TabFM: A zero-shot foundation model for tabular data (research.google)

Google Research has announced TabFM, a zero-shot foundation model specifically designed for tabular data integration within BigQuery ML, aimed at streamlining classification and regression tasks. Historically, deploying traditional supervised learning algorithms like XGBoost has been labor-intensive, requiring extensive manual feature engineering and hyperparameter optimization. TabFM addresses these challenges by leveraging in-context learning (ICL), allowing users to generate predictions on new datasets without the need for traditional model training, thus removing significant bottlenecks in machine learning workflows. Significantly, TabFM is trained using synthetic datasets generated by structural causal models, which accurately reflect complex feature relationships typically found in real-world tables. This model effectively interprets the two-dimensional nature of tabular data, enabling scalable zero-shot predictions, outperforming conventional algorithms in terms of ease of use and performance. With its integration into Google BigQuery, users can execute advanced predictions through a simple SQL command, democratizing access to advanced machine learning capabilities and empowering data scientists to generate high-quality insights rapidly and efficiently.