expressive power
The ability of a model or learning framework to represent a wide variety of functions and relationships in the data. In AI, higher expressive power often correlates with the model's capacity to learn complex patterns but may also lead to overfitting.
- A Hierarchy of Graphical Models for Counterfactual Inferences
- A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers
- Bridging Theory and Practice in Link Representation with Graph Neural Networks
- Compositional Reasoning with Transformers, RNNs, and Chain of Thought
- Constant Bit-size Transformers Are Turing Complete
- Discovering Symbolic Partial Differential Equation by Abductive Learning
- E2Former: An Efficient and Equivariant Transformer with Linear-Scaling Tensor Products
- EUGens: Efficient, Unified and General Dense Layers
- Fast attention mechanisms: a tale of parallelism
- FuncGenFoil: Airfoil Generation and Editing Model in Function Space
- Future Link Prediction Without Memory or Aggregation
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks
- On topological descriptors for graph products
- Quantum Doubly Stochastic Transformers
- Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks
- Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation
- The Computational Complexity of Counting Linear Regions in ReLU Neural Networks
- The Logical Expressiveness of Temporal GNNs via Two-Dimensional Product Logics
- TransMLA: Migrating GQA Models to MLA with Full DeepSeek Compatibility and Speedup