View article

In machine learning (ML) and deep learning, once a model has been trained and deployed, learning new classes is difficult. In this context, class incremental learning has traditionally been accomplished by retraining an already existing deep neural network (DNN) across newly added classes using either fine tuning techniques, partially freezing the DNN, or by joint training. The drawback of traditional approaches is that either accuracy dramatically declines when adding new classes one at a time or compute power gradually increases. In this work, three conventional and ten continuous ML algorithms have been compared based on online learning techniques for class incremental learning in the plant domain. The basic premise is to initially learn healthy plants leaf data from publicly available Plant Village dataset and subsequently incrementally learn new plants leaf diseases in the first set of experiments. In a separate second series of experiments, the models were trained with plants leaf data from Plant Seedling dataset available in the public domain. Then, the models incrementally learned new plants on the go. The current comparison is motivated by the aforementioned caveats. Task agnostic accuracy and forgetting metrics were utilized to evaluate and compare results of all the algorithms used in the experiments.

Experimental results revealed that incremental ML models may learn new classes of plant diseases and plant types (as classes) incrementally and continually with only an insignificant forgetting, without retraining the model from scratch. These models tend to save the compute resources and training time without any downtime in …