The demand for AI in the enterprise is insatiable, but the problem lies in creating infrastructure for support and its development and maintenance. IDC 2020 interview found that the lack of data for AI training and low-quality data remain the main barriers to its implementation, along with data security, governance, performance, and latency issues. In fact, a third of the businesses that took part in the survey report spending about a third of their AI lifecycle time integrating and preparing data, compared to actual data science and analytics efforts.
Josh Tobin, a former OpenAI Fellow, observed this trend while teaching a deep learning course at UC Berkeley in 2019 with Vicki Chung. He and Cheung saw the history of AI reach a tipping point: Over the past decade, companies have invested in AI to keep up with technology trends or help with analytics. However, while some vendors claim to be “democratizing AI”, it is still very difficult for most companies to create AI-powered products.
“The main challenge when building or implementing a machine learning framework is that the field is growing incredibly fast. For example, just a few years ago, natural language processing was considered out of reach for industrial applications, but today it is rapidly becoming commonplace,” Tobin said. “That’s why we’re building a platform for continuous improvement in machine learning.”
Tobin and Chung, who previously led infrastructure at OpenAI and were a founding engineer at Duolingo, are co-founders Portal, a service designed to help AI teams decide when to retrain their AI systems and what data to use when retraining. Tobin claims that Gantry, which connects to existing applications, data labeling services, and data warehouses, can summarize and visualize data in the training, evaluation, and deployment phases.
Gantry came out of the stealth today with $28.3 million, a combination of a $23.9 million Series A round and a previously undisclosed $4.4 million seed round. Amplify and Coatue jointly led Series A with investors including OpenAI President and Co-Founder Greg Brockman and Peter Abbil, co-founder of an industrial robotics startup. covariant.
“Our product helps machine learning engineers use the data that goes through their running machine learning product to figure out how the application actually works, find ways to improve it, and implement those improvements,” Tobin said.
AI systems learn to make predictions by collecting datasets (eg historical weather data) and learning the relationships between different data points (eg temperatures tend to be warmer on sunny days) in those datasets. But AI systems tend to be unreliable in the real world because real world data is almost never static, so the training sample is not representative of the real world for a long time. For example, a stock forecasting system could break down because the pandemic is changing buying behavior. Volvo’s self-driving car system was infamous embarrassed kangaroos because the kangaroos jumping made it hard to judge how close they were.
Tobin and Chung believe that the answer to this question is Gantry’s “continuous” learning system, a framework that can adapt the system to an ever-changing data stream. According to Tobin, Gantry is designed to serve as a single source of truth about the performance of an AI system, allowing users to learn how the system is performing and how to improve it, using workflow tools to define the metrics and data slices on which they are calculated.
“Gone are the days of poor corporate customer service—customers now expect the same seamless, consistent, and intuitive experience they’ve come to expect from today’s technology companies. Machine learning allows you to scale these capabilities. However, building products based on machine learning is costly and poses a risk to the brand and customer interactions, as models can unexpectedly and harmfully fail when interacting with users,” he added. “Gantry helps enterprises create a lower-risk, lower-cost machine learning customer experience by providing the infrastructure and controls needed to securely maintain and replicate the features of machine learning-based products.”
Gantry is part of a new category of software known as MLOps (Machine Learning Operations) that seeks to optimize the AI system lifecycle by automating and standardizing development workflows. Analyst company Cognilytica, driven by the rapid adoption of artificial intelligence predicts that the global market for MLOps solutions will be worth $4 billion by 2025, up from $350 million in 2019.
Tobin acknowledges that other tools such as Arize, Arthurand violinist, perform some of the same things as Gantry. But he argues that they focus on a broader range of AI issues, while Gantry touches on—but goes beyond—observability, monitoring, and explainability. For example, Gantry can be used to detect bias in AI-based applications, Tobin argues, even if the applications use “unstructured” data such as text and images.
Tobin declined to say how many users or customers Gantry has. But he says the funding will go in part to acquiring customers as well as growing Gantry’s team of 22.
“We believe the potential headwind in technology is more than offset by a strong tailwind in machine learning,” Tobin added when asked about the current economic climate and what that could mean for Gantry. “Furthermore, as the belts tighten and companies become more conscious about their costs, investing in tools to help improve team performance and product performance and reliability becomes even more important.”
Credit: techcrunch.com /