Simplifying Deep Learning: Research Overview by Prof. Yao Quanming

Prof. Yao Quanming, an assistant professor at Tsinghua University, shares his research agenda titled “The Minimalist Path of Deep Learning,” highlighting his background, awards, and contributions to automated machine learning and drug design.

He identifies three research directions: (1) improving model adaptability through automated architecture design, (2) enhancing learning efficiency via few‑shot learning, and (3) reducing drug‑design costs by modeling structured data.

The talk is organized into four parts: (1) the need for a minimalist learning approach, (2) future visions for scientific‑intelligence scenarios, (3) a minimal closed‑loop case study in drug‑interaction prediction, and (4) practical attempts aligned with Ant Group’s needs.

He explains that current large‑model scaling laws rely on ever‑greater data, compute, and parameters, leading to three bottlenecks—data scarcity, compute limits, and trust issues—making further performance gains increasingly costly.

To address these challenges, he advocates three forms of simplification: model simplification, training‑process simplification, and interpretability simplification, arguing that a simpler model with comparable performance can reduce data consumption and improve explainability.

In the drug‑interaction domain, his team built a closed‑loop pipeline that predicts adverse interactions using graph neural networks (EmerGNN), achieving superior results to large language models while using far fewer parameters (≈2×10⁵ vs. ≈8×10⁷) and offering better interpretability through attention‑weighted subgraph analysis.

Validation is performed by cross‑checking high‑attention links against existing literature, achieving 60‑70% coverage, and future work includes metabolic‑network validation and domain‑shift handling.

Additional topics include multi‑agent coordination, privacy preservation in large models, and leveraging knowledge‑space privacy to mitigate hallucinations, illustrating broader implications of minimalist AI across recommendation, finance, and biomedical applications.

The presentation concludes with a call for continued exploration of symbolic‑based learning primitives to break scaling‑law constraints and advance practical AI solutions.