We define a notion of complexity, motivated by considerations of circuit complexity, which quantifies the nonlinearity of the computation of a neural network, as well as a complementary measure of the effective dimension of feature representations. We investigate these observables both for trained networks for various datasets as well as explore their dynamics during training. These observables can be understood in a dual way as uncovering hidden internal structure of the datasets themselves as a function of scale or depth. The entropic character of the proposed notion of complexity should allow to transfer modes of analysis from neuroscience and statistical physics to the domain of artificial neural networks.

Gradient descent has been a central training principle for artificial neural networks from the early beginnings to today's deep learning networks. The most common implementation is the backpropagation algorithm for training feed-forward neural networks in a supervised fashion. Backpropagation involves computing the gradient of a loss function, with respect to the weights of the network, to update the weights and thus minimize loss. Although the mean square error is often used as a loss function, the general stochastic gradient descent principle does not immediately connect with a specific loss function. Another drawback of backpropagation has been the search for optimal values of two important training parameters, learning rate and momentum weight, which are determined empirically in most systems. The learning rate specifies the step size towards a minimum of the loss function when following the gradient, while the momentum weight considers previous weight changes when updating current weights. Using both parameters in conjunction with each other is generally accepted as a means to improving training, although their specific values do not follow immediately from standard backpropagation theory. This paper proposes a new information-theoretical loss function motivated by neural signal processing in a synapse. The new loss function implies a specific learning rate and momentum weight, leading to empirical parameters often used in practice. The proposed framework also provides a more formal explanation of the momentum term and its smoothing effect on the training process. All results taken together show that loss, learning rate, and momentum are closely connected. To support these theoretical findings, experiments for handwritten digit recognition show the practical usefulness of the proposed loss function and training parameters.

Lipschitz constants of neural networks have been explored in various contexts in deep learning, such as provable adversarial robustness, estimating Wasserstein distance, stabilising training of GANs, and formulating invertible neural networks. Such works have focused on bounding the Lipschitz constant of fully connected or convolutional networks, composed of linear maps and pointwise non-linearities. In this paper, we investigate the Lipschitz constant of self-attention, a non-linear neural network module widely used in sequence modelling. We prove that the standard dot-product self-attention is not Lipschitz, and propose an alternative L2 self-attention that is Lipschitz. We derive an upper bound on the Lipschitz constant of L2 self-attention and provide empirical evidence for its asymptotic tightness. To demonstrate the practical relevance of the theory, we formulate invertible self-attention and use it in a Transformer-based architecture for a character-level language modelling task.

The interest in developing smart cities has increased dramatically in recent years. In this context an intelligent transportation system depicts a major topic. The forecast of traffic flow is indispensable for an efficient intelligent transportation system. The traffic flow forecast is a difficult task, due to its stochastic and non linear nature. Besides classical statistical methods, neural networks are a promising possibility to predict future traffic flow. In our work, this prediction is performed with various recurrent neural networks. These are trained on measurements of induction loops, which are placed in intersections of the city. We utilized data from beginning of January to the end of July in 2018. Each model incorporates sequences of the measured traffic flow from all sensors and predicts the future traffic flow for each sensor simultaneously. A variety of model architectures, forecast horizons and input data were investigated. Most often the vector output model with gated recurrent units achieved the smallest error on the test set over all considered prediction scenarios. Due to the small amount of data, generalization of the trained models is limited.

We propose ARIANN, a low-interaction framework to perform private training and inference of standard deep neural networks on sensitive data. This framework implements semi-honest 2-party computation and leverages function secret sharing, a recent cryptographic protocol that only uses lightweight primitives to achieve an efficient online phase with a single message of the size of the inputs, for operations like comparison and multiplication which are building blocks of neural networks. Built on top of PyTorch, it offers a wide range of functions including ReLU, MaxPool and BatchNorm, and allows to use models like AlexNet or ResNet18. We report experimental results for inference and training over distant servers. Last, we propose an extension to support n-party private federated learning.

In this paper, we propose a structural design methodology called \textit{data-driven topology design}, which aims to obtain high-performance material distributions for a multi-objective optimization problem from the initially given material distributions in a given design domain. Its basic idea is iterating the following processes: (i) selecting the material distributions from a dataset according to Pareto optimality, (ii) generating new material distributions using a deep generative model with the selected material distributions as the training data, and (iii) integrating the generated material distributions into the dataset. Because of the nature of a deep generative model, the generated material distributions are diverse and inheriting features of the training data, which are material distributions on the Pareto front at that specific point. Therefore, it is expected that some of the generated material distributions are superior to the training data, whereas some are inferior, and the Pareto front is improved by integrating the generated material distributions into the dataset. The Pareto front is further improved by iterating the above processes. Data-driven topology design is used to enhance a support system for determining appropriate formulations of topology optimization problems, and its usefulness is demonstrated through numerical examples.

