Meetup 18/10/2018 - Artificiële intelligentie en mobiliteit
- 3. The Problem
Access to Deep Learning Solutions
● DL very powerful
● Limited availability of experts
● Implications?
○ Subject matter expertise
○ Screen applications
○ Waste of experts’ time
● Lots of untapped potential!
- 5. What does it look like today?
1. No more feature engineering
2. Repurposable through Transfer
Learning !
- 6. What is the problem?
1. Seemingly easy
a. Play with hyperparameters to make a
model ‘learn’
b. Make the blue curve go up!
2. Hyperparameter space is huge
3. Building intuition requires insight & experience
4. Paying us to make a computer learn
- 7. Can’t a computer learn how to learn?
● What Google suggests
○ Neural Architecture Search
○ 100-1,000x computational power
● Stick with Transfer Learning!
- 8. Can’t a computer learn how to learn?
● Important aspects in CV
○ Collecting & cleaning data
○ Selecting a model
○ Configure your model
○ Hyperparameter Optimization (HPO)
● Computer Vision is an ideal candidate!
- 9. How does it work?
● Black box optimization
○ We can’t evaluate the derivative of the objective function
→ Like searching in the dark
- 10. How does it work?
● Many options!
○ 1 evaluation = how well does the model learn with this set of hyperparameters?
■ Excludes many optimization strategies!
- 11. How does it work?
● Bayesian Optimization
○ Build a probability model of the objective
function
○ Treat OF as random
○ Prior → Posterior → Acquisition function
○ Spearmint - SMAC - Hyperopt - RoBO
- 12. How does it work?
● Problem with Bayesian
○ Sequential
○ Random takes many evaluations
○ Solution: behave like a gambler on slot
machines
→ Hyperband
- 13. How does it work?
● Hyperband
○ Run random candidate settings
○ Limited iterations per candidate
○ Revisit good candidates
→ 70 times faster than random
- 14. How does it work?
● Hyperband on steroids = BOHB
○ Bayesian Optimization (BO)+ Hyperband (HB)
○ Speed of HB + optimization performance of BO
- 15. Our Experience
With existing tools
● Client work
● (+) Within 5% accuracy range
● (-) Costly
● (-) Closed-source
● (-) Can’t export models
- 16. Our Experience
With own research
● Traffic Analysis
○ Future project
○ Implemented simple POC
○ Combining with Synthetic Data
● Feel we can do better
○ Interested parties to collaborate?