Salesforce open-sources the tool it uses to build Einstein’s AI models

A day after Oracle Corp. released the code for an internally developed tool designed to ease artificial intelligence projects, Inc. is joining the fray too.

Salesforce today open-sourced TransmogrifAI, a homegrown piece of software for automating AI initiatives. The tool addresses a different aspect of the task than Oracle’s tool. Whereas the database giant’s goal is to simplify the task of connecting models to applications, Salesforce is focusing on making it easier to build those models.

Developing an AI is no trivial task. That’s particularly true in Salesforce’s case, which processes thousands of organizations’ data throughout the Einstein set of machine learning features that it has built out in recent years.

Salesforce must implement a standalone model for every customer to address to requirement of keeping companies’ records separate. There’s a strong technical incentive for doing so as well: different types of information require different types of AI to process optimally. But manually creating a neural network from scratch for every project is hardly practical, which is how TransmogrifAI was born.

The tool enables engineers to automatically generate a machine learning model tailored for the specific data at hand. It’s touted as a more flexible alternative to existing AI automation solutions such as Google LLC’s cloud-based AutoML service. According to Salesforce, those offerings often only support a narrow set of use cases and are designed for processing unstructured data such as photos, rather than structured business records.

TransmogrifAI automates all the major tasks involved in creating an AI. That starts with the initial data preparation phase, during which the tool can extract important details such as names and phone numbers while discarding inapplicable ones.

From there, TransmogrifAI organizes the data into a specialized numeric form usable by AI models. The tool then runs a variety of different algorithms against the prepared records until it finds the one most suitable for the project.

Salesforce claims tthe tool has brought down the time it takes its engineers to create models from weeks to hours. The company added that TransmogrifAI is also scalable, with the ability to support projects with anywhere from a few thousand to millions of records.

According to Salesforce, it has used TransmogrifAI internally to deploy “thousands of customer-specific machine learning models” that make some 3 billion predictions every day. The fact that the tool is already field-tested at scale should appeal to large risk-averse enterprises working on their own AI projects.

TransmogrifAI is written in the Scala programming language and runs on Spark, the popular open-source data processing engine. Salesforce has published the code for the tool on a dedicated website along with developer documentation.

Photo: Salesforce



UPDATED 15:09 EST . 16 AUGUST 2018

Be the first to comment

Leave a Reply