Towards Transparent Deep Learning: Assessing Model Accuracy and Interpretability in CNNs

In this work, various CNN architectures that include the baseline, width-scaled, as well as depth-scaled designs were measured and their performance based on top-(1, 3, 5) accuracy. The baseline reported 71.47% Top-1, 91.99% Top-3, and 97.33% Top-5 with 361,034 parameters. Using scaling, the width of the model increased up to 75.24% (Top-1), 93.40% (Top-3), and 97.63% (Top-5) with parameterization increased to 1,430,666. The depth-scaled model achieved the best accuracy in Top-(1,3,5) as 81.99%, 95.67%, and 98.70%, respectively; however, such a model imposes a whopping 24,492,234 parameters on the complexity level. Advanced visualization techniques like Integrated Grad-CAM and LIME have been used for a better understanding of model behaviour. The results depict the fact that deeper models capture more detailed information, but this also introduces noise. The width-scaled model shows a balance between accuracy and interpretability as applied. Further analysis will further unveil the hard trade-of between accuracy and interpretability across the different variants of CNNs. In baseline, width-scaled, and depth-scaled models, two properties were analysed. In this work, the baseline model utilized its simplest architecture as a reference point, then scaled the width and depth of each net to try to improve performance. Accuracy and interpretability have been investigated through visualization techniques: integrated gradients and LIME. Some interesting findings are that while depth scaling is profoundly important to accuracy, it results in significantly increased cost to interpretability. On the other hand, width scaling balances reasonable accuracy with a level of interpretability at least comparable to less sophisticated models. Such findings feed into the discussion over how to design transparent and explainable deep learning models that will allow for more trustworthy and accountable AI solutions to be designed for real-world applications.