Robustness to adversarial attacks is an important concern due to the fragility of deep neural networks to small perturbations and has received an abundance of attention in recent years. Distributionally Robust Optimization (DRO), a particularly promising way of addressing this challenge, studies robustness via divergence-based uncertainty sets and has provided valuable insights into robustification strategies such as regularization. In the context of machine learning, the majority of existing results have chosen $f$-divergences, Wasserstein distances and more recently, the Maximum Mean Discrepancy (MMD) to construct uncertainty sets. We extend this line of work for the purposes of understanding robustness via regularization by studying uncertainty sets constructed with Integral Probability Metrics (IPMs) - a large family of divergences including the MMD, Total Variation and Wasserstein distances. Our main result shows that DRO under \textit{any} choice of IPM corresponds to a family of regularization penalties, which recover and improve upon existing results in the setting of MMD and Wasserstein distances. Due to the generality of our result, we show that other choices of IPMs correspond to other commonly used penalties in machine learning. Furthermore, we extend our results to shed light on adversarial generative modelling via $f$-GANs, constituting the first study of distributional robustness for the $f$-GAN objective. Our results unveil the inductive properties of the discriminator set with regards to robustness, allowing us to give positive comments for several penalty-based GAN methods such as Wasserstein-, MMD- and Sobolev-GANs. In summary, our results intimately link GANs to distributional robustness, extend previous results on DRO and contribute to our understanding of the link between regularization and robustness at large.

A traditional approach to initialization in deep neural networks (DNNs) is to sample the network weights randomly for preserving the variance of pre-activations. On the other hand, several studies show that during the training process, the distribution of stochastic gradients can be heavy-tailed especially for small batch sizes. In this case, weights and therefore pre-activations can be modeled with a heavy-tailed distribution that has an infinite variance but has a finite (non-integer) fractional moment of order $s$ with $s<2$. Motivated by this fact, we develop initialization schemes for fully connected feed-forward networks that can provably preserve any given moment of order $s \in (0, 2]$ over the layers for a class of activations including ReLU, Leaky ReLU, Randomized Leaky ReLU, and linear activations. These generalized schemes recover traditional initialization schemes in the limit $s \to 2$ and serve as part of a principled theory for initialization. For all these schemes, we show that the network output admits a finite almost sure limit as the number of layers grows, and the limit is heavy-tailed in some settings. This sheds further light into the origins of heavy tail during signal propagation in DNNs. We prove that the logarithm of the norm of the network outputs, if properly scaled, will converge to a Gaussian distribution with an explicit mean and variance we can compute depending on the activation used, the value of s chosen and the network width. We also prove that our initialization scheme avoids small network output values more frequently compared to traditional approaches. Furthermore, the proposed initialization strategy does not have an extra cost during the training procedure. We show through numerical experiments that our initialization can improve the training and test performance.

Randomized smoothing has achieved state-of-the-art certified robustness against $l_2$-norm adversarial attacks. However, it is not wholly resolved on how to find the optimal base classifier for randomized smoothing. In this work, we employ a Smoothed WEighted ENsembling (SWEEN) scheme to improve the performance of randomized smoothed classifiers. We theoretically analyze the expressive power of the SWEEN function class and show that SWEEN can be trained to achieve near-optimal risk in the randomized smoothing regime. We also develop an adaptive prediction algorithm to reduce the prediction and certification cost of SWEEN models. Extensive experiments show that SWEEN models outperform the upper envelope of their corresponding candidate models by a large margin. Moreover, SWEEN models constructed using a few small models can achieve comparable performance to a single large model with a notable reduction in training time.

Time series forecasting is ubiquitous in real-world applications. Examples include forecasting of demand to fill up inventories, economic growth forecasts to inform policies, and predicting stock prices to guide financial decisions. Forecasting is also a fruitful area for machine learning research, and pure and hybrid machine learning approaches have recently achieved state-of-the-art performance [1, 2]. In practice, forecasting involves a number of steps: we first need to specify, fit and select an appropriate model, and then evaluate and deploy it. There are various open-source toolboxes that help us implement these steps. However, most existing toolboxes are limited in important respects.