BSW 2018

Posted on Mon 14 May 2018 in Projects

Deep Learning Deep Dives

Michelle Archuleta

  • Cross polination between ML (NN) and NLP
  • What if color was discrete?
  • Words seem discrete --> sparse matrix
  • instead, think of words on a continuum, represented as a vector (Mikolov 2013)
  • King - Man + Woman = Queen
  • read picture, populate caption (!)

Dean Wyatte

  • Deep Learning-Assisted Creativity and Art (Deep Learning Meetup- see paper)
  • Understanding vs. Predictive Power (along diagonal). Instead of 'Understanding', what about 'Explorability' instead
  • Texture Synthesis
  • to see: Halt and Catch Fire- development of PC
  • Generative Adversarial Networks (GANs)- generator vs. discriminator (inverted CNN to generate images)
  • interpolation between celebrrity faces(!)
  • PSGAN- Periodic Spatial GAN- learn periodic structures in texture, dynamically choose picture size (Bergmann et. al. 2017)
  • to-do.. Google Magenta drum beats: (latent space of drum beats)- interpolation between drum beats

Jeff Rose

  • applied Deep Learning..
  • "Earthscope"- ask questions, investigate processes over the Earth (measurement, rate of change, city growth, population growth, migration)
  • Scale of data is so large- can't do it by hand. New realm of inquiry empowered by ML/AI
  • Data Source: WorldView-3 bird (DG, Westminster), vis/UV/IR. 1 pixel per foot.
  • Judge size of crowd?.. precursers to a revolution.. build up of traffic?
  • Inventory of coal power plants in China, periodic update..
  • Count airplanes in the world.. economic activity
  • Classification: identify a cat
  • Object detection: bounding boxes, tags. Dog and cat in same image
  • Semantic Segmentation: just pixel info
  • Instance Segmentation: which areas contain cat-ness
  • in satellite imagery, segmentation traces mask over every item (calculate volume), or localizes cars to count them
  • cloud computing enables this, over 1000s of servers, GPUs
  • Earthscope output: elevation, parcels, streets, land/water, elevation, customers. Database queries and aggregations to ask about stats: "How many solar panels added in the last 24 hrs"
  • Earth: 500M $km^2$
  • Habitable land: 63M $km^2$
  • 1 sec per $km^2$ of compute time
  • = 5 hrs processible time. So, Earthscope can update every day!
  • Challenges:
    • training data (use TomNod, public draws boxes on images)
    • detection resolution
    • refresh rates
    • business model (source imagery cost): data sold by $m^2$? No longer makes sense, we want it ALL.

python data