Convolutional neural networks are among the most successful architectures in deep learning. This success is at least partially attributable to the efficacy of spatial invariance as an inductive bias. Locally connected layers, which differ from convolutional layers in their lack of spatial invariance, usually perform poorly in practice. However, these observations still leave open the possibility that some degree of relaxation of spatial invariance may yield a better inductive bias than either convolution or local connectivity. To test this hypothesis, we design a method to relax the spatial invariance of a network layer in a controlled manner. In particular, we create a \textit{low-rank} locally connected layer, where the filter bank applied at each position is constructed as a linear combination of basis set of filter banks. By varying the number of filter banks in the basis set, we can control the degree of departure from spatial invariance. In our experiments, we find that relaxing spatial invariance improves classification accuracy over both convolution and locally connected layers across MNIST, CIFAR-10, and CelebA datasets. These results suggest that spatial invariance in convolution layers may be overly restrictive.

The process of neural network pruning with suitable fine-tuning and retraining can yield networks with considerably fewer parameters than the original with comparable degrees of accuracy. Typically, pruning methods require large, pre-trained networks as a starting point from which they perform a time-intensive iterative pruning and retraining algorithm. We propose a novel pruning in-training method that prunes a network real-time during training, reducing the overall training time to achieve an optimal compressed network. To do so, we introduce an activation density based analysis that identifies the optimal relative sizing or compression for each layer of the network. Our method removes the need for pre-training and is architecture agnostic, allowing it to be employed on a wide variety of systems. For VGG-19 and ResNet18 on CIFAR-10, CIFAR-100, and TinyImageNet, we obtain exceedingly sparse networks (up to 200x reduction in parameters and >60x reduction in inference compute operations in the best case) with comparable accuracies (up to 2%-3% loss with respect to the baseline network). By reducing the network size periodically during training, we achieve total training times that are shorter than those of previously proposed pruning methods. Furthermore, training compressed networks at different epochs with our proposed method yields considerable reduction in training compute complexity (1.6x -3.2x lower) at near iso-accuracy as compared to a baseline network trained entirely from scratch.

Drug discovery projects entail cycles of design, synthesis, and testing that yield a series of chemically related small molecules whose properties, such as binding affinity to a given target protein, are progressively tailored to a particular drug discovery goal. The use of deep learning technologies could augment the typical practice of using human intuition in the design cycle, and thereby expedite drug discovery projects. Here we present DESMILES, a deep neural network model that advances the state of the art in machine learning approaches to molecular design. We applied DESMILES to a previously published benchmark that assesses the ability of a method to modify input molecules to inhibit the dopamine receptor D2, and DESMILES yielded a 77% lower failure rate compared to state-of-the-art models. To explain the ability of DESMILES to hone molecular properties, we visualize a layer of the DESMILES network, and further demonstrate this ability by using DESMILES to tailor the same molecules used in the D2 benchmark test to dock more potently against seven different receptors.

Can meta-learning discover generic ways of processing time-series (TS) from a diverse dataset so as to greatly improve generalization on new TS coming from different datasets? This work provides positive evidence to demonstrate this using a broad meta-learning framework which we show subsumes many existing meta-learning algorithms as specific cases. We further identify via theoretical analysis the meta-learning adaptation mechanisms within N-BEATS, a recent neural TS forecasting model. Our meta-learning theory predicts that N-BEATS iteratively generates a subset of its task-specific parameters based on a given TS input, thus gradually expanding the expressive power of the architecture on-the-fly. Our empirical results emphasize the importance of meta-learning for successful zero-shot forecasting to new sources of TS, supporting the claim that it is viable to train a neural network on a source TS dataset and deploy it on a different target TS dataset without retraining, resulting in performance that is at least as good as that of state-of-practice univariate forecasting models.

As neural networks are increasingly employed in machine learning practice, organizations will have to determine how to share limited training resources among a diverse set of model training tasks. This paper studies jointly training multiple neural network models on a single GPU. We presents an empirical study of this operation, called pack, and end-to-end experiments that suggest significant improvements for hyperparameter search systems. Our research prototype is in TensorFlow, and we evaluate performance across different models (ResNet, MobileNet, DenseNet, and MLP) and training scenarios. The results suggest: (1) packing two models can bring up to 40% performance improvement over unpacked setups for a single training step and the improvement increases when packing more models; (2) the benefit of a pack primitive largely depends on a number of factors including memory capacity, chip architecture, neural network structure, and batch size; (3) there exists a trade-off between packing and unpacking when training multiple neural network models on limited resources; (4) a pack-based Hyperband is up to 2.7x faster than the original Hyperband training method in our experiment setting, with this improvement growing as memory size increases and subsequently the density of models packed.

