X
Innovation

AI startup Abacus goes live with commercial deep learning service, takes $13M Series A financing

With $18.25 million in venture capital money from Eric Schmidt, Jerry Yang, and other luminaries, Abacus Tuesday opened its deep learning service for general customer availability, with the intent of making it easy for companies to customize and scale AI in the cloud.
Written by Tiernan Ray, Senior Contributing Writer
abacus-dot-ai-founders-july-2020.jpg

Abacus co-founders, from left, Siddartha Naidu, previously a principal engineer for Amazon's fulfillment team and also a developer of the BigQuery software at Google; Bindu Reddy, previously head of "AI Verticals" for Amazon's AWS; and Arvind Sundararajan, previously engineering lead for Google's ad delivery technology. 

Abacus

Just under six months after coming out of stealth mode, startup Abacus dot AI of San Francisco Tuesday announced the company's service for commercial deep learning has gone live with customers such as 1-800-Flowers, and the company has gotten a Series A investment round totaling $13 million from major investors including Index Ventures. 

"This is a crowded space and very few AI/ML services actually manage to get customers to production and actually realize a positive ROI," Bindu Reddy, co-founder and chief executive officer, told ZDNet in email.

As with its last major announcement, in January, the company also demonstrated technology for a novel approach in deep learning, in this case offering a new technique for de-biasing AI models. The software can do things such as tweak an existing program so it more fairly and accurately categorizes whether Black people in a photo are smiling. 

Abacus, which began life a year ago March as Reality Engines, is using a variety of approaches to deep learning, including, in particular, generative adversarial networks, or GANs, to offer a kind of push-button service in the cloud that lets companies train and test and deploy novel AI programs without the hassle of traditional laboratory deep learning work. 

Also: Reality Engines offers a deep learning tour de force to challenge Amazon et al in Enterprise AI

The service sits somewhere between the incredibly complex, and resource-intensive efforts of AI giants Google and others, and the simplistic, cookie-cutter, push-button chatbots and other watered-down AI offerings that can be found in abundance on every cloud platform.

Abacus uses GANs to design, train, and test deep learning algorithms tuned to clients' needs, and then the finished model can be deployed on a customer's cloud platform of choice. The stated goal of the company is for clients to obtain the models "on a self-serve basis," what Abacus calls an autonomous product.

The name change to Abacus is more reflective of what the company is building, said Reddy, which she described as "an autonomous AI service that helps everyone create and deploy a deep learning service in production."

"Given that, we decided to change the name to something short and simple," added Reddy.

Among the 35 initial customers, 1-800-Flowers, an early adoptee, will continue to use the service as it goes general availability. The company is using Abacus to "easily create deep learning systems for various use-cases, including personalized recommendations, churn reduction and fraud detection," according to 1-800-Flowers' chief marketing officer Amit Shah. 

After developing the model based on 1-800-Flowers' data set and its use case, said Shah, Abacus sets up all the infrastructure needed to deploy the model in production and to periodically re-train it and increase its capacity for an increasing volume of prediction requests.

As it has in the past, Abacus peppered its business announcement with a display of novel technology. In this case, the de-biasing technology is what is known as a "post-hoc" approach to de-biasing. It is in a beta stage and is being turned on for a few select customers, said Reddy. 

The technology uses generative adversarial networks as the main route to a solution. Bias in this context can be defined by a bunch of measures of performance of a neural network, such as, for example, the rate of true or false positives for a protected group, such as Black people, versus a non-protected group, such as white people. Closing that gap, coming to parity, is the statistical definition of minimizing bias in this context. 

To achieve parity to reduce bias, the authors developed a novel adversarial approach where a second neural network is used as a critic to predict the bias in the first neural network. The critic network's predictions nudge the weights of the first network until those weights settle into a less-biased configuration, via the typical deep learning optimization method known as gradient descent. The work is laid out in a paper posted on June 15 on the arXiv pre-print server

Also: IBM offers explainable AI toolkit, but it's open to interpretation

Colin White, a research scientist at Abacus who helped to develop the approach, explained to ZDNet that post-hoc de-biasing is desirable when the processing demand of re-training models from scratch to eliminate bias is prohibitive. 

"Big companies have released very large language or object detection models which were trained with hundreds of GPUs, and it is common for other orgs to fine-tune for their specific ML use-cases," said White. "But in order to de-bias these behemoth models, we would need to use a post-hoc technique, unless we have hundreds of GPUs."

In that sense, post-hoc de-biasing is a form of fine-tuning of existing models. Especially when the original training data is not available, such ad-hoc processing may be one of the only ways to fix a biased network. (Abacus is hosting a virtual meet-up on July 20th to discuss the topic of bias.)

Abacus's new venture capital money brings with it new investors such as Mike Volpi, co-founder of Index Ventures, and a board member of companies such as Elastic and Pure Storage. Also joining are several investors in last year's seed round of financing, including former Google CEO Eric Schmidt, investor and former Amazon executive Ram Shriram, and Yahoo! co-founder and onetime CEO Jerry Yang. The Series A brings total funding in the two rounds to $18.25 million. Marty Chavez, former CFO of Goldman Sachs, is joining the company's board of advisors. 

Abacus employs 23 people. Founders include Reddy, who was previously head of "AI Verticals" for Amazon's AWS; CTO Arvind Sundararajan, who previously was the engineering lead for Google's ad delivery technology; and director of research Siddartha Naidu, who was a principal engineer for Amazon's fulfillment team and also a developer of the BigQuery software at Google prior to that.

Abacus noted its investor cohort includes "several prominent women investors," among which are Mariam Naficy, founder and CEO of online marketplace Minted; Neha Narkhede and Erica Ruliffson Schultz, the CTO and the president of field operations at streaming platform Confluent, respectively; Jeannette Catherine, hereditary princess of Fürstenberg; and investor Xuezhao Lan, who is founding and managing partner at San Francisco's Basis Set Ventures.

CEO Reddy told ZDNet the areas of focus this year for the company will include efforts to "refine and revise our techniques to build accurate models". It ill also continue to build out new features "including data pipelines, re-training policies, and model monitoring to put models in production easily," and to gain more clients who want to be early adopters and "take an AI-first approach," as she put it.

For her part, Reddy told ZDNet among her goals as chief is to "to re-create the vibe, energy, and pace we used to have in the office, virtually." 

The pandemic, said Reddy, has presented new challenges to the team "when it is easy to get stuck on an issue or feel frustrated when you are up against a wall."

With everyone working from home, new approaches to socializing have included ad-hoc slack chats, team members chatting virtually over lunch, and "weekly TGIFs on Zoom." Reddy said Abacus also has a weekly show and tell session for staff to talk about what they're building.

"I still think this doesn't compare well to being in the same room and being able to brainstorm and have ad-hoc discussions that are intellectually stimulating," Reddy told ZDNet.


Editorial standards