|

Salesforce’s Model Tester Adds to Firm’s DIY AI Toolkit

Salesforce wants chatbots to expect the unexpected with the latest patent for it’s DIY AI toolkit.

Photo of a Salesforce patent
Photo via U.S. Patent and Trademark Office

Sign up to uncover the latest in emerging technology.

Salesforce wants to make sure its AI models aren’t acting up. 

The company filed a patent application for a way to provide an “automated testing pipeline for neural network models.” Salesforce’s tech allows developers to estimate whether or not their models will remain accurate when facing “unseen queries,” using a customer service bot as its primary example.

Typically, people will test their models using a portion of the data originally used in training, Salesforce said. “However, this evaluation procedure wouldn’t be an optimal estimate for smaller datasets or when real-time distribution varies significantly from the original training dataset distribution.”  

To overcome this, Salesforce’s tech essentially curates both easy and hard evaluation datasets from real-time customer data to make sure the model can handle a diverse array of user queries. A hard evaluation dataset may use samples that are significantly different from what the model was trained on; an easy one would line up more closely. 

This system passes customer data through several filters to create these datasets. First, it goes through a “dependency parser,” which identifies and filters specific actions or verbs that represent meaningful commands. Next, a pre-trained language model ranks the queries based on how similar they are to the data the model was trained on. A “bag of words” classifier then removes those queries, ensuring the testing data is sufficiently different from the original training data. 

From this process, both the easy and hard datasets are created, and used to test the model’s capabilities. Salesforce also notes that this pipeline includes a “human-in-the-loop” feedback cycle to communicate when a model may not be performing up to par, which lets developers make adjustments.

Salesforce’s main AI offering is Einstein, its suite of solutions that allows customers to build generative AI experiences with their data. While other firms plunge their resources into building billion-parameter models, Salesforce’s strategy has been to teach a man to fish — the “man” being AI-hungry enterprise clients and the “fish” being the models themselves, said Bob Rogers, Ph.D., the co-founder of BeeKeeperAI and CEO of Oii.ai

This patent could add another tool to that kit, ensuring the models created from its umbrella of offerings are working as intended — and well, said Rogers. “I think Salesforce wants Einstein to really generate more leads and faster. And if that’s not happening, and people are putting a lot of time and money into it, then that’s kind of a miss for Salesforce.” 

With this patent, Rogers said, “I’m wondering if they were trying to get to a way to automatically help their customers not deploy bad models, because the self-service concept is a big core idea of Einstein.” 

The filing’s focus on improving customer service chatbots may also hint at Salesforce’s interest in AI-based customer service interactions, said Rogers, especially as it unveiled its fully-autonomous Einstein Service Agent last week. “It says something about where Salesforce thinks the traction for Einstein may be going.” 

But creating tools that allow people to build their own AI is “harder than it sounds,” said Rogers. And in a landscape crowded with companies like Google, Microsoft, and OpenAI that provide easy-to-use services, more often people rely on other models to do the hard work. “At the end of the day, most AI utilization is still people saying, ‘solve my problem for me,’” said Rogers.