Detecting Out-of-Distribution Text Using Topological Features of Transformer-Based Language Models

🎉 Exciting News! 🎉

I'm thrilled to share that our paper has been accepted for presentation at IJCAI 2024 in the AI Safety workshop!

📄 Paper Title: Detecting Out-of-Distribution Text Using Topological Features of Transformer-Based Language Models

🔍 What’s It About?

Our research focuses on improving how AI models, like BERT, detect when they encounter text that is very different from what they were trained on (known as out-of-distribution or OOD text). This is important for making AI systems safer and more reliable.

🔧 How Did We Do It?

We used a technique called Topological Data Analysis (TDA) to analyze the attention patterns within BERT, a popular AI language model. Think of TDA as a way to look at the "shape" of data to find patterns.

🆚 What Did We Compare?

We compared our TDA approach with a more traditional method that looks at BERT's summary of text (CLS embeddings).

💡 Key Findings:

• Our TDA approach did a great job distinguishing between in-domain texts (like HuffPost news articles) and far out-of-domain texts (like IMDB movie reviews).

• However, it was less effective at distinguishing between in-domain texts and near out-of-domain texts (like CNN/DailyMail news articles) or similar-domain texts (like business news articles from HuffPost).

This research could help improve AI's ability to recognize when it's facing unfamiliar or unexpected text, making AI applications more robust and trustworthy.

Looking forward to sharing more insights at IJCAI 2024!

You can follow the slide to get more out of the paper. I gave this talk at Sikkim university.