Recently, several studies proposed methods to utilize some restricted classes of optimization problems as layers of deep neural networks. However, these methods are still in their infancy and require special treatments, i.e., analyzing the KKT condition, etc., for deriving the backpropagation formula. Instead, in this paper, we propose a method to utilize fixed-point iteration (FPI), a generalization of many types of numerical algorithms, as a network layer. We show that the derivative of an FPI layer depends only on the fixed point, and then we present a method to calculate it efficiently using another FPI which we call the backward FPI. The proposed method can be easily implemented based on the autograd functionalities in existing deep learning tools. Since FPI covers vast different types of numerical algorithms in machine learning and other fields, it has a lot of potential applications. In the experiments, the differentiable FPI layer is applied to two scenarios, i.e., gradient descent iterations for differentiable optimization problems and FPI with arbitrary neural network modules, of which the results demonstrate the simplicity and the effectiveness.

In this paper, we consider the problem of assessing the adversarial robustness of deep neural network models under both Markov chain Monte Carlo (MCMC) and Bayesian Dark Knowledge (BDK) inference approximations. We characterize the robustness of each method to two types of adversarial attacks: the fast gradient sign method (FGSM) and projected gradient descent (PGD). We show that full MCMC-based inference has excellent robustness, significantly outperforming standard point estimation-based learning. On the other hand, BDK provides marginal improvements. As an additional contribution, we present a storage-efficient approach to computing adversarial examples for large Monte Carlo ensembles using both the FGSM and PGD attacks.

We propose a novel topological layer for general deep learning models based on persistent landscapes, in which we can efficiently exploit underlying topological features of the input data structure. We use the robust DTM function and show differentiability with respect to layer inputs, for a general persistent homology with arbitrary filtration. Thus, our proposed layer can be placed anywhere in the network architecture and feed critical information on the topological features of input data into subsequent layers to improve the learnability of the networks toward a given task. A task-optimal structure of the topological layer is learned during training via backpropagation, without requiring any input featurization or data preprocessing. We provide a tight stability theorem, and show that the proposed layer is robust towards noise and outliers. We demonstrate the effectiveness of our approach by classification experiments on various datasets.

Convolutional neural networks (CNNs) often perform well, but their stability is poorly understood. To address this problem, we consider the simple prototypical problem of signal denoising, where classical approaches such as nonlinear diffusion, wavelet-based methods and regularisation offer provable stability guarantees. To transfer such guarantees to CNNs, we interpret numerical approximations of these classical methods as a specific residual network (ResNet) architecture. This leads to a dictionary which allows to translate diffusivities, shrinkage functions, and regularisers into activation functions, and enables a direct communication between the four research communities. On the CNN side, it does not only inspire new families of nonmonotone activation functions, but also introduces intrinsically stable architectures for an arbitrary number of layers.

In recent years, a variety of effective neural network-based methods for anomaly and cyber attack detection in industrial control systems (ICSs) have been demonstrated in the literature. Given their successful implementation and widespread use, there is a need to study adversarial attacks on such detection methods to better protect the systems that depend upon them. The extensive research performed on adversarial attacks on image and malware classification has little relevance to the physical system state prediction domain, which most of the ICS attack detection systems belong to. Moreover, such detection systems are typically retrained using new data collected from the monitored system, thus the threat of adversarial data poisoning is significant, however this threat has not yet been addressed by the research community. In this paper, we present the first study focused on poisoning attacks on online-trained autoencoder-based attack detectors. We propose two algorithms for generating poison samples, an interpolation-based algorithm and a back-gradient optimization-based algorithm, which we evaluate on both synthetic and real-world ICS data. We demonstrate that the proposed algorithms can generate poison samples that cause the target attack to go undetected by the autoencoder detector, however the ability to poison the detector is limited to a small set of attack types and magnitudes. When the poison-generating algorithms are applied to the popular SWaT dataset, we show that the autoencoder detector trained on the physical system state data is resilient to poisoning in the face of all ten of the relevant attacks in the dataset. This finding suggests that neural network-based attack detectors used in the cyber-physical domain are more robust to poisoning than in other problem domains, such as malware detection and image processing.