Selecting an optimizer is a central step in the contemporary deep learning pipeline. In this paper, we demonstrate the sensitivity of optimizer comparisons to the metaparameter tuning protocol. Our findings suggest that the metaparameter search space may be the single most important factor explaining the rankings obtained by recent empirical comparisons in the literature. In fact, we show that these results can be contradicted when metaparameter search spaces are changed. As tuning effort grows without bound, more general optimizers should never underperform the ones they can approximate (i.e., Adam should never perform worse than momentum), but recent attempts to compare optimizers either assume these inclusion relationships are not practically relevant or restrict the metaparameters in ways that break the inclusions. In our experiments, we find that inclusion relationships between optimizers matter in practice and always predict optimizer comparisons. In particular, we find that the popular adaptive gradient methods never underperform momentum or gradient descent. We also report practical tips around tuning often ignored metaparameters of adaptive gradient methods and raise concerns about fairly benchmarking optimizers for neural network training.

Recently, machine learning (ML) has introduced advanced solutions to many domains. Since ML models provide business advantage to model owners, protecting intellectual property (IP) of ML models has emerged as an important consideration. Confidentiality of ML models can be protected by exposing them to clients only via prediction APIs. However, model extraction attacks can steal the functionality of ML models using the information leaked to clients through the results returned via the API. In this work, we question whether model extraction is a serious threat to complex, real-life ML models. We evaluate the current state-of-the-art model extraction attack (the Knockoff attack) against complex models. We reproduced and confirm the results in the Knockoff attack paper. But we also show that the performance of this attack can be limited by several factors, including ML model architecture and the granularity of API response. Furthermore, we introduce a defense based on distinguishing queries used for Knockoff attack from benign queries. Despite the limitations of the Knockoff attack, we show that a more realistic adversary can effectively steal complex ML models and evade known defenses.

Automatic text recognition from ancient handwritten record images is an important problem in the genealogy domain. However, critical challenges such as varying noise conditions, vanishing texts, and variations in handwriting make the recognition task difficult. We tackle this problem by developing a handwritten-to-machine-print conditional Generative Adversarial network (HW2MP-GAN) model that formulates handwritten recognition as a text-Image-to-text-Image translation problem where a given image, typically in an illegible form, is converted into another image, close to its machine-print form. The proposed model consists of three-components including a generator, and word-level and character-level discriminators. The model incorporates Sliced Wasserstein distance (SWD) and U-Net architectures in HW2MP-GAN for better quality image-to-image transformation. Our experiments reveal that HW2MP-GAN outperforms state-of-the-art baseline cGAN models by almost 30 in Frechet Handwritten Distance (FHD), 0.6 on average Levenshtein distance and 39% in word accuracy for image-to-image translation on IAM database. Further, HW2MP-GAN improves handwritten recognition word accuracy by 1.3% compared to baseline handwritten recognition models on the IAM database.

We introduce a pruning algorithm that provably sparsifies the parameters of a trained model in a way that approximately preserves the model's predictive accuracy. Our algorithm uses a small batch of input points to construct a data-informed importance sampling distribution over the network's parameters, and adaptively mixes a sampling-based and deterministic pruning procedure to discard redundant weights. Our pruning method is simultaneously computationally efficient, provably accurate, and broadly applicable to various network architectures and data distributions. Our empirical comparisons show that our algorithm reliably generates highly compressed networks that incur minimal loss in performance relative to that of the original network. We present experimental results that demonstrate our algorithm's potential to unearth essential network connections that can be trained successfully in isolation, which may be of independent interest.

Knowledge about frequency and location of snow avalanche activity is essential for forecasting and mapping of snow avalanche hazard. Traditional field monitoring of avalanche activity has limitations, especially when surveying large and remote areas. In recent years, avalanche detection in Sentinel-1 radar satellite imagery has been developed to overcome this monitoring problem. Current state-of-the-art detection algorithms, based on radar signal processing techniques, have highly varying accuracy that is on average much lower than the accuracy of visual detections from human experts. To reduce this gap, we propose a deep learning architecture for detecting avalanches in Sentinel-1 radar images. We trained a neural network on 6345 manually labelled avalanches from 117 Sentinel-1 images, each one consisting of six channels with backscatter and topographical information. Then, we tested the best network configuration on one additional SAR image. Comparing to the manual labelling (the gold standard), we achieved an F1 score above 66%, while the state-of-the-art detection algorithm produced an F1 score of 38%. A visual interpretation of the network's results shows that it only fails to detect small avalanches, while it manages to detect some that were not labelled by the human expert.

Deep reinforcement learning (RL) agents often fail to generalize to unseen environments (yet semantically similar to trained agents), particularly when they are trained on high-dimensional state spaces, such as images. In this paper, we propose a simple technique to improve a generalization ability of deep RL agents by introducing a randomized (convolutional) neural network that randomly perturbs input observations. It enables trained agents to adapt to new domains by learning robust features invariant across varied and randomized environments. Furthermore, we consider an inference method based on the Monte Carlo approximation to reduce the variance induced by this randomization. We demonstrate the superiority of our method across 2D CoinRun, 3D DeepMind Lab exploration and 3D robotics control tasks: it significantly outperforms various regularization and data augmentation methods for the same purpose.

