New CRISP framework enhances interpretability of deep neural networks

Researchers have introduced CRISP, a novel framework designed to enhance the interpretability of deep neural networks. CRISP reconstructs the final-layer activation vectors of binary neural networks into Tsetlin Machine clauses, providing a symbolic trace from input features to hidden neurons. This method was evaluated on several benchmark datasets, including MNIST, KMNIST, Fashion-MNIST, SVHN, and CIFAR10, using different neural network architectures. The findings indicate that while CRISP's reconstruction fidelity is a limiting factor, it maintains significant teacher-head accuracy and offers a clause-level pathway for inspecting binary neural network representations. AI

IMPACT Provides a new method for inspecting the internal workings of binary neural networks, potentially aiding in debugging and understanding model behavior.

RANK_REASON The cluster contains a research paper detailing a new framework for neuro-symbolic propositions. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• BCCNN
• BinaryConnect: Training Deep Neural Networks with binary weights during propagations
• BNN
• CIFAR-10
• CRISP
• Fashion-MNIST
• KMNIST
• MNIST database
• The Street View House Numbers Dataset
• Tsetlin Machine
• Universiti Teknologi MARA

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →