Summary: | Deep Metric Learning (DML) supports the non-linearity problem faced when unsupervised learning is used; whereby multi-input data corresponds to one output. Hence, DML is suitable for managing large imbalanced datasets which are often faced in the production industry in which a handful of defective products were manufactured. One of the methods used in DML is the autoencoders. The use of autoencoder information can be utilized in overcoming the convergence problem which arises when random samples are used for model training [1]. Two separate autoencoders were trained for normal products and defective products respectively. Next, the Triplet network is trained to learn an embedding of the feature vector representation of the products. The embedding prevents the convergence problem by moving each sample closer to its reconstruction restored with the same class’s autoencoder and further from the opposite class’s autoencoder [1]. Eventually, it allocates each sample to the associated autoencoder’s class, which recovers the sample’s nearest reconstruction in the embedding space. Bachelor of Engineering (Computer Science)
|