GPT-4 with vision, known as GPT-4V, empowers users to instruct the model to analyse images provided by the user. This integration of image analysis into large language models (LLMs) represents a significant advancement that is now being made widely accessible. The inclusion of additional modalities, such as image inputs, into LLMs is considered by some as a crucial frontier in the field of artificial intelligence research and development, as highlighted in various sources. Multimodal LLMs hold the potential to expand the capabilities of language-focused systems by introducing novel interfaces and functionalities. This, in turn, is now allowing them to address new tasks and offer unique experiences to their users.
GPT-4V, similar to GPT-4, completed its training in 2022, with early access becoming available in March 2023. The training process for GPT-4V was akin to that of GPT-4, involving initial training to predict the next word in text using a large dataset of text and image data from the internet and licensed sources. Subsequently, reinforcement learning from human feedback (RLHF) was used to fine-tune the model, ensuring its outputs align with human preferences.
Large multimodal models like GPT-4V combine both text and vision capabilities, which introduces unique limitations and risks. GPT-4V inherits the strengths and weaknesses of each modality while also presenting new capabilities resulting from the fusion of text and vision, as well as the intelligence derived from its large scale. To gain a comprehensive understanding of the GPT-4V system, a combination of qualitative and quantitative evaluations were employed. Qualitative assessments involved internal experimentation to rigorously assess the system’s capabilities, and external expert red-teaming was sought to provide valuable insights from external perspectives.
This system card provides insights into how OpenAI prepared GPT-4V’s vision capabilities for deployment. It covers the early access period for small-scale users, safety measures learned during this phase, evaluations to assess the model’s readiness for deployment, feedback from expert red team reviewers, and the precautions taken by OpenAI before the model’s broader release.
The above image demonstrates examples of GPT-4V’s unreliable performance for medical purposes. The capabilities of GPT-4V present both exciting prospects and new challenges. The approach taken in preparing for its deployment has focused on evaluating and addressing risks associated with images of individuals, which include concerns like person identification and the potential for biased outputs from such images, leading to representational or allocational harms.
Furthermore, the model’s significant leaps in capabilities within high-risk domains, such as medicine and scientific proficiency, have been thoroughly examined. There are multiple fronts, where researchers As we move forward, it is essential to continue refining and expanding the capabilities of GPT-4V, paving the way for even more remarkable advancements in the realm of AI-driven multimodal systems!
Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 30k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.
Janhavi Lande, is an Engineering Physics graduate from IIT Guwahati, class of 2023. She is an upcoming data scientist and has been working in the world of ml/ai research for the past two years. She is most fascinated by this ever changing world and its constant demand of humans to keep up with it. In her pastime she enjoys traveling, reading and writing poems.