iNNvestigate Neural Networks!

Authors: Maximilian Alber, Sebastian Lapuschkin, Philipp Seegerer, Miriam Hägele, Kristof T. Schütt, Grégoire Montavon, Wojciech Samek, Klaus-Robert Müller, Sven Dähne, Pieter-Jan Kindermans

JMLR 2019 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the versatility of i NNvestigate, we provide an analysis of image classifications for variety of state-of-the-art neural network architectures. ... For quantitative evaluations of (image) classification task we further provide an implementation of the method perturbation analysis (Samek et al., 2017).
Researcher Affiliation Academia Maximilian Alber EMAIL Technische Universit at Berlin, Machine Learning Group 10623 Berlin, Germany Sebastian Lapuschkin EMAIL Fraunhofer Heinrich Hertz Institute, Video Coding and Analytics 10587 Berlin, Germany Philipp Seegerer EMAIL Miriam H agele EMAIL Kristof T. Sch utt EMAIL Gr egoire Montavon EMAIL Technische Universit at Berlin, Machine Learning Group 10623 Berlin, Germany Wojciech Samek EMAIL Fraunhofer Heinrich Hertz Institute, Video Coding and Analytics 10587 Berlin, Germany Klaus-Robert M uller EMAIL Technische Universit at Berlin, Machine Learning Group 10623 Berlin, Germany Korea University, Department of Brain and Cognitive Engineering Seoul 02841, Korea Max Planck Institute for Informatics 66123 Saarbr ucken, Germany Sven D ahne EMAIL Pieter-Jan Kindermans EMAIL Technische Universit at Berlin, Machine Learning Group 10623 Berlin, Germany
Pseudocode Yes The corresponding Python code is: 1 import i n n v e s t i g a t e 2 model = create a keras model () 3 analyzer = i n n v e s t i g a t e . c r e a t e a n a l y z e r ( analyzer name , model ) 4 analyzer . f i t ( X train ) # i f needed 5 a n a l y s i s = analyzer . analyze ( X test )
Open Source Code Yes i NNvestigate is available at repository: https://github.com/albermax/innvestigate. ... Installation & license: i NNvestigate is published as open-source software under the BSD2-license and can be downloaded from: https://github.com/albermax/innvestigate.
Open Datasets Yes one can train required patterns for the methods on large data sets like Imagenet (Deng et al., 2009) in less than an hour using one GPU.
Dataset Splits No The paper mentions 'X_train' and 'X_test' in the example Python code snippet, implying a train/test split might be used for the analyzer fitting and analysis. However, it does not specify any concrete details about these splits, such as percentages, sample counts, or the methodology for creating them.
Hardware Specification No It is built as a Python 2 or 3 application on top of the popular and established Keras (Chollet et al., 2015) framework. This allows to use the library on various platforms and devices like CPUs and GPUs. ... one can train required patterns for the methods on large data sets like Imagenet (Deng et al., 2009) in less than an hour using one GPU.
Software Dependencies No It is built as a Python 2 or 3 application on top of the popular and established Keras (Chollet et al., 2015) framework. ... At the time of publication only the Tensor Flow (Abadi et al., 2016) Keras-backend is supported.
Experiment Setup No The paper discusses applying the iNNvestigate library to analyze various neural network architectures (VGG16, Inception V3, ResNet50, etc.). It mentions that the analyzer can be fitted to data ('analyzer.fit(X_train)'), but it does not provide specific hyperparameters (e.g., learning rate, batch size, number of epochs) or detailed training configurations for the analysis methods or the pre-trained neural networks used.