Machine Learning for Mere Mortals

We are here with about 250 others learning from Googler's and GDE's (Google Developer Experts) who are focused on transferring ML know-how to application developers. A couple are seasoned data science practitioners who have retained from prior generation technologies and bring a rich ability to advocate and teach today's ML/AI to a mostly younger cohort, although the diversity is very inclusive. page

ML technologies are now mature enough for cloud based ML engineering assisted by AI. This looks like an important democratization to a far wider cohort of developers. With the departure of Dianne Green--who is now on the supervisory board of SAP--Thomas Kurian who succeeded as head of Google Cloud has tuned strategy to more familiar SQL console based ML cloud development. Many of the data science engineering ops tasks have been productized to allow emphasis on model development and deployment by developers who understand how to develop and their business domain.

I noticed by the way Google's cloud console (scroll down the left menu) now makes its platform available for application developers to act as data scientists. The real feature engineering and raw ML expertise is still going on but concentrated in the platform by the research data scientists and engineering cognescenti.

e.g. most data scientist ML models were not getting deployed into prod [same story as blockchain]. The solution is to make ML easier for business developers via cloud platform based ML, and a shift from local CPU model training to GPU/TPU in the far faster cloud.

Some examples. Picture light because this wiki is hosted in Singapore to test latency issues and limited to 4GB (high images are not my friend).

# Focus on more devices and privacy. - Exported models can be deployed to devices and browsers. - KubeFlow deployments with model routing via Istio - Federated Learning will have an impact on [privacy and confidentiality] deployments - Tips PoC with ReST, Prod with RPC (protobuffs).

# Switch to Big Table for ML - Google Search database become overwhelmed with sharding data, the internal version was down for 1 week (not the consumer search) -- giving rise to Big Table for high throughput and many reads. ML has the same consumption pattern, hence it is used for Big Query ML.

# Deep Learning for Robot Navigation - Real world robots are hard because the environment keeps changing (think a lab or Starbucks). Thus ML the model needs a human to intervene and update the model. So what do we do, We look to biology and use genetic learning that breaks each layer down and uses neural learning and AutoRL. This uses rewards to train each level.

- 1 Practice Run AutoRL - 2 Given neighborhood and a point, learn to predict point's importance (example indoor navigation at Googleplex). - 3 In unseen environments: Use important points as hubs.

# AI/AR Android Apps - Google's Federated Learning is beginning to look JRP (just real practical) even on Android apps with AR. - Make a warehousing app that inspects boxes w/o an operator opening them up. (tensorflow/tensorflow/lite/examples/android/app) build with bazel. @SwebCEO goshtastic,blogspot.com github.com/mrcity/mlworkshop slideshare: StephenWylie3

# TF 2.0 designed for Production

- Prod AL/ML pipeline - Move pre-processing into the graph. From CPU to GPU/TPUs (from TF1 to TF2). - Data preprocessing does not need to be redone at deployment. -Start by subclassing (e.g. subclass for custom data preprocessing layer, i.e Existing Model (Pre-stem) no backprop. - What else can I move into the graph? - Benefits: Data preprocessing does not need to be redone at deployment. - Deconvolution see CIFAR-10 (paper)

# Convolutional Neural Networks with Swift

- Swift added functional programming concepts to old C languages. - Programming DNN ML uses the backprop implements the convolutional chain. - Swift-clang is highly optimized for compilations and thus efficient backprop. page