Application of embeddings for multi-class classification with optional extendability

This study investigates the feasibility of an expandable image classification method utilizing a convolutional neural network to generate embeddings for use with simpler machine learning algorithms. The possibility of utilizing this approach to add new classes by additional training without modifying the topology of the vectorization network was shown on two datasets: MNIST and Fashion-MNIST. To achieve this, a straight classificational CNN was trained on both datasets using three first classes. The respective trained networks were then modified to generate embeddings instead of classification results. The added embedding generation layers for both networks were then trained using Triplet Loss to extract the features from the output of the convolutional layers, while maintaining distinction between classes. Several simpler machine learning algorithms were then trained to classify on the produced embeddings. To test the expandability hypothesis, fourth class was added to training datasets of both networks, and the embedding generation networks were subjected to additional training, with corresponding other machine learning algorithms retrained from scratch. The accuracy of machine learning algorithms on 3-class and 4-class networks was measured with the respective test datasets converted to embeddings. The time expenses were analyzed for both simple classification networks and the proposed method. The findings indicate that this approach can reduce retraining time and complexity, particularly for more complex image classification tasks, and also offers additional capabilities such as similarity search in vector databases. However, for simpler tasks, conventional classification networks remain more time-efficient. Ref. 8, fig. 4, tab. 2.