Rahnamayan, ShahryarZaman Farsa, Davood2021-08-312022-03-292021-08-312022-03-292021-08-01https://hdl.handle.net/10155/1332Deep learning can cope with complex tasks; however, it suffers from the crucial problem of high dimensionality which makes it intractable to learn the patterns and classify the data. Importantly, having a fixed model cannot achieve high accuracy on every issue and the optimal solution is embodied in the architecture of the model tailored, specifically, for that issue. Motivated by these challenges, we propose a multi-objective evolutionary scheme that evolves autoencoders for the goal of dimension reduction. Throughout the evolution of each model, multiple minimization objectives are considered. These objectives can obtain compressed models that extract significant features while having the highest classification accuracy. In our case study, we use the extracted features of Histopathology images from a DenseNet trained to classify 30 subtypes of carcinomas from TCGA repository. As a result, we increased classification accuracy over 8% and compressed the representation of gigapixel images above 46,000 times, simultaneously.enEvolving autoencoderDigital pathologyDimension reductionNeural architecture searchMulti-objective optimizationThesis