AutoDL: Automated and Interpretable Deep Learning – Research Highlights from Baidu Big Data Lab

Deep learning is widely used in object recognition, medical imaging, autonomous driving, voice assistants, machine translation, and advertising, but faces challenges such as high computational cost, complex model configuration, and lack of interpretability.

The Baidu Big Data Lab focuses on two key problems: (1) end‑to‑end automated design of deep‑learning models (AutoDL) and (2) improving model interpretability and safety.

AutoDL Design aims to automatically search network architectures and hyper‑parameters. Existing techniques such as Bayesian optimization, multi‑armed bandits, RNN‑based reinforcement learning, and genetic algorithms have been applied, achieving strong results on CIFAR and ImageNet.

For visual style‑transfer, the lab introduced a neural architecture search (NAS) space of size 2^31, enabling real‑time video style transfer (e.g., styleNAS‑5opt processes 20 frames per second) and eliminating the need for masks.

The lab also built a pose‑generation network and a recommender‑system inference pipeline (JIZHI) that dynamically allocates resources, prunes long‑tail features, and optimizes latency, throughput, and instance count.

AutoDL Transfer introduces three versions of transfer learning: DELTA (feature‑map attention and selective distillation), a method to mitigate negative transfer by dropping gradient components that conflict with the loss, and RIFLE, which periodically re‑initialises fully‑connected layers to escape local minima.

These methods are exposed via PaddleHub APIs, allowing both script‑based one‑line fine‑tuning and programmatic use with less than 20 lines of code.

The EasyDL platform provides a no‑code solution for industrial AI, covering data preprocessing, automated model search, hyper‑parameter tuning, model deployment on cloud, private cloud, or edge devices, and includes a rich library of pre‑trained models for vision, NLP, and even bio‑informatics.

Interpretability tools such as NormLIME, PaddleX, and visual attribution maps enable users to understand which pixels or features drive predictions, helping to verify model correctness beyond raw accuracy.

Looking ahead, the team envisions AutoDL becoming the default pipeline for all machine‑learning tasks, turning model design into a combinatorial search problem that requires minimal human intervention.

Q&A session covered topics such as explainable knowledge graphs and acceleration techniques for recommender‑system inference pipelines.