Regularization Through Simultaneous Learning: A Case Study for Hop Classification

Overfitting remains a prevalent challenge in deep neural networks, leading to suboptimal real-world performance. Employing regularization techniques is a common strategy to counter this challenge, improving model generalization. This paper proposes Simultaneous Learning, a novel regularization approach drawing on Transfer Learning and Multi-task Learning principles, applied specifically to the classification of hop varieties - an integral component of beer production. Our approach harnesses the power of auxiliary datasets in synergy with the target dataset to amplify the acquisition of highly relevant features. Through a strategic modification of the model's final layer, we enable the simultaneous classification of both datasets without the necessity to treat them as disparate tasks. To realize this, we formulate a loss function that includes an inter-group penalty. We conducted experimental evaluations using the InceptionV3 and ResNet50 models, designating the UFOP-HVD hop leaf dataset as the target and ImageNet and PlantNet as auxiliary datasets. Our proposed method exhibited a substantial performance advantage over models without regularization and those adopting dropout regularization, with accuracy improvements ranging from 5 to 22 percentage points. Additionally, we introduce a technique for interpretability devised to assess the quality of features by analyzing correlations among class features in the network's convolutional layers.