NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions

Authors: Tue Cao, Nhat Hoang-Xuan, Hieu Pham, Phi Le Nguyen, My Thai

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive empirical studies validate the fidelity of our proposed Neur Flow. Additionally, we showcase its utility in practical applications such as image debugging and automatic concept labeling.
Researcher Affiliation Academia 1 Institute for AI Innovation and Societal Impact (AI4LIFE), Hanoi University of Science and Technology, Hanoi, Vietnam 2 University of Florida, Gainesville, Florida, USA 3 College of Engineering & Computer Science, Vin University, Hanoi, Vietnam
Pseudocode Yes E DETAILED ALGORITHMS E.1 IDENTIFYING CORE CONCEPT NEURONS AND CONSTRUCTING NEURON CIRCUIT Algorithms 1, 2, and 3 provide detailed pseudocode for identifying core concept neurons, determining the semantic groups, and constructing the neuron circuit respectively. Algorithm 1 Identifying core concept neurons Algorithm 2 Determining semantic groups Algorithm 3 Forming neuron circuit
Open Source Code Yes 1Source code: https://github.com/tue147/neurflow
Open Datasets Yes Our experiments are performed on Res Net50 (He et al., 2016) and Goog Le Net Szegedy et al. (2015) using the ILSVRC2012 validation set (Russakovsky et al., 2015).
Dataset Splits Yes Our experiments are performed on Res Net50 (He et al., 2016) and Goog Le Net Szegedy et al. (2015) using the ILSVRC2012 validation set (Russakovsky et al., 2015).
Hardware Specification No Explanation: The paper mentions
Software Dependencies No Explanation: The paper mentions
Experiment Setup Yes Unless otherwise specified, the input parameters are τ = 16, N = 50, and k = 50, where the top 50 images with the highest activation on the target neuron are considered as its concept.