Hands-On with Tabular Foundation Models
summary
Hands-on notebooks run TabICL-style models and compare them rigorously to XGBoost and Random Forest using sklearn, tabicl, xgboost, pandas and an API option; experiments use a small medical set (~500 rows), a medium e‑commerce set (~5K rows) and a large noisy set (~50K rows). Claims: TFMs outperform trees on small tables, compete on medium, and lag or cost more on large tables; the session shows benchmarks, SHAP‑style diagnostics, practical tokenization/latency tradeoffs and working code that you can run, tweak, and break.
AI-written summary · about
chapters (12)
- 0:00Tabular foundation models intro
- 1:51Why TFMs emerged
- 4:18How TFMs work
- 10:38Hands-on with tabular foundation models
- 28:32First TFM prediction
- 39:01TFMs vs gradient boosting
- 50:02SHAP for TFM predictions
- 1:01:29Feature importance: TFMs vs trees
- 1:12:55Fine-tuning and scaling TFMs
- 1:19:41TFM benchmark results
- 1:24:15TFMs: limitations and future
- 1:28:00Q&A
related talks
29mThe Foundation Model Revolution for Tabular Data
30mTabular data Transformed? TabPFN Brings Deep Learning to the Table
31mTFM based Shared Intelligence Layer
37mCreating correct and capable classifiers
57mGood Practices for Applied Machine Learning
31mMove beyond academia: Introducing an industry-first tabular benchmark
34mRethinking Feature Importance-Evaluating SHAP & TreeSHAP for Tree-Based MLM
25mHumans helping machines help humans run machines