Sharif Digital Repository

Deep Neural Networks, are predictive models in Machine Learning, that during the last decade they've had a great success. However being in an over-parametrized and highly non-convex regime, the analytical examinations of these models is quite a challenging task to do. The empirical developments of Neural Networks, and their distinguishing performance in prediction problems, has motivated researchers, to formalize a theoretical foundations for these models and provide us with a framework, in which one can explain and justify their behavior and properties. this framework is of great importance because it would help us to come to a better understanding of how these models work and also enables... 

DeepFake is a type of forgery that uses deep learning algorithms to make changes to audio and video content that the audience is unable to detect. Nowadays, due to the threats posed by the use of DeepFake to move people's faces in video, researchers' attention has been drawn to designing methods to detect this type of forgery. Detection methods are usually classified into two types. The first case is the extraction of features to detect forgery distortions, for example, the extraction of facial orientations to detect inconsistencies. The second case is the use of deep learning networks for feature extraction and classification, of which the EfficientNet network is an example. Despite the... 

One of the approaches used in learning theory is using information theoretic tools. The general idea of this approach is that if we show the algorithm did not memorize the dataset, we could guarantee generalization. Noisy channel model is one of the important concepts in this approach. A noisy channel is a lossy process which maps the data to a compressed format.There are two ways to use noisy channel model in literature: input compression and model compression. One of the main results of this thesis is to show that the input compression methods can not explain the generalization of algorithms (despite previous belief). On the contrast by fixing some of the problems faced in the model... 

Few-shot learning methods have achieved notable performance in recent years. However, fewshot learning in large-scale settings with hundreds of classes is still challenging. In this dissertation, we tackle the problems of large-scale few-shot learning by taking advantage of pre-trained foundation models. We recast the original problem in two levels with different granularity. At the coarse-grained level, we introduce a novel object recognition approach with robustness to sub-population shifts. At the fine-grained level, generative experts are designed for few-shot learning, specialized for different superclasses. A Bayesian schema is considered to combine coarse-grained information with... 

The aim of this thesis is to review the theory of deep learning with an experimental approach. In this thesis, we review researches that examine the impact of input selection on outputs in deep learning systems; Inputs we can control (samples, architecture, model size, optimizer, etc.) and outputs we can observe (the performance of the neural network, its test error, its parameters, etc.). Among the reviewed cases are the generalizability of deep learning systems, the effect of model components on its accuracy, interpolation and hyperparameters, as well as new phenomena in this field for which new frameworks have been defined  

Membership inference attacks are one of the most important privacy-violating attacks in machine learning, as well as infrastructure of more serious attacks such as data extraction attacks. Since membership inference attack is used as a measure to evaluate the level of privacy protection of machine learning models, different researches have investigated and provided new methods for this attack. However, the accuracy of these attacks has not been investigated on models trained with the latest techniques such as data augmentation and regularization techniques. In this research, we see that the Lira attack, the latest membership inference attack, which has much more power compared to previous...