Αποθηκεύτηκε σε:
| Κύριοι συγγραφείς: | , , |
|---|---|
| Μορφή: | Recurso digital |
| Γλώσσα: | Αρχαία αγγλική γλώσσα |
| Έκδοση: |
Zenodo
2026
|
| Θέματα: | |
| Διαθέσιμο Online: | https://doi.org/10.5281/zenodo.20392477 |
| Ετικέτες: |
Προσθήκη ετικέτας
Δεν υπάρχουν, Καταχωρήστε ετικέτα πρώτοι!
|
Πίνακας περιεχομένων:
- <p>The performance of deep neural networks is critically influenced by the choice of optimization algorithm and its hyperparameters, particularly the learning rate (η). However, the interplay between optimizers and learning rates across different task paradigms, such as classification and regression, remains under-explored. This paper addresses this gap by presenting a systematic empirical study of four common Optimizers (Adaptive Moment Estimation (Adam), Adaptive Gradient Algorithm (Adagrad), Root Mean Square Propagation(RMSProp), and Stochastic Gradient Descent(SGD) with Nesterov Accelerated Gradient momentum) across a spectrum of learning rates . We evaluate 32 configurations on two distinct tasks: a Convolutional Neural Network (CNN) for image classification and a Multilayer Perceptron (MLP) for tabular regression. Our findings demonstrate that optimizer performance, stability, and learning rate sensitivity are highly task-dependent. We show that the optimal configuration diverges significantly between tasks: the CNN classification task achieved peak accuracy with RMSProp at a low η =10−4, whereas the MLP regression task achieved the highest R2 score (0.9102) with the non-adaptive SGD at a moderate η =10−2. These results provide strong empirical evidence that optimization strategies must be carefully tailored to the specific task and architecture.</p>