December 2022 Summaries

Labelbox has introduced several updates to its suite of tools, including Labelbox Catalog, Annotate, and Model, aimed at enhancing AI teams' capabilities in data labeling and model training. Key features include improved similarity search powered by vector embeddings, which allows users to efficiently identify specific data points within vast datasets, thus refining model performance by addressing edge cases and rare examples. The platform now supports native PDF and text document annotation, utilizing multimodal annotation to extract complex data from documents, which is particularly beneficial in industries like healthcare and finance. To optimize pre-labeling workflows, Labelbox has introduced an automation efficiency score that quantifies the impact of pre-labeling on time and cost savings, enabling teams to better measure and enhance their processes. Additionally, the platform offers tools for curating and versioning hyperparameters and datasets, facilitating model comparison and iteration. These advancements collectively aim to streamline AI development by providing more robust, efficient, and insightful data handling and model training processes.

December saw the introduction of new features and improvements in Labelbox aimed at enhancing the search and organization of visual and natural language data. Key updates include the ability to preview HTML data rows for efficient data selection, conduct search queries on plain text, and utilize AI-powered natural language search on both text and images to quickly locate relevant data. These enhancements facilitate the identification and prioritization of data for labeling, thereby improving model performance. The updates are being gradually rolled out across Labelbox's Catalog, Annotate, and Model products, with an option for early access via user ID submission. Additionally, the HTML editor now supports detailed previews, while display settings have been improved to allow viewing of data row IDs like global key or external ID for precise data selection. The latest Python SDK version introduces a method to clear global keys and support for creating large data batches, complemented by new guides and tutorials on data curation and active learning.

Labelbox Model's recent developments aim to enhance the speed and quality of machine learning model deployment by automating the process of identifying model failures and labeling errors. The platform now offers auto-generated metrics such as precision, recall, and confusion matrices, allowing users to evaluate model performance and make necessary adjustments efficiently. Users can upload model predictions and ground truths to access these metrics, as well as upload custom metrics if needed. The updated features include an interactive NxN confusion matrix and histograms that help pinpoint areas of model underperformance or labeling discrepancies. Additionally, the embedding projector tool aids in error analysis by visualizing data patterns and outliers, supporting up to 50,000 data points for in-depth analysis. The platform's new capabilities in visualizing segmentation masks and the ability to adjust confidence and IOU thresholds further streamline the debugging process, enabling faster identification and resolution of model and data issues.

Labelbox users now have the capability to upload and annotate documents and conversational text to train language models tailored to specific business needs, with enhancements made to the video editor and data review processes. The platform supports native upload and annotation of PDFs for tasks like Named Entity Recognition (NER) and Optical Character Recognition (OCR), which are crucial in industries such as financial services and healthcare. Users can annotate conversational texts to understand user intent and sentiment, leveraging natural language processing advancements. Labelbox also introduced a cuboid tool to capture 3D space in images, enhancing the ability to annotate dimensions and orientations. The video editor now features improvements like click-and-drag classification and beta bounding box tracking to streamline video labeling. Additionally, a new batch-based queueing system has been implemented to optimize data labeling and review workflows, enabling users to manage data rows more effectively. This rollout, initially available to Free, Education, and Starter users, began reaching Pro and Enterprise customers in December 2023, with a migration path for older projects planned for Q1 2023.