Emotions recognition is the task of recognizing people's emotions. Usually it is achieved by analyzing expression of peoples faces. There are two ways for representing emotions: The categorical approach and the dimensional approach by using valence and arousal values. Valence shows how negative or positive an emotion is and arousal shows how much it is activated. Recent deep learning models, that have to do with emotions recognition, are using the second approach, valence and arousal. Moreover, a more interesting concept, which is useful in real life is the "in the wild" emotions recognition. "In the wild" means that the images analyzed for the recognition task, come from from real life sources(online videos, online photos, etc.) and not from staged experiments. So, they introduce unpredictable situations in the images, that have to be modeled. The purpose of this project is to study the previous work that was done for the "in the wild" emotions recognition concept, design a new dataset which has as a standard the "Aff-wild" database, implement new deep learning models and evaluate the results. First, already existing databases and deep learning models are presented. Then, inspired by them a new database is created which includes 507.208 frames in total from 106 videos, which were gathered from online sources. Then, the data are tested in a CNN model based on CNN-M architecture, in order to be sure about their usability. Next, the main model of this project is implemented. That is a Regression GAN which can execute unsupervised and supervised learning at the same time. More specifically, it keeps the main functionality of GANs, which is to produce fake images that look as good as the real ones, while it can also predict valence and arousal values for both real and fake images. Finally, the database created earlier is applied to this model and the results are presented and evaluated.

Glioma is one of the most common types of brain tumors arising in the glial cells in the human brain and spinal cord. In addition to the threat of death, glioma treatment is also very costly. Hence, automatic and accurate segmentation and measurement from the early stages are critical in order to prolong the survival rates of the patients and to reduce the costs of health care. In the present work, we propose a novel end-to-end cascaded network for semantic segmentation that utilizes the hierarchical structure of the tumor sub-regions with ResNet-like blocks and Squeeze-and-Excitation modules after each convolution and concatenation block. By utilizing cross-validation, an average ensemble technique, and a simple post-processing technique, we obtained dice scores of 90.34, 81.12, and 78.42 and Hausdorff Distances (95th percentile) of 4.32, 6.28, and 3.70 for the whole tumor, tumor core, and enhancing tumor, respectively, on the online validation set.

In the context of HCI, building an automatic system to recognize affect of human facial expression in real-world condition is very crucial to make machine interact naturallisticaly with a man. However, existing facial emotion databases usually contain expression in the limited scenario under well-controlled condition. Aff-Wild is currently the largest database consisting of spontaneous facial expression in the wild annotated with valence and arousal. The first contribution of this project is the completion of extending Aff-Wild database which is fulfilled by collecting videos from YouTube on which the videos have spontaneous facial expressions in the wild, annotating videos with valence and arousal ranging in [-1,1], detecting faces in frames using FFLD2 detector and partitioning the whole data set into train, validate and test set, with 527056, 94223 and 135145 frames. The diversity is guaranteed regarding age, ethnicity and values of valence and arousal. The ratio of male to female is close to 1. Regarding the techniques used to build the automatic system, deep learning is outstanding since almost all winning methods in emotion challenges adopt DNN techniques. The second contribution of this project is that an end-to-end DNN is constructed to have joint CNN and RNN block and gives the estimation on valence and arousal for each frame in sequential data. VGGFace, ResNet, DenseNet with the corresponding pre-trained model for CNN block and LSTM, GRU, IndRNN, Attention mechanism for RNN block are experimented aiming to find the best combination. Fine tuning and transfer learning techniques are also tried out. By comparing the CCC evaluation value on test data, the best model is found to be pre-trained VGGFace connected with 2 layers GRU with attention mechanism. The models test performance is 0.555 CCC for valence with sequence length 80 and 0.499 CCC for arousal with sequence length 70.

A BSTRACT Recently, the paradigm of unfolding iterative algorithms into finite-length feed-forward neural networks has achieved a great success in the area of sparse recovery. Benefit from available training data, the learned networks have achieved state-of-the-art performance in respect of both speed and accuracy. However, the structure behind sparsity, imposing constraint on the support of sparse signals, is often an essential prior knowledge but seldom considered in the existing networks. In this paper, we aim at bridging this gap. Specifically, exploiting the iterative reweighted null 1 minimization (IRL1) algorithm, we propose to learn the cluster structured sparsity (CSS) by rewegihting adaptively. In particular, we first unfold the Reweighted Iterative Shrinkage Algorithm (RwIST A) into an end-to-end train-able deep architecture termed as RW-LIST A. Then instead of the element-wise reweighting, the global and local reweighting manner are proposed for the cluster structured sparse learning. Numerical experiments further show the superiority of our algorithm against both classical algorithms and learning-based networks on different tasks. 1 I NTRODUCTION Sparsity is an important inherent property that describes the low-dimensionality of signals.