Machine Intelligence for Architects

How can we apply machine learning in the field of architecture and design? During his presentation at AIA Boston, our CEO Dan Kuster discussed ways we can use AI to enable generative design. He also explored social factors, the use of sensors, and how we can optimize the business of architecture with machine learning.

Transcript

1. MACHINE INTELLIGENCE FOR ARCHITECTS Daniel Kuster, Ph.D. founder + CEO at cambr.io
 @djkust
2. USING COMPUTER-AIDED DESIGN TO AMPLIFY HUMAN CREATIVITY + INTELLIGENCE Daniel Kuster, Ph.D. founder + CEO at cambr.io
 @djkust
3. Cambrio is a digital R&D lab, focused on AI and machine learning. We use computers + data + scientific thinking to solve problems. We love hard problems. 
 If you have one, please reach out to discuss! <hello@cambr.io>
4. My first job was detasseling corn. Hot. Tedious. Boring. Manual labor.
5. Like the stone age.
7. Well, that isn’t quite accurate.
9. You’ll pay me to design things using a computer? me, 20 years ago
11. ~1996 AutoCAD R12:
 draw in 2D planes 3D building
12. ~2004 + Use CAD to design many different things!
 Human creativity + exhaustive search
13. How to represent objects in the computer?
 How to generate candidates?
 What are the constraints? Why bother? How does “computer-aided” add value?
14. How you do that? atomize problems tasks metric search apply constraints
15. How do you know the constraints? must learn the physics of the problem…

16. physics code? …this could take years
17. Hey man, what if we could automate the discovery of “the physics”? machine intelligence
18. Problems solved in many domains!
19. Common pieces Problem worth solving Tasks to optimize Data Hypothesis Metrics Rules for learning
20. Let’s illustrate, using a benchmark problem
21. Problem: 
 USPS wants to automate mail sorting using zip codes wearing headphones! boring job ;)
22. Task: Optical character recognition
23. Data: 60k labeled images of digits from USPS
24. Hypothesis: 
 digits can be recognized as patterns of patterns
25. Metric: 
 actual label = predicted label? @genekogan
26. Better hypotheses -> stronger results @alecrad 1998: ~12.0% error (linear classifier)
 2007: ~0.5% error (kNN with data aug)
 2012: ~0.2% error (convnets)
27. Classification -> generative models
28. Classification -> generative models
29. Classification -> generative models
30. Classification -> generative models
31. Classification -> generative models
32. Classification -> generative models
33. Classification -> generative models
34. Classification -> generative models
35. OK, what about architecture?
36. All the pieces are ready Problems worth solving Tasks to optimize Data Hypothesis Metrics Rules for learning
37. All the pieces are ready Problem: initial layout and design Task: boring shuffling/iteration
 (e.g., 40-person conf rooms tend to be similar) Data: BIM Hypothesis: optimizing for human factors yields better design Metrics: daylight, view, noise, distance to coworkers, … Rules for learning: simulate + compare
38. Building design process: constraints
 how it is done (manually) today credit: studio192.co
39. credit: studio192.co Building design process: constraints
 how it is done (manually) today …iterate, translate into Revit, render views, etc.
40. Data is a good industry readiness indicator…
 it is hard to fake the culture
41. BIM: a boundary object for architecture + IoT/sensors after building is built
42. BIM: a boundary object for architecture
43. AutoDesk MaRS project: 
 https://vimeo.com/193915345
44. AutoDesk MaRS project: 
 https://vimeo.com/193915345
45. AutoDesk MaRS project: 
 https://vimeo.com/193915345
46. AutoDesk MaRS project: 
 https://vimeo.com/193915345
47. AutoDesk MaRS project: 
 https://vimeo.com/193915345
48. AutoDesk MaRS project: 
 https://vimeo.com/193915345
49. AutoDesk MaRS project: 
 https://vimeo.com/193915345 Generate Evaluate ~Exhaustive search
50. AutoDesk MaRS project: 
 https://vimeo.com/193915345 Evaluate thousands of designs vs. metrics
 (minimize travel distance)
51. AutoDesk MaRS project: 
 https://vimeo.com/193915345 Evaluate thousand of designs vs. metrics
 (maximize exterior views)
52. How to evaluate each candidate for each metric? 1. Simulate the physics
 (how AutoDesk did it this year)
 good for stuff we already know, like daylight
 2. Learn from the data 
 (via AI / ML in the future)
 good for subjective + social factors, sensors, etc
53. Q&A: 
 What problems are you facing that might be solved using data? Problems worth solving Tasks to optimize Data Hypothesis Metrics Rules for learning
54. IMAGE CREDITS US National Archives: detasseling
 NASA: space station
 NIH: lung cancer CT scan
 AutoDesk: MaRS project + BIM cartoon
 indico.io/blog: MNIST diagram
 MufasuCAD.com: old AutoCAD screenshot
 virascoop.com: alpacas
 kasisto.com: iOS chatbot
 wired.com: filling potholes
 Wikipedia: Field work, Walkman, Grandview, Octane workstation, protein render, markets screen, med evac, lymphocyte, offshore rig
 Matan et al. “Reading Handwritten Digits: A Zip Code Recognition System” AT&T Bell Labs (1991): MNIST examples
 Teterboro News: mail sorting
 Gene Kogan + Alec Radford: MNIST visualizations
 FastCoDesign: automotive clay model