There has been a recent burst in “AutoML” techniques as a means to automate the creation of ML models without necessary domain expertise. This demonstration looks well beyond AutoML’s current narrow focus on automated model building, to tackling automation across the full end-to-end AI/ML lifecycle. In industry settings, the AI/ML lifecycle typically includes a series of labor-intensive tasks such as preparing data, training models, deploying the selected model in cloud, monitoring performance, identifying faults, and taking corrective actions when failures or new business requirements occur. Enormous opportunities exist for scaling, automating, and accelerating this AI/ML lifecycle.
In this session, we demonstrate tools and research results in driving automation across the entire AI/ML lifecycle: from assessing data readiness and recommending mitigations, to semantically-driven automation based on concept discovery and knowledge augmentation, to advanced ML model building with business and fairness constraints, to novel pipelines for industry-critical modalities, to automation for monitoring models in deployment, recognizing deficiencies and recommending corrective actions. We will also demonstrate practical methods for scaling in multi-cloud environments with federated learning, and accelerated cloud-based inference of widely-popular classical ML algorithms such as XGBoost and LightGBM.