Startup Weights & Biases Inc., whose software is used by researchers at OpenAI and Nvidia Corp. to develop new artificial intelligence models, today disclosed that it has closed a $135 million investment at a $1 billion valuation.

The Series C funding round was provided by Felicis Ventures, BOND, Insight Partners and Coatue. Weights & Biases has raised $200 million from investors to date.

Enterprise-grade AI models are developed through trial and error. Developers create multiple versions of the same neural network, test them and then select the one that proves most suitable for their project. For example, if a company plans to run its AI model on a battery-powered connected device, it might choose the version of the AI model that can perform computations using the least electricity. 

Weights & Biases has developed a machine learning operations or MLOps platform that increases the efficiency of the trial-and-error process through which AI software is developed. According to the startup, the platform improves developer productivity by solving a key challenge involved in AI initiatives: organizing and processing project data.

After a company chooses which version of a neural network to use, deploying the software in production isn’t as simple as installing a file. On its own, a neural network can’t produce useful results. There are numerous associated data assets and other files that need to be combined with the neural network for it to work as intended. Finding exactly which files are associated with what AI model is tricky, especially in a large enterprise where developers may have thousands of neural networks.

According to Weights & Biases, its platform speeds up the task to help software teams deploy AI models in production faster. The platform collects all the files associated with a neural network and organizes them in one centralized interface. The result, the startup says, is less manual work for developers. 

One of the key items that Weights & Biases keeps track of is an AI’s training dataset. There are often multiple versions of the same training dataset in a project, which makes it difficult to keep track of which one should be used. The platform’s centralized interface displays details on the effectiveness of each training dataset to help developers quickly find the one that can help them meet project requirements most effectively. 

The training dataset from which an AI learns how to perform a task is only one of several components that influences the neural network’s performance. Hyperparameters are another. Hyperparameters are configuration settings that define key operational details of an AI, such as how many artificial neurons it has and the way training data is processed. Weights & Biases organizes an AI’s configuration settings, too, to make managing machine learning projects easier.

Weights & Biases provides its core feature set alongside complementary capabilities designed to simplify some of the other steps involved in building AI software. The startup’s platform visualizes technical information on neural networks to help inform developers’ work. Software teams have access to graphs that detail the accuracy of an AI model, the amount of processing capacity it requires to run and other key details.

Weights & Biases also helps with hyperparameter tuning, the process of tweaking an AI model’s settings to maximize its accuracy and speed. The platform tweaks hyperparameters using a technique known as Bayesian optimization.

The complexity of modern AI software means that neural networks’ settings can’t be optimized with manual methods alone. Usually, developers create scripts that try a large number of setting combinations in a partly randomized way. 

The Bayesian optimization technique used by Weights & Biases takes a more focused approach. The platform tries out a combination of settings and then, based on the results from the experiment, narrows down its search to ensure that the next attempt will be more successful. This technique can produce better results than more randomized approaches.

Weights & Biases says its platform is used by more than 100,000 machine learning practitioners at OpenAI, Nvidia, Qualcomm Inc., Lyft Inc. and other major tech firms. The startup has doubled its headcount since the start of 2021 to keep up with demand. Now that it has an additional $135 million on the books, Weights & Biases intends to continue expanding its operations and will in parallel work to develop more features for its platform. 

“This funding allows us to continue building out our platform, adding new and better functionality, and scaling with the growing needs of the ML community,” said Weights & Biases co-founder and Chief Executive Officer Lukas Biewald.

This round is the latest of several raised recently by startups working to simplify AI development for enterprises. Domino Data Lab Inc. received $100 million this month in a round a backed by Nvidia. Earlier, DataRobot closed a $100 million investment that valued it at $6.3 billion. 

Image: Weights & Biases