Abstract
Traditional approaches for training classical neural networks require that all possible classes that the model might encounter be sampled and presented during initial training. Such a requirement limits the domain of problems solved. For instance, building a vehicle that can traverse uncharted territories would be challenging due to the difficulty of constructing a model that can account for all unknown situations. In this thesis, we are presenting a proof-of-concept framework and technique for a novel approach to continual lifelong learning that utilizes feature similarities and dissimilarities in a given batch of data to solve never seen-before tasks. Our approach has the advantage that it can be applied to both Euclidean data as well as graphs and can sustain notable accuracy across introductions of new classes without any retraining/rehearsal. The heart of our technique, Lign, is the leveraging of neural network fine tuning and pruning, commonly used in transfer learning, to temporarily remove certain weights from a network that detect key features of a previously solved tasks and reuses them to understand new problems. After pruning, the neural network can be utilized as an embedder with the aid of clustering techniques to label data based on genetic learned features. Lign_MNIST (are structured model tested on the MNIST data set) was able to demonstrate feature comparison and learning when shown unknown digits. Results were also found with other models that were tested on the CIFAR-100 and Cora data sets that provide further insights into the inner working of the technique.