Over the years, as Convolutional Neural Networks (CNNs) have revolutionized the field of image processing by achieving state-of-the-art results, questions on their internal working mechanics have arisen. As of this moment in time, the internal workings of CNNs remain a “black box”, making it challenging to grasp and understand their decision-making processes. This study aims to unveil the mysteries of the black box by exploring how images are represented within CNNs. Accomplished with the TinyImageNet Data and VGG16 architecture, we extracted features from the penultimate layer and utilized K-means clustering to group the features. Through these clusters, we were able to uncover meaningful patterns and similarities among a vast amount of images, gaining insight into the internal mechanics of CNNs and the features they prioritize. Although parameters were adjusted to accommodate the purpose of understanding CNNs, a comparative analysis with other clustering methods is conducted to reveal more information about their efficacy and mechanics. This study has not only revealed more information about the internal workings of CNNs, but it also hopes to open the gates to more interpretable deep learning models in the future.

Ruiyi Zhang. Decoding the Black Box through a Comparative Study on Clustering Features in Convolutional Neural Networks. Academic Journal of Computing & Information Science (2023), Vol. 6, Issue 12: 118-124. https://doi.org/10.25236/AJCIS.2023.061213.

[1] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition.” in International Conference of Learning Representations (ICLR), 9 2015.

[2] Le, Ya, and Xuan S. Yang. "Tiny ImageNet Visual Recognition Challenge." 2015. Activeloop. 1312.6034.pdf (arxiv.org)

[3] S. Cai, L. Qiu, X. Chen, Q. Zhang, and Longteng Chen. “Semantic-Enhanced Image Clustering.” arXiv:2208.09849v2, 2023, https://arxiv.org/pdf/2208.09849.pdf

[4] J. Chang, L. Wang, G. Meng, S. Xiang, and Chunhong Pan. “Deep Adaptive Image Clustering.” Proceedings of the International Conference on Computer Vision (ICCV), 2017, pp. 5879-5887.

[5] S. Minaee, Y. Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, and Demetri Terzopoulos. “Image Segmentation Using Deep Learning: A Survey.” arXiv:2001.05566v5, 2020, https://arxiv.org/pdf/2001.05566.pdf

[6] V. K. Dehariya, S. K. Shrivastava and R. C. Jain, "Clustering of Image Data Set Using K-Means and Fuzzy K-Means Algorithms," 2010 International Conference on Computational Intelligence and Communication Networks, Bhopal, India, 2010, pp. 386-391, doi: 10.1109/CICN.2010.80.

[7] K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps.” arXiv:1312.6034v2, 2014, https://arxiv.org/pdf/1312.6034.pdf

[8] A Mahendran and A. Vedaldi, “Visualizing deep convolutional neural networks using natural pre-images.” arXiv:1512.02017v3, 2016, https://arxiv.org/pdf/1512.02017.pdf

[9] A. Dosovitskiy and T. Brox, “Inverting Visual Representations with Convolutional Networks.”, Proceedings of the International Conference on Computer Vision (ICCV), 2016, pp. 4829-4837.

[10] S. Ruder, “An overview of gradient descent optimization algorithms.” arXiv:1609.04747v2, 2017, https://arxiv.org/pdf/1609.04747.pdf

[11] Jin, X., Han, J. (2011). K-Means Clustering. In: Sammut, C., Webb, G.I. (eds) Encyclopedia of Machine Learning. Springer, Boston, MA. https://doi.org/10.1007/978-0-387-30164-8_425