Exactness at Inference: A Representational Criterion for Out-of-Distribution Generalization
The paper argues OOD generalization requires exact representational equivalence to the data-generating mechanism.
The authors frame extrapolation as an inference-time property, not a training-time fit. Tensor Logic at zero temperature is presented as a passing case because it is equivalent to discrete logic while remaining continuous in arithmetic. They contrast that with thresholded neural networks and Logic Tensor Networks, which they say fail the criterion. The paper also ties failed extrapolation and inability to bind novel entities to the same lack of exact representability. ArXiv · AI/CL/LG's note
The authors frame extrapolation as an inference-time property, not a training-time fit. Tensor Logic at zero temperature is presented as a passing case because it is equivalent to discrete logic while remaining continuous in arithmetic. They contrast that with thresholded neural networks and Logic Tensor Networks, which they say fail the criterion. The paper also ties failed extrapolation and inability to bind novel entities to the same lack of exact representability. ArXiv · AI/CL/LG's note
score 4