Does a Neural Network Really Encode Symbolic Concepts?
Authors: Mingjie Li, Quanshi Zhang
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical studies have verified that a well-trained DNN usually encodes sparse, transferable, and discriminative concepts, which is partially aligned with human intuition. |
| Researcher Affiliation | Academia | 1Shanghai Jiao Tong University. Correspondence to: Quanshi Zhang. Quanshi Zhang is the corresponding author. He is with the Department of Computer Science and Engineering, the John Hopcroft Center, at the Shanghai Jiao Tong University, China. |
| Pseudocode | No | The paper includes mathematical formulations and descriptions of processes but does not feature any clearly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | The code is released at https://github.com/sjtu-xai-lab/ interaction-concept. |
| Open Datasets | Yes | To this end, we trained various DNNs3 on tabular datasets (the tic-tac-toe dataset3 and the wifi dataset3), image datasets (the MNIST-3 dataset3 and the Celeb Aeyeglasses dataset3), and a point-cloud dataset (the Shape Net dataset3). |
| Dataset Splits | No | The paper mentions using different 'samples' from categories for analysis and discusses 'average explanation ratio ρ(k)', but it does not specify explicit training, validation, and test dataset splits with percentages or counts for reproducibility. |
| Hardware Specification | No | The paper describes the datasets and models used but does not provide any specific hardware details such as CPU/GPU models, memory, or cloud computing instances used for running experiments. |
| Software Dependencies | No | The paper mentions various DNN architectures (e.g., MLP-5, Res MLP-5, Le Net, Alex Net, Res Net, VGG, Point Net, OPT-1.3B) and methods, but it does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For tabular datasets, we used the UCI tic-tac-toe endgame dataset... We trained the following two MLPs on each tabular dataset. MLP-5 contained five fully connected layers with 100 neurons in each hidden layer... For image data, we used the following three datasets... We trained Le Net, Alex Net, Res Net, VGG... Based on the Shape Net dataset... we trained Point Net and Point Net++. |