Neurosymbolic Learning and Geometric Reasoning Advances

The field of artificial intelligence is witnessing significant developments in neurosymbolic learning and geometric reasoning. Researchers are exploring new ways to combine deep learning with symbolic reasoning to achieve better data efficiency, interpretability, and generalizability. Recent works have focused on developing frameworks that can harness the power of GPUs to accelerate neurosymbolic learning, enabling the solution of complex problems that were previously infeasible. Additionally, there is a growing interest in geometric reasoning, with studies demonstrating the ability of graph neural networks and transformers to learn and reason about geometric constraints. Noteworthy papers include: Lobster, a unified framework for neurosymbolic learning that achieves an average speedup of 5.3x over state-of-the-art frameworks. CTSketch, a novel algorithm for scalable neurosymbolic learning that pushes the boundaries of what is possible with current hardware. GEOPARD, a transformer-based architecture for predicting articulation from a single static snapshot of a 3D shape, which yields state-of-the-art results in articulation inference. These advancements have the potential to revolutionize various fields, from computer vision to natural language processing, and are expected to have a significant impact on the development of more sophisticated AI systems.

Sources

Lobster: A GPU-Accelerated Framework for Neurosymbolic Programming

Concise One-Layer Transformers Can Do Function Evaluation (Sometimes)

Approximation results on neural network operators of convolution type

Integer multiplication is at least as hard as matrix transposition

Tape Diagrams for Monoidal Monads

CTSketch: Compositional Tensor Sketching for Scalable Neurosymbolic Learning

$\mathsf{P}$-completeness of Graph Local Complementation

Neural Approaches to SAT Solving: Design Choices and Interpretability

Geometric Reasoning in the Embedding Space

On shallow feedforward neural networks with inputs from a topological space

Investigating Simple Drawings of $K_n$ using SAT

SymDQN: Symbolic Knowledge and Reasoning in Neural Network-based Reinforcement Learning

GEOPARD: Geometric Pretraining for Articulation Prediction in 3D Shapes

Spline-based Transformers

Built with on top of