Refine
Document Type
- Conference Proceeding (2)
- Part of a Book (1)
Language
- English (3)
Is part of the Bibliography
- yes (3)
Keywords
Institute
- Fakultät IuI (2)
- Fakultät WiSo (1)
Receptive Field Analysis for Optimizing Convolutional Neural Network Architectures Without Training
(2023)
The number of layers in convolutional neural networks (CNN) is often overshot, when a convolutional neural network architecture is designed for an image-based task. These CNN-architectures are therefore unnecessarily costly to train and deploy. The increase in network-complexity also results in diminishing returns in terms of the predictive quality. The receptive field of a convolutional layer is strictly limiting the features it can process. We can consistently predict unproductive layers that will not contribute qualitatively to the test performance in a given CNN architecture, by analyzing the receptive field expansion over the network. Since the receptive field is a property of the architecture itself, this analysis does not require training the model. We refer to this analysis technique as Receptive Field Analysis (RFA). In this work, we demonstrate that RFA can be used to guide the optimization of CNN architectures by predicting the presence of unproductive layers. We show that RFA allows the deduction of design decisions and simple design strategies that reliably improve the parameter efficiency of the model on the given task. We further demonstrate that these RFA-guided strategies can reliably improve the predictive performance, computational efficiency or strike a balance between the two. Finally, we show that RFA can also be used to define an interval of feasible input resolutions for any modern architecture, in which the model will operate with high efficiency, while being able to extract any pattern from the image. This allows practitioners to pick efficient input resolutions when adapting models for novel tasks.
Venous leg ulcers and diabetic foot ulcers are the most common chronic wounds. Their prevalence has been increasing significantly over the last years, consuming scarce care resources. This study aimed to explore the performance of detection and classification algorithms for these types of wounds in images. To this end, algorithms of the YoloV5 family of pre-trained models were applied to 885 images containing at least one of the two wound types. The YoloV5m6 model provided the highest precision (0.942) and a high recall value (0.837). Its mAP_0.5:0.95 was 0.642. While the latter value is comparable to the ones reported in the literature, precision and recall were considerably higher. In conclusion, our results on good wound detection and classification may reveal a path towards (semi-) automated entry of wound information in patient records. To strengthen the trust of clinicians, we are currently incorporating a dashboard where clinicians can check the validity of the predictions against their expertise.
Should You Go Deeper? : Optimizing Convolutional Neural Network Architectures without Training
(2021)
When optimizing convolutional neural networks (CNN) for a specific image-based task, specialists commonly overshoot the number of convolutional layers in their designs. By implication, these CNNs are unnecessarily resource intensive to train and deploy, with diminishing beneficial effects on the predictive performance.The features a convolutional layer can process are strictly limited by its receptive field. By layer-wise analyzing the size of the receptive fields, we can reliably predict sequences of layers that will not contribute qualitatively to the test accuracy in the given CNN architecture. Based on this analysis, we propose design strategies based on a so-called border layer. This layer allows to identify unproductive convolutional layers and hence to resolve these inefficiencies, optimize the explainability and the computational performance of CNNs. Since neither the strategies nor the analysis requires training of the actual model, these insights allow for a very efficient design process of CNN architectures, which might be automated in the future.