The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Transformers for Tabular Data at Capital One with Bayan Bruss - #591
45 snips
Sep 12, 2022 Bayan Bruss, Senior Director of Applied ML Research at Capital One, explores the intricacies of applying deep learning to tabular data in the financial sector. He addresses the challenges faced, such as messy data and fraud detection, emphasizing the underappreciated significance of this domain. The discussion highlights the need for modern techniques like transformers and transfer learning, aiming to boost model performance and interpretability. Additionally, they delve into the potential of multimodal deep learning for enhancing predictive models.
AI Snips
Chapters
Transcript
Episode notes
Tabular Data & Deep Learning
- Deep learning innovation has focused on areas with low baselines, like computer vision.
- Tabular data's strong baselines (e.g., XGBoost) limit deep learning's perceived impact.
Bridging the Gap
- Advancements in deep learning for computer vision and NLP could apply to tabular data.
- This could bridge the gap and enable using deep learning tools for tabular data.
Counterfactual Explanations
- Counterfactual explanations help understand model decisions by identifying necessary input changes for different predictions.
- This deep learning technique, requiring differentiable models, can also be applied to tabular data.