The title is an homage to 'Do Androids dream of electric sheep?', but with the introduction of GPT-4V, interesting cases began to emerge. However, since the number of tokens used for image input and processing is enormous, it can only be used for high value-added services due to cost reasons, and the field that naturally came to mind was called radiology.
Simply put, it is a field that diagnoses by imaging and photographing parts or the whole body with various media such as CT, MRI, X-Ray, etc. It is a field that is often used in hard-to-reach places such as health checkups, orthopedics, and obstetrics and gynecology.
In some cases, GPT-4-generated radiology report summaries were preferable to report summaries written by trained radiologists.
Example
🧑⚕ : Cardiomegaly and mild interstitial pulmonary edema.
🤖: Stable cardiomegaly with prominent peripheral opacity that may indicate scarring or edema.
Additionally, according to a study published by Microsoft titled <Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine> , the general AI model GPT-4 was found to perform at the expert level in the medical field. In particular, it showed surprising results that surpassed models specialized for medical applications. The study reported that GPT-4 showed excellent performance in various medical problem-solving benchmarks, and especially showed superior results than existing models in problems requiring medical knowledge.
GPT-4 announced that it was able to achieve the best performance through a method called 'Medprompt' without any special fine-tuning, which is significant. This is the first time that GPT-4 achieved over 90% accuracy on the MedQA dataset, and reduced the misdiagnosis rate by 27% compared to Med-PaLM2 released by Google.
Many experts in the AI field believe that domain-specific fine-tuning is necessary to make general basic models perform well in a specific domain. However, fine-tuning can be expensive, requiring expert or professionally labeled datasets and computation for model parameter updates. This process is resource-intensive and expensive, making it a difficult challenge, especially for small and medium-sized organizations.
This study demonstrates the value of further exploring the possibilities of prompting in transforming general models into expert-level models. More interestingly, the proposed prompting methods have also been shown to be valuable in job competency tests across a variety of professional fields without the need for updating expertise.
In simple terms, it means that fine-tuning itself is expensive, but you can achieve decent performance with just the prompt settings. In other words, if you have a great basic model without fine-tuning, you can achieve expert-level performance in a specific domain just by setting the prompt well.
This paper and published research results show that general AI models such as GPT-4 can act as experts in specific fields. This opens up new possibilities for small and medium-sized businesses and organizations with limited resources to utilize advanced AI capabilities. The advancement of AI technology will continue, and it is expected to bring about innovative changes in various industries.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives. --- I post articles related to IT 💻, economy 💰, and humanities 🎭. If you are curious about my thoughts, perspectives or interests, please subscribe. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
