Intelligent Titanium TI-ONE: Tencent Cloud's One-Stop Machine Learning IDE

At the Tencent "Cloud + Future" Summit in May 2018, Dr. Wang Caihua, Technical Lead of Tencent Cloud AI Platform, presented TI-ONE (Intelligent Titanium), a one-stop machine learning platform on Tencent Cloud. The name "Titanium" was chosen for its lightweight and high-strength properties, similar to Iron Man's suit, reflecting the platform's powerful yet efficient characteristics.

Why TI-ONE is Needed:

While AI has become increasingly important (as Andrew Ng described it as "the new electricity"), machine learning algorithms have a lengthy lifecycle: from data acquisition and preprocessing to framework selection, algorithm development, model training, prediction, and finally model serving on the cloud. TI-ONE was created to accelerate this lifecycle and expedite model serving.

Key Features of TI-ONE:

Integrated data preprocessing platform to improve efficiency

Support for mainstream ML frameworks with built-in common algorithms; drag-and-drop interface for algorithm development

Automatic hyperparameter tuning with multi-level collaboration support

One-click model deployment and serving with online inference

Graphical development interface serving as an IDE for machine learning on Tencent Cloud

Architecture:

TI-ONE has a layered architecture: COS storage layer at the bottom, GaiaStack resource scheduling layer above (providing commercial features), followed by the framework layer integrating TensorFlow, PyTorch, XGBoost, Angel (Tencent's self-developed framework), and enhanced Spark. The platform supports both deep learning algorithms (CNN, RNN, DBN) and traditional ML algorithms (GBDT, FFM) for applications like image recognition, speech recognition, recommendation systems, and real-time risk control.

Workflow Example:

Users can develop ML algorithms through: 1) obtaining data from COS or local filesystem, 2) data preprocessing, 3) splitting data into training and validation sets, 4) selecting algorithms via drag-and-drop (e.g., logistic regression), 5) setting parameters, and 6) training to obtain a model. After running, the platform provides confusion matrix and AUC values for evaluation.

Automatic Hyperparameter Tuning:

TI-ONE generates multiple parameter combinations and runs them in parallel, selecting the best result. For example, when training a random forest, users can specify the number of trees and features per tree, and TI-ONE will find the optimal combination.

Collaboration Features:

TI-ONE supports collaboration at three levels: 1) Model-level sharing - share trained models with colleagues for comparison, 2) Workflow-level sharing - share the entire ML lifecycle for reuse with minor modifications, 3) Service-level sharing - share deployed models with backend teams for problem diagnosis.

Deployment and Serving:

TI-ONE provides one-click deployment tools to convert trained models into Applications, load multiple instances with version support, and enable third-party users to call via REST API.

Special Differentiating Features:

1. Angel Framework: Tencent's self-developed ML framework that overcomes Spark's limitation of storing models on single nodes. Through optimization of underlying mathematical libraries, Angel can support trillion-parameter models. Performance comparisons show Angel is 20+ times faster than Spark for some algorithms.

2. Graph Computing Algorithms: Based on GraphX, Tencent added numerous graph algorithms including node evaluation, community discovery, and statistical feature algorithms, all optimized to support billion-scale relationship chains.

3. Custom Algorithm Support: Advanced users can define and upload their own algorithms to TI-ONE for execution, providing great flexibility.

Commercial Features (via GaiaStack):

TI-ONE offers dedicated clusters with multi-tenant support, resource and data isolation, hot upgrades without service interruption, automatic master-standby failover, and automatic load balancing when traffic increases.

Users:

TI-ONE is used internally by Tencent Game, WeChat, MyApp, QQ Music, and other business units.