inductive biases
Inductive biases refer to the set of assumptions a learning algorithm makes to generalize from the training data to unseen data. In AI, these biases greatly influence how models learn and adapt to new information, determining how effectively they can generalize across different tasks or domains.
- A compressive-expressive communication framework for compositional representations
- Attention on the Sphere
- Brain-Like Processing Pathways Form in Models With Heterogeneous Experts
- Connecting Neural Models Latent Geometries with Relative Geodesic Representations
- Deep Continuous-Time State-Space Models for Marked Event Sequences
- Disentangled Representation Learning via Modular Compositional Bias
- DualFocus: Depth from Focus with Spatio-Focal Dual Variational Constraints
- E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization
- Equi-mRNA: Protein Translation Equivariant Encoding for mRNA Language Models
- Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings
- Hybrid Latent Representations for PDE Emulation
- Improving Time Series Forecasting via Instance-aware Post-hoc Revision
- Infinite Neural Operators: Gaussian processes on functions
- Is Your Diffusion Model Actually Denoising?
- Learning normalized image densities via dual score matching
- Meta Guidance: Incorporating Inductive Biases into Deep Time Series Imputers
- Neural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation
- Object-centric binding in Contrastive Language-Image Pretraining
- OmniFC: Rethinking Federated Clustering via Lossless and Secure Distance Reconstruction
- On Inductive Biases That Enable Generalization in Diffusion Transformers
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity
- Partner Modelling Emerges in Recurrent Agents (But Only When It Matters)
- Probing Neural Combinatorial Optimization Models
- Scalable Evaluation and Neural Models for Compositional Generalization
- Stable Port-Hamiltonian Neural Networks
- Task-Optimized Convolutional Recurrent Networks Align with Tactile Processing in the Rodent Brain
- Task-Optimized Convolutional Recurrent Networks Align with Tactile Processing in the Rodent Brain
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment
- Transformers for Mixed-type Event Sequences
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization
- Where Graph Meets Heterogeneity: Multi-View Collaborative Graph Experts
- Which Algorithms Have Tight Generalization Bounds?
- Why Masking Diffusion Works: Condition on the Jump Schedule for Improved Discrete Diffusion