TFM based Shared Intelligence Layer
summary
Claim: replacing permutation‑invariant aggregates with sequence‑aware Transaction Foundation Models (TFMs) and a shared intelligence layer gives big, measurable uplift. Implementation: custom tokenizer → NEMO‑trained 29M‑parameter TFM on ~24M IBM records (fraud rate 0.2%) → extract TFM embeddings; add TFM embeddings to raw XGBoost features; combine with bank‑wide graph embeddings from GraphSAGE. Results reported: raw XGBoost average precision ≈12%, raw+TFM ≈15%, TFM+GNN pushes to ≈66%, combined stack ≈72%. Uses NVIDIA stack (CUDA, RAPIDS, DOCA, MagnumIO) and includes a live demo.
AI-written summary · about
chapters (13)
- 0:00Shared intelligence in banking
- 4:43Siloed banking models challenges
- 6:14NVIDIA AI data center platform
- 10:29Regulated industries interpretability
- 12:18Revolut foundation model use cases
- 13:46Custom tokenizer for transaction data
- 15:13Transformer models for tabular data
- 18:02Training TFM on 24M records
- 21:16Extracting TFM embeddings
- 23:42Raw features with TFM embeddings
- 24:26Graph Neural Networks for transactions
- 28:44Results and future of TFM
- 30:29Q&A
related talks
01h 28mHands-On with Tabular Foundation Models
38mTensorFlow Extended (TFX): Machine Learning Pipelines and Model Understanding
13mAn introduction to MLOps with TensorFlow Extended (TFX)
33mTrillion-Token Pretraining
24mMulti-tasking fraud detection From Trees to MLPs
14mDoes your app use ML? Make it a product with TFX
18mBig‑T Notation: Engineering for Token Efficiency in the Age of Enterprise AI
28mEnhancing Fraud Detection