AAAI‑2024 Highlights: Alibaba Cloud’s Deep Tabular Learning & Multi‑Modal Fusion

Recent papers from Alibaba Cloud’s AI platform PAI were accepted at AAAI‑2024, one of the most prestigious international conferences in artificial intelligence, highlighting the platform’s advances in fundamental and applied AI research.

Unlocking Deep Tabular Learning (AMFormer)

The authors identify arithmetic feature interaction as a crucial inductive bias for deep tabular learning. By embedding this bias into a Transformer‑based architecture called AMFormer, they achieve superior modeling accuracy, data‑efficiency, and generalization on both synthetic and real‑world tabular datasets.

MuLTI: Efficient Video‑Language Understanding

MuLTI addresses the high computational cost of multimodal video‑language models by introducing a Text‑Guided MultiWay‑Sampler and a multiple‑choice modeling pre‑training task. These innovations reduce GPU memory usage while preserving performance, achieving state‑of‑the‑art results on several video‑question‑answering and retrieval benchmarks.

M2SD: Multiple Mixing Self‑Distillation for Few‑Shot Class‑Incremental Learning

M2SD proposes a dual‑branch architecture with virtual classes and a multiple‑mixing self‑distillation strategy to expand the feature space for new categories while preserving knowledge of old ones. Extensive experiments on few‑shot class‑incremental benchmarks demonstrate significant improvements in accuracy and robustness.

M2Doc: Plug‑in Multi‑Modal Fusion for Document Layout Analysis

M2Doc introduces early‑fusion and late‑fusion modules that combine visual and textual features within existing object detectors. This plug‑in design yields consistent performance gains on document layout benchmarks such as DocLayNet and M6Doc, achieving state‑of‑the‑art results when combined with detectors like DINO.