Scientists from the University of Science and Technology of China and Tencent's YouTu Lab have developed a tool to combat "hallucination" in artificial intelligence (AI) models.
Hallucination is the tendency of the AI model to produce high-confidence outputs that do not appear to be based on the information found in the training data. This problem is common in large language model (LLM) research. Their influence can be seen in models such as OpenAI's ChatGPT and Anthropic's Claude.
The USTC/Tencent team has developed a tool called “Woodpecker” that they claim can correct hallucinations in multimodal large language models (MLLM).
This subset of AI includes models such as GPT-4 (specifically its visual variant GPT-4V) and systems that include text-based language modelling, as well as vision and/or other processing operations.
The team notes that it used three separate AI models to perform hallucination correction, separate from the MLLM that corrects for hallucinations.
These include GPT-3.5 turbo, Grounding DINO and BLIP-2-FlanT5. Together, these models work as evaluators to detect hallucinations and give instructions for the corrected model to reproduce its output in accordance with its data.
To correct hallucinations, the AI models running “Woodpecker” use a five-step process that includes “key concept extraction, question formulation, visual information verification, visual claim generation, and hallucination correction.”
The researchers claim that these techniques provide additional transparency and offer a “30.66%/24.33% accuracy increase over baseline MiniGPT-4/mPLUG-Owl.” They evaluated many “off the shelf” MLLMs using their method and concluded that Woodpecker “can be easily integrated into other MLLMs.”
The evaluation version of the Woodpecker is available on Gradio Live, where enthusiasts can check out the vehicle in action. Join the discussion by sharing your thoughts in the comments section.
Hallucination is the tendency of the AI model to produce high-confidence outputs that do not appear to be based on the information found in the training data. This problem is common in large language model (LLM) research. Their influence can be seen in models such as OpenAI's ChatGPT and Anthropic's Claude.
The USTC/Tencent team has developed a tool called “Woodpecker” that they claim can correct hallucinations in multimodal large language models (MLLM).
This subset of AI includes models such as GPT-4 (specifically its visual variant GPT-4V) and systems that include text-based language modelling, as well as vision and/or other processing operations.
The team notes that it used three separate AI models to perform hallucination correction, separate from the MLLM that corrects for hallucinations.
These include GPT-3.5 turbo, Grounding DINO and BLIP-2-FlanT5. Together, these models work as evaluators to detect hallucinations and give instructions for the corrected model to reproduce its output in accordance with its data.
To correct hallucinations, the AI models running “Woodpecker” use a five-step process that includes “key concept extraction, question formulation, visual information verification, visual claim generation, and hallucination correction.”
The researchers claim that these techniques provide additional transparency and offer a “30.66%/24.33% accuracy increase over baseline MiniGPT-4/mPLUG-Owl.” They evaluated many “off the shelf” MLLMs using their method and concluded that Woodpecker “can be easily integrated into other MLLMs.”
The evaluation version of the Woodpecker is available on Gradio Live, where enthusiasts can check out the vehicle in action. Join the discussion by sharing your thoughts in the comments section.