- want to predict poverty using satellite images
- poverty prediction currently done through day and night images since this is less expensive than on ground surveys (day + night -> poor)
- can model using just day imagery (day -> night, day + generated night -> poor). benefit here is less data needed.
- but even better, maybe can do day -> latent -> poverty and directly learn what the latent is instead of assuming night captures the latent
- train the latent based on the night. experimentation here.
- day -> poverty (discriminative)
- night -> poverty (discriminative)
- day / night -> poor or not (discriminative) imgs -> VGG features -> fully connected -> sigmoid
- for now doing binary classification, can do 5-class, or regression.
- should get more data (right now 1k, we want 41k).
- You may wish to consider certain semi-supervised learning baselines such as self-consistency to compare against the efficacy of the GAN-based approach.
- day -> night, night-> poverty (generative) train two step day -> night is generative (pix2pix) day + generated night -> poverty is discriminative
- take night, train autoencoder, take those latents take day, train to generate those latents take latent, train to get poverty then given day, can generate latents,
- conditional multiclass gan (adversarially learned inference) generator conditioned on daytime image generates latent embedding discriminator classifies embedding as fake/poor/rich
- generator -> conditioned on poverty, generates day (or night or day & night) discriminator -> given day / night / both, predicts poverty (fake | poor | rich) see if match up