Next week, Microsoft and OpenAI plan to launch their latest in revolutionary language learning models (LLM). GPT-4, unlike it's predecessors, is multi-modal, meaning it can translate written text to not just text-based responses, but other mediums like pictures and videos.
As models become more advanced, the XAI implications become greater, as the world is starting to see in various litigation cases around the world. Multi-model models are inherently black box, with the only embedded traceability being the features it is trained upon. It is yet to be seen if GPT-4 will have an XAI layer to aid users in understanding the basis of the responses they receive - something that could greatly reduce mistakes, misunderstandings, and misuse.
With LLMs, there are several simple, yet impactful ways XAI can be implemented - but must be considered at the start of the model design and feature engineering phases of development. For LLMs, some XAI approaches are:
Source Citations - By including a feature importance XAI layer (such as LIME), a model can trace back to specific features that led to a specific response. Source citation can greatly reduce copyright/plagiarism risks.
Delayed Responses - While the rapid response feature of LLMs is exciting and a great feature for customers, delayed responses that take a slight longer amount of time to understand source material that drove the response can help caveat information for users. A delayed response could be in the form of a response report that includes XAI information like feature importance, source citations, and other key information.
Visualization - Visuals go a long way with humans and just seeing how a model is extracting and responding can aid in interpretability. Consider including a visual that shows real-time how a response is being formulated.
Documentation - Strong documentation, while seemingly a given, can be an afterthought in rapid development environments. Strong documentation that details, in plain speak, assumptions, guardrails, and controls can help give users understanding on the models decision making.
With the rapidly growing AI industry, XAI will continue to be needed. The XAI Foundation will continue updating with the latest models and techniques to encourage the AI industry and practitioners to use XAI.
Comentários