Using Logic Programming and Kernel-Grouping for Improving Interpretability of Co...
summary
NeSyFOLD-G converts last-layer CNN kernels into a compact logic program: it clusters similar filters by cosine similarity (kernel-grouping), binarizes each group’s activations, then feeds the binarization table to FOLD-SE-M to induce defaults-and-exceptions rules. Semantic concepts are assigned by intersecting receptive fields with segmentation masks and averaging IOU scores (semantic labeling). Across Places, GTSRB and MNIST the grouped rules keep comparable accuracy and fidelity while producing a markedly smaller rule-set; queryable justifications come from s(CASP).
AI-written summary · about
chapters (10)
- 0:00Research problem
- 2:31Convolutional neural networks
- 4:40FOLD and rule-based learning
- 6:48NessyFold framework
- 11:50Kernel grouping motivation
- 14:20Semantic labeling of kernels
- 20:52Experimental setup
- 22:20Results and metrics
- 25:12Conclusion and future work
- 26:15Q&A
related talks
26mStructured Predictions an historical overview
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
36mEnhanced and Efficient Reasoning in Large Language Models
18mMaking Logic a First-Class Citizen in Generative ML for Networking
27mPattern Analysis: Learned vs. Engineered
18mActive Learning for Neurosymbolic Program Synthesis
02h 42m[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)