TLDR: Obz AI is a new software ecosystem that integrates explainable AI (XAI) and observability into computer vision (CV) systems. It provides a Python library for applying XAI, detecting outliers, and logging model activities, alongside a full-stack analytics dashboard for real-time monitoring and analysis. This helps make “black box” deep learning models more transparent and trustworthy for practical deployments.
Obz AI is a new software system designed to make deep learning models used in computer vision (CV) more understandable and easier to monitor. Deep learning models, like those used for image classification or medical imaging, are often called “black boxes” because it’s hard to see how they make their decisions. This lack of transparency can be a problem, especially in critical areas like healthcare or autonomous driving.
To solve this, Obz AI provides a complete solution that combines explainable AI (XAI) techniques with tools for managing and monitoring these models. It helps machine learning engineers understand why a model makes a certain prediction, detect unusual data or model behaviors, and continuously keep an eye on AI models in real-time. By making the decision-making mechanisms of deep models interpretable, Obz AI promotes observability and responsible deployment of computer vision systems.
How Obz AI Works
The system is made up of several key parts. First, there’s a Python library that works directly with a user’s AI model. This library has three main modules:
-
XAI: This module includes various methods to explain how a model arrived at its decision, producing “heatmaps” that show which parts of an image were most important for a prediction.
-
Data Inspector: This helps analyze the input data to spot anomalies or outliers. For example, it can extract features from images and use them to identify unusual patterns.
-
Obz Client: This part collects and saves all the information about the AI model’s activities, including predictions, explanations, and features.
Beyond the Python library, Obz AI also has a robust data management system that stores both structured information (like logs and project details) and large unstructured files (like images and explanation maps). A backend system, built with FastAPI, manages all interactions, ensuring data is stored correctly and securely.
Finally, there’s a user-friendly web dashboard. This dashboard is where developers and project managers can see high-level summaries, inspect data and model behaviors, and manage their projects. It’s specifically designed for computer vision, allowing users to trace an individual log from the input image to its extracted features, outlier status, and generated explanations. Users can also view XAI visualizations, where an original image and its explanation heatmap are shown side-by-side, highlighting important regions for a prediction.
Also Read:
- Enhancing Cybersecurity with Transparent AI: Introducing the L-XAIDS Framework
- Fairer Explanations for AI: Introducing MC3G for Actionable Recourse
Real-World Applications
Obz AI has been demonstrated with two types of applications. One involves natural images from the ImageNet dataset, where it uses advanced techniques to detect outliers and explain predictions for a vision transformer model. The other application uses medical images from the LIDC-IDRI dataset, focusing on lung nodule malignancy prediction, where it helps monitor features and identify unusual scans.
In essence, Obz AI aims to make computer vision systems more transparent and trustworthy by integrating explainability and observability into the entire machine learning pipeline. You can find more details about this work in the research paper.
