Classification-by-Components: Probabilistic Modeling of Reasoning over a Set of Components

Authors: Sascha Saralajew, Lars Holdijk, Maike Rees, Ebubekir Asan, Thomas Villmann

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The evaluation of the approach on MNIST shows that CBCs are viable classifiers. Additionally, we demonstrate that the inherent interpretability offers a profound understanding of the classification behavior such that we can explain the success of an adversarial attack. The method s scalability is successfully tested using the IMAGENET dataset.
Researcher Affiliation Collaboration 1Dr. Ing. h.c. F. Porsche AG, Weissach, Germany, sascha.saralajew@porsche.de 2University of Applied Sciences Mittweida, Mittweida, Germany, thomas.villmann@hs-mittweida.de
Pseudocode No The paper describes the network architecture and training process conceptually, but does not include any formal pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at www.github.com/saralajew/cbc_networks.
Open Datasets Yes We evaluate CBCs on MNIST [40] and IMAGENET [41].
Dataset Splits Yes The final test accuracies of both models are quasi equivalent and on average over three runs (97.33 0.19) %. The top-5 validation accuracy of 82.4% is on par with earlier CNN generations such as Alex Net with 82.8% [48].
Hardware Specification Yes The CBC was evaluated using one NVIDIA Tesla V100 32 GB GPU.
Software Dependencies No The paper mentions using specific optimizers (Adam) and activation functions (Swish, ReLU) and architectures (ResNet-50) by name and citation, but it does not specify software versions for any libraries or frameworks (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes The CBCs use the cosine similarity with Re LU activation as detection probability function. They are trained with the margin loss defined as Eq. (2) with φ (x) = Re LU(x + β), where β is a margin parameter, using the Adam optimizer [42]. The CNN feature extractors are implemented without the use of batch normalization [43], with Swish activation [44], and the convolutional filters constraint to a Euclidean norm of one. We trained the components and reasoning probabilities from scratch using random initialization. Moreover, the margin parameter β was set to 0.3.