🔛🔝Salesforce Einstein Brings AutoML Models to a Massive Scale
How do you see the application of AutoML in services like Einstein?
📝 Editorial
When Salesforce launched the Einstein platform in 2016, I was a bit skeptical about it. My perspective at the time was that it was going to be difficult to create predictive models that could adapt to the different configurations of Salesforce data across hundreds of thousands of customers. Think about it, every company has unique ways to model sales and marketing pipelines, so how could you possibly create low-touch predictive models that learn those things on the flight? During that time, techniques like AutoML were still seen as a pipe dream, so don’t judge me too harshly 😉. Well, four years after its launch, the Salesforce Einstein platform is processing an astonishing 80 billion predictions per day, which makes it one of the largest B2B machine learning platforms in production.
The success of Einstein is directly tied to the recent improvements in AutoML stacks. Specifically, Einstein is built on a proprietary AutoML stack known as TransmogrifAI (Edge#2 is about AutoML and TransmogrifAI). The use of AutoML allows Einstein to tune the hyperparameters of models in order to adapt to specific Salesforce datasets. There are plenty of examples of AutoML services in the market but the production deployments of those methods remained constrained to very narrow scenarios. From that perspective, Einstein must be one of the best examples of AutoML in production and a glimpse into the future of machine learning in SaaS products.
How do you see the application of AutoML in services like Einstein?
🗓 Next week in TheSequence Edge:
Edge#43: the concept of Bidirectional Long-Short Term Memory Networks; DeepMind’s Differentiable Neural Computer; Google Fairness Indicators as a key component of any ML model.
Edge#44: deep dive into AI behind DeepMind’s Agent57, which outperformed humans in 57 Atari games.
🔎 ML Research
Assessing Alexa Conversations
Amazon Research published a paper detailing a deep learning model used to estimate how customers would rate their satisfaction with dialogue interactions ->read more on Amazon Research blog
Better Semi-Supervised Learning
Salesforce published a paper introducing CoMatch, a new semi-supervised learning method for image classification ->read more on Salesforce Research blog
Smart Scrolling
Google Research published a detailed blog post describing the architecture of a neural network used to improve the navigation experience of its recently released Recorder app ->read more on Google Research blog
🤖 Cool AI Tech Releases
Horovod v0.21
Uber released the new version of Horovod, its framework for large scale deep learning training ->read more on Uber blog
💬 Useful Tweet
💸 Money in AI
Loyalty automation platform Glue raised $8 million. They help customize loyalty strategies, using AI for analyzing data points from 100,000 organizations and automating interactions between the company and clients in a mobile app.
Industrial robotic startup Percepto raised $45 million in a Series B funding round. They use proprietary computer vision and deep learning algorithms to better navigate the robots in real-time. The collected data is used in ML models for better decision making for any given mission. Reinforcement learning is actively used to match the specific needs of the client.
Speech-to-code startup Serenade raised $2.1 million in the seed round. They created an engine that transforms natural speech into code.
Autonomous delivery startup Gatik raised $25 million in a Series A round. It seems that the pandemic greatly hastens the adoption of autonomous vehicles and autonomous vehicles for delivery.
TheSequence is a summary of groundbreaking ML research papers, engaging explanations of ML concepts, exploration of new ML frameworks, and platforms. It also keeps you up to date with the news, trends, and technology developments in the AI field.
5 minutes of your time, 3 times a week– you will steadily become knowledgeable about everything happening in the AI space.
Jesus, Don't be too harsh on yourself, the majority of us did not see the intersection of semi controlled operating environments like Salesforce and Workday and AutoML.
Einstein's success proves that industrious companies can bring together a bounded set of information and automation to make predictions a production quality service. We can't call it PaaS, that has been taken. Maybe AIaaS, will catch on.
I expect to see more and more of these offerings coming to market over time.
I have spoken extensively about AutoML being a technology that will help move predictions to a higher level of quality and accuracy. When all the low hanging fruit can be automated away with AutoML, human talent will concentrate on higher quality, and value, targets in their work.
Thanks for sharing all the great thoughts and content. I read every issue of the newsletter. Have a great Sunday!