PulseAugur
EN
LIVE 16:12:49

Compositionality Emerges in Neural Networks Within Specific Depth-Connectivity Sweet Spot

Researchers have identified that compositionality in neural networks emerges within a specific, narrow range of depth and connectivity. This phenomenon is highly dependent on the sparsity of connections rather than just weight sparsity. When these conditions are not met, gradient descent leads to fractured solutions instead of compositional ones. To address this, the study introduces similarity-based pruning (SP) for compositional connectivity and a depth predictor to identify optimal depths for compositionality, supported by a theoretical framework. AI

IMPACT Understanding the conditions for compositionality could lead to more generalizable and robust AI models.

RANK_REASON The cluster contains a research paper detailing findings on neural network compositionality.

Read on arXiv cs.LG →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

Compositionality Emerges in Neural Networks Within Specific Depth-Connectivity Sweet Spot

COVERAGE [2]

  1. arXiv cs.LG TIER_1 English(EN) · Dat H. Do, Rushi Shah, Duc V. Le, Dianbo Liu ·

    Compositionality Emerges in a Narrow Depth-Connectivity Regime: Architecture Constraints and Solution Manifolds

    arXiv:2606.19941v1 Announce Type: new Abstract: Compositionality is believed to be the foundation for generalization, enabling models to reuse meaningful primitives in novel combinations. Yet, models trained with standard gradient-based optimization rarely, and often only weakly,…

  2. arXiv cs.LG TIER_1 English(EN) · Dianbo Liu ·

    Compositionality Emerges in a Narrow Depth-Connectivity Regime: Architecture Constraints and Solution Manifolds

    Compositionality is believed to be the foundation for generalization, enabling models to reuse meaningful primitives in novel combinations. Yet, models trained with standard gradient-based optimization rarely, and often only weakly, exhibit compositional internal structure, and i…