OmniXAI, which stands for Omni eXplainable AI, is a python-based machine learning library that provides deeper insights into explainable AI (XAI) models. It is an open-source framework by Salesforce, which offers a range of explanatory methods as ‘model-agnostic’ and ‘model-specific’ AI decisions.
Primitive AI models can get very complex for human understanding. As a consequence, a lot of crucial applications deter using them. The complexity of AI models based on deep neural networks has caused a surge in developing more XAI models. These models offer more transparency and persuasiveness to enhance model performance.
However, existing XAI libraries pose some restrictions. These libraries can handle only a limited type of data and models. Each XAI library has a different interface, making it difficult to switch from one to another. Furthermore, there is a lack of visualization and comparative explanation in the existing frameworks.
OmniXAI is designed as a one-stop comprehensive library that addresses these flaws. It makes XAI accessible for all machine learning processes; it is not restricted to feature engineering, model development, data exploration, or decision-making. With OmniXAI, users can deploy a ‘model-agnostic’ approach. With this approach, the framework can provide insights without any prior knowledge of the AI model. And in the ‘model-specific’ approach, it enables the framework to generate explanations with some prior knowledge of the model.
Salesforce’s open-source library, OmniXAI, is also compatible with PyTorch, TensorFlow, and others. Users can choose from several explanation methods for various elements of the model. OmniXAI facilitates this process by providing a standard interface that generates explanations with few lines of code.
The primary design philosophy of OmniXAI is to allow users to apply many explanation methods simultaneously and visualize the resulting generated explanations. Salesforce is still occupied with developing and improvising OmniXAI with more algorithms and compatibility with more data types.