01
The paperYoshimura, Cai, Wang & Ratti · 2018
- Convnet to classify works belonging to 34 different architects. 77% accuracy.
- Using images scraped from he web, and some photos taken by researchers
- Architects then clustered together and compared to conventional architecture theory
- Difficulty in classifying modern architects - detecting space, not features
- Uses NASNet - Zoph and Schlens (2018)
- Gradient-weighted Class Activation Mapping (Grad-CAM). Enables us to understand the focus of the machine eye for classification
- Dimension reduction and clustering - Linear principal component analysis. k-means clustering
- 34 architects - all past Prisker Prize winners
- 20,000 total sample images
- Goolge Tensorflow. Two GeForce GTX 1070Ti. Training complete in 8 hours.
- 30 epochs.
- Tends to confuse Koolhaas, Holl, Perrault with other architects
- Correctly distinguishes Kahn, Siza & van de Rohe
- “Top 1 accuracy indicates the probability of whether the image can correctly match with the target label. Conversely, the top 5 accuracy suggests the probability of whether the correct image can appear with the target label among five pictures ordered according to their highest probability.”
- Almost 70% of architects can be distinguished with more than 80% probability (Top 5 accuracy)
- 45% distinguished with 90% probability (Top 5 accuracy)
- Indoor scenes more distinguishable to machine eye (~5% more accurate)
- k-means clustering:
- Foster, Renzo, Rogers (High-Tech Design)
- Lloyd Wright, "Normal House" (Prarie style)
- Ando, Eisenman, Miralles (Deconstructivists)
- Mayne, Gehry
- Does not distinguish Architect's work over time
We help professional services firms work out where AI genuinely helps, what has to be structured first, and what to leave alone.
Get in touch