June 2023 Summaries

Model Foundry, a new solution from Labelbox, is set to enhance AI development by allowing machine learning teams to efficiently leverage and compare foundation models through A/B testing and comprehensive model comparison processes. With the rise of off-the-shelf and foundation models, AI teams can streamline their workflows by selecting suitable pre-trained models for specific tasks, speeding up development and reducing costs. Model Foundry facilitates this by providing a platform where users can test, evaluate, and experiment with a range of models, including those for computer vision and natural language processing, in a no-code environment. By offering detailed performance metrics and a model comparison view, the platform helps teams assess model effectiveness, track ML experiments, and maintain an accessible store of records for future reference. This ensures that AI projects align with business goals and that the most efficient models are utilized. The platform's features also include auto-generated metrics, custom metrics, and a streamlined A/B testing framework to optimize model performance. Through a practical example comparing GPT-4 and Claude, the blog illustrates how Model Foundry can aid in evaluating models' predictive accuracy and generative capabilities on specific datasets, thereby demonstrating the platform's utility in real-world scenarios.

Labelbox's latest advancements, particularly the introduction of Model Foundry and improvements to its image editor, are designed to enhance efficiency and accuracy in computer vision tasks, especially image segmentation. These updates leverage cutting-edge foundation models like Meta's Segment Anything to automate and accelerate the traditionally labor-intensive process of labeling, reducing both time and costs. The Model Foundry allows users to access various models, including SAM and YOLOv8, enabling pre-labeling of data and streamlining the labeling workflow. The integration of features like AutoSegment 2.0 and the brush tool further refines this process by offering AI-assisted segmentation with precise control over pixel-level details, allowing labelers to focus on reviewing and correcting AI-generated labels. Together, these innovations aim to minimize human errors, improve model performance, and facilitate collaboration between AI models and human reviewers, ultimately transforming how data labeling is conducted in the field of computer vision.

Karen Yang's blog post explores the use of Labelbox's Model Foundry to leverage large language models (LLMs) like ChatGPT for automating product categorization and summarization. The post highlights the labor-intensive nature of traditional product tagging and categorization and demonstrates how LLMs can streamline this process by instantly classifying products based on their descriptions. Using an Etsy dataset, the experiment showcases how LLMs can accurately predict product categories and generate concise summaries from long-form descriptions, thus reducing manual effort and allowing experts to focus on model evaluation. The blog provides insights into setting up the model with Labelbox, creating effective prompts, and evaluating the model's performance through both qualitative and quantitative analyses. It emphasizes the model's ability to perform well in a zero-shot setting, while also noting areas for improvement due to category ambiguity. The process underlines the importance of a quick iteration cycle for refining model prompts and configurations, ultimately enhancing the automation workflow for product categorization tasks.

Labelbox has announced early access to its Foundry add-on for Labelbox Model, which integrates foundational AI models, such as GPT-4, into its platform to enhance data labeling, enrichment, and model development. These foundation models, which excel across various data modalities, promise to accelerate labeling tasks by over 88%, reducing human effort significantly, and enabling tasks that previously took days to be completed in hours. By leveraging these models, Labelbox aims to outperform traditional and programmatic labeling methods, providing a more efficient and cost-effective solution. Despite their capabilities, foundation models still require human oversight for quality assurance. Labelbox's approach integrates cutting-edge technologies into its platform, including similarity search and zero-shot classification, to optimize every aspect of data-centric AI development. The Foundry add-on allows access to top AI models from major providers, acting as a copilot for experts and offering tools to test and evaluate models across different prompts and parameters. This development marks a significant evolution in AI systems, promising enhanced performance across enterprise applications.