 -4 -2 -2 0 0 2 2 4 4 6 6 x

7

The world obeys mathematical relationships – from physics to business operations Modern AI can deduce these hidden patterns automatically from data Machine Intelligence

Test and Find Structure High error k 1 θ 2 + k 2 2 – k 3 ω 2 2 + k 4 ω 1 ω 2 cos( θ 1 – k 5 θ 2 ) k 1 θ 2 – k 2 ω 1 + k 6 cos( θ 2 ) + k 7 cos( θ 1 ) – k 8 cos( k 9 θ 2 ) – k 10 k 1 ω 1 ω 2 – k 2 cos( θ 1 – θ 2 ) cos( k 11 – k 12 θ 2 ) Accurate Complex Simple

Model search Parsimony / Simplicity Error Metric Error Model front Population Better of Models Models Complexity Variation

The science under the hood High Error Optimal Solutions Accurate Simple Complex

Robot Scientist Algorithms distill laws of physics from chaotic systems (published in Science 2009) Schmidt M. , Lipson H. (2009) "Distilling Free-Form Natural Laws from Experimental Data," Science, Vol. 324, no. 5923, pp. 81 - 85.

Getting the right result Evolutionary Search Neural networks Test-set Accuracy Computational Effort

Massively Parallel Computation tests billions of independent models on the data Search Search Search Search Search Search Kernel Kernel Kernel Kernel Kernel Kernel ... Search Search Search Search Search Search Kernel Kernel Kernel Kernel Kernel Kernel CPU Cores CPU Cores CPU Cores Compute Server 1 Compute Server 2 Compute Server N ● Low bandwidth -- transferring solutions ● High latency -- no control flow dependencies

Machine intelligence in action • Predict finish positions of the 2016 Kentucky Derby • Expose relationships between running style, speed, and trainer record • Predicted winner, and 4 out of top 5 horses – Winning Exacta (30:1 odds), – Winning Trifecta (87:1) – Winning Superfecta (542:1) • 1. Nyquist Standardized live odds probability • 2. Gun Runner Speed over the past two races • 3. Exaggerator Post position • 4. Creator Racing style • 5. Mohaymen Track conditions http://performancegenetics.com/machine-learning-algorithm-crushed-kentucky-derby/

Example

Demand forecasting for pharmaceuticals 7/21/2016 Confidential and Proprietary. 17

Optimizing crop yield 7/21/2016 Confidential and Proprietary. 18

Determining causes of customer churn 7/21/2016 Confidential and Proprietary. 19

Conclusions • Machine intelligence extracts meaning from data • Some companies employing machine intelligence today • Many new applications ahead of us Michael Schmidt Founder & CTO michael@nutonian.com www.nutonian.com Blog: http://blog.nutonian.com Twitter: @Nutonian