Issue #8: Deep Learning Platform, TinyML, Privacy ↔ Contact Tracing

SubscribePrevious Issues

Data Exchange podcast

  • An open source platform for training deep learning models  Evan Sparks describes the newly open sourced Determined Training Platform for training deep learning models. It includes components for distributed training and hyperparameter tuning, experiment tracking and tools for collaboration and governance, a scheduler specialized for DL workflows, and more.
  • Why TinyML will be huge  Pete Warden explains the enormous impact of deep learning on embedded systems in tiny devices, and why he has chosen to focus on building tools for ultra-low-power systems.
  • Human-in-the-loop machine learning  Rob Munro describes his long standing interest in ML and NLP, and building real-world human-in-the-loop systems. He shares insights about creating an effective customer-centered view of ML-based products.

Machine Learning tools and infrastructure

  • Introducing RaySGD   Distributed model training is difficult to set up and expensive to run. RaySGD is a new library that makes distributed PyTorch and TensorFlow model training simple and cheap.
  • A system for massively parallel hyperparameter tuning   An interesting new paper from the group of researchers behind Hyperband and other state-of-the-art hyperparameter tuning algorithms. Their new algorithm outperforms existing state-of-the-art hyperparameter optimization methods, and is suitable for massive parallelism.
  • Machine learning and microcontrollers   Google’s Pete Warden has a great series of screencast videos on how to enable ultra-low power machine learning at the edge.  No machine learning or microcontroller experience is necessary, and you can train models small enough to fit into any environment.
  • Reinforcement Learning in Public Policy   A group of researchers from Harvard and Salesforce Research used RLlib to derive AI-driven tax policies based on economic simulations. RL is used to model interactions between different players in the economy, including workers and governments.

COVID-19

  • Moscow’s Facial Recognition Tech Will Outlast the Coronavirus   Here’s a compelling, high-stakes reminder that computer vision and other technology needed for contact tracing can also be used for mass surveillance.
  • Data collection and unification for forecasting epidemics  AI is the headline in this 60 Minutes segment, but I believe the key is the combining of multiple data sources. A similar data unification project in the UK (“Data can save us from COVID-19”) has produced real-time dashboards that give senior policy makers the information they need to make sound decisions.
  • Epidemic Modeling 103  Bruno Gonçalves describes how you can add confidence intervals and stochastic effects to your CoVID-19 models, to address common limitations of an epidemic model. Bruno was a recent guest on the Data Exchange podcast.
  • What Happens Next?  Speaking of epidemic simulations, the visualizations on this site are great teaching tools.  Check out these playable simulations to gain a greater understanding of what’s ahead in the next few months and years.

Virtual Conferences

Here’s an  are updates on events I’m involved with:

  • The Future of Machine Learning and AI   Two award-winning researchers – Michael Jordan and Ion Stoica – will speak on May 13 on the interplay between machine learning, decision science, and economics, and on the growing importance of distributed computing. Register here.
  • Automatic Forecasting   At the recent MLOps Virtual conference which I hosted, Perry Stephenson described how he uses the Databricks Platform to develop custom forecasting models for many different groups within Atlassian. He provides a very practical approach for how one can build, deploy, and manage many different ML models.
  • Understand the Ray ecosystem in a few minutes   I recorded this brief video to explain why Ray is generating buzz among machine learning enthusiasts and Python developers.

Work and Hiring


Recommendations



[Image: Newsletter from Pixabay.]