![](bgb.png)
ground truth bounding box, in the second column is
the Non-Quantized CAM, and in the third column is
the Quantized CAM. The first row displays a predic-
tion from MobileNetV2 on a Leatherback Turtle, the
second row shows a prediction from MobileNetV3 on
a Tench fish, the third row presents a prediction from
ResNet18 on a Tiger Shark, and finally, the fourth
row exhibits the ShuffleNetV2 with a prediction on
a Goldfish. In each case, it is evident that the CAMs
generated by the Quantized CNNs produce more con-
cise bounding boxes for the respective images. This
observation aligns with the reported WSOL IoU val-
ues.
5 CONCLUSION
To conclude, we have visualized and statistically
compared activations from Quantized and Non-
Quantized CNNs and identified differences within the
activations themselves. Moreover, we have compared
Quantized CNN activations in a WSOL task and com-
pared the visualizations to their Non-Quantized CNN
counterparts. Through this visualization, we have
identified that Quantized CNNs utilize different fea-
tures and regions for image classification using the
ImageNet dataset. From this, we have demonstrated
that Quantized CNNs exhibit higher performance in
WSOL tasks and can be deployed in real-time using
EigenCAM, and are statisically different. Thus, quan-
tization should be considered in more academic pa-
pers, as it not only offers a more efficient network but
also provides a more interpretable network in some
cases.
For our future work, we will apply gradient-free
methodologies to activations, incorporating ViTs, in
order to explore the distinctions between Quantized
and Non-Quantized ViT models.
ACKNOWLEDGEMENTS
This work is supported by the Engineering and Phys-
ical Sciences Research Council [EP/S023917/1].
REFERENCES
Bany Muhammad, M. and Yeasin, M. (2021). Eigen-cam:
Visual explanations for deep convolutional neural net-
works. SN Computer Science, 2:1–14.
Chattopadhay, A., Sarkar, A., Howlader, P., and Balasub-
ramanian, V. N. (2018). Grad-cam++: Generalized
gradient-based visual explanations for deep convolu-
tional networks. In 2018 IEEE winter conference on
applications of computer vision (WACV), pages 839–
847. IEEE.
Ghimire, D., Kil, D., and Kim, S.-h. (2022). A survey on
efficient convolutional neural networks and hardware
acceleration. Electronics, 11(6).
Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W.,
and Keutzer, K. (2021). A survey of quantization
methods for efficient neural network inference.
He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep resid-
ual learning for image recognition.
Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B.,
Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V.,
Le, Q. V., and Adam, H. (2019). Searching for mo-
bilenetv3.
Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard,
A., Adam, H., and Kalenichenko, D. (2017). Quan-
tization and training of neural networks for efficient
integer-arithmetic-only inference.
Liang, T., Glossner, J., Wang, L., Shi, S., and Zhang,
X. (2021). Pruning and quantization for deep neu-
ral network acceleration: A survey. Neurocomputing,
461:370–403.
Liu, J., Tripathi, S., Kurup, U., and Shah, M. (2020). Prun-
ing algorithms to accelerate convolutional neural net-
works for edge applications: A survey. arXiv preprint
arXiv:2005.04275.
Ma, L., Hong, H., Meng, F., Wu, Q., and Wu, J. (2023).
Deep progressive asymmetric quantization based on
causal intervention for fine-grained image retrieval.
IEEE Transactions on Multimedia, pages 1–13.
Ma, N., Zhang, X., Zheng, H.-T., and Sun, J. (2018). Shuf-
flenet v2: Practical guidelines for efficient cnn archi-
tecture design.
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,
Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,
Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-
Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,
Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).
PyTorch: An Imperative Style, High-Performance
Deep Learning Library. In Wallach, H., Larochelle,
H., Beygelzimer, A., d’Alch
´
e Buc, F., Fox, E., and
Garnett, R., editors, Advances in Neural Information
Processing Systems 32, pages 8024–8035. Curran As-
sociates, Inc.
Petsiuk, V., Das, A., and Saenko, K. (2018). Rise: Ran-
domized input sampling for explanation of black-box
models.
Ramaswamy, H. G. et al. (2020). Ablation-cam: Vi-
sual explanations for deep convolutional network via
gradient-free localization. In proceedings of the
IEEE/CVF winter conference on applications of com-
puter vision, pages 983–991.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,
S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).
ImageNet Large Scale Visual Recognition Challenge.
International Journal of Computer Vision (IJCV),
115(3):211–252.
Sabih, M., Hannig, F., and Teich, J. (2020). Utilizing ex-
Evaluating the Use of Interpretable Quantized Convolutional Neural Networks for Resource-Constrained Deployment
119