Google ML summit — 2019

I attended Google developer’s ML summit last week and it was a brilliant experience to meet machine learning developers at Google and learn the newest innovations from different industries.

Here are a few things I would like to highlight out of the day-long enlightening presentations.

1. Depth supervision by Forrestor Cole:

• A depth map is used to generate training data. It depends on the distance from the camera.
• Images are obtained from Slam or ARcore camera trajectory.
• The ground truth depth is achieved from multi-view stereo. It aims at creating a pseudo depth map and optical flow from FlowNet.
• The presentation videos were obtained from mannequin challenge.
• This algorithm is further implemented for object insertion.
• It is also used in inpainting humans.
• For future work, check this out https://mannequin-depth.github.io/

2. ML deployments with TensorFlow by Hannes Hapke

• Creating an ML infrastructure in tensor flow which acts as an alternative to flask- a tensorflow serving
• In this approach, we first export the algorithm to a docker image
• Fetch the docker image which is a container deployment
• Serve it to — Tensorflow serving path
• Feed version labels
• This method utilizes GRPC remote procedure protocol
• It is just similar to a REST application
• In this approach, Base64 is a protobuf format
• gRPC — Google remote procedure call
• Advantages are that model interferences are faster and the required field is defined easily.

3. ESR GAN by Victor Dibia

• GANs used for generating face masks
• Moreover, the improvement in pixels was a great thing when I compare it to previous papers on CycleGAN and DCGAN
• Acquainted with the concept of

1. Bucketing training from how far they to the training extrapolation
2. few-shot learning
3. GAN in outlier detection
4. Triplet loss

4. Fusing AI with AR by Stephen Wylie

• Using ARcore and Sceneform
• Establish an android app
• The ranking algorithm here tech — https://github.com/mrcity/mlworkshop/
• Use cases:

1. teachable machine
2. scratch MIT
3. Life sciences
4. grab a coffee
5. pair codes GitHub federate learning