AI Startup: Weights & Biases Eases Hyperparameter Tracking


San Francisco startup Weights & Biases (WB) is making a name for itself in the AI tools market. As an “experiment tracking platform for deep learning”, WB allows users to collaborate, “visualize model performance”, and track model versions, code, workflows, training models, configurations, and more. Since the basic process of training models is rinse and repeat, WB makes it easy to conduct model experimentation using different configurations and hyperparameters without getting bogged down in the complexity.

One area that WB excels at is working with hyperparameters. Hyperparameters are elements (knobs) of a model that must be tuned to control behavior and achieve a certain outcome. Hyperparameter Tuning or Hyperparameter Optimization as Google calls it is an important task in which the “best hyperparameters” are selected for a specific model. Every model requires a different set of hyperparameters. Some popular hyperparameters are learning rate, number of dense layers, dropout rate, and number of hidden units. Usually, the more layers and neurons in a deep learning model, the more knobs there are to tune.

In one specific experiment, the WB team created a simple image classification model using Keras. During the experimentation, the WB platform tracked all the various hyperparameters and configurations, which in turn, helped them measure model efficiency and accuracy. At the conclusion of the test, model accuracy reached 95%.

Source: W&B Classifying ASL Digits

Data scientist and blogger Jesus Rodriguez writes that while frameworks like Keras and TensorFlow “natively include hyperparameters optimization algorithms,” more advanced tooling is required for real-world models in order to track experiments and results. The DB platform is feature-rich, supporting AI frameworks like PyTorch and Tensorflow. Other features include support for Docker, GPU monitoring, and multi-GPU hyperparameter sweeps that test different configurations of a model. All it takes to monitor PyTorch models is five lines of code.

The startup has an impressive list of customers including OpenAI, GitHub, and Qualcomm. Also, the CEO of GitHub and the Chief Scientist for Salesforce are investors in Weights & Biases.


  • Company: Weights & Biases
  • HQ: San Francisco
  • Founded: 2017
  • Raised: $20M
  • # of Employees: 17
  • Founders: Lukas Biewald (CEO), Shawn Lewis (CTO), and Chris Van Pelt (CVP)
  • Product: Hyperparameter Optimization and ML Model Optimization
  • Customers: Toyota, GitHub, OpenAI, Qualcomm, and others
Scroll to Top