Hasty Briefsbeta

Bilingual

TabFM: A zero-shot foundation model for tabular data

2 days ago
  • #Tabular Data
  • #Zero-Shot Learning
  • #Foundation Model
  • Google introduces TabFM, a foundation model for zero-shot classification and regression on tabular data, eliminating manual training and feature engineering.
  • TabFM uses in-context learning (ICL) to process entire datasets as prompts, learning relationships at inference time without weight updates.
  • The model employs a hybrid architecture combining strengths from TabPFN and TabICL to handle tabular data's two-dimensional, orderless structure.
  • Due to scarcity of real-world tabular datasets, TabFM is pre-trained on hundreds of millions of synthetic datasets generated with structural causal models.
  • Benchmarks on TabArena show TabFM outperforms traditional supervised methods like XGBoost and specialized models, with two configurations (1B and 3.1B parameters).
  • TabFM will be integrated into Google BigQuery, allowing users to make predictions via a simple SQL command without ML expertise.