Performance of neural network models relies on the availability of large datasets with minimal levels of uncertainty. Transfer Learning (TL) models have been proposed to resolve the issue of small dataset size by letting the model train on a bigger, task-related reference dataset and then fine-tune on a smaller, task-specific dataset. In this work, we apply a transfer learning approach to improve predictive power in noisy data systems with large variable confidence datasets. We propose a deep neural network method called Filtered Transfer Learning (FTL) that defines multiple tiers of data confidence as separate tasks in a transfer learning setting. The deep neural network is fine-tuned in a hierarchical process by iteratively removing (filtering) data points with lower label confidence, and retraining. In this report we use FTL for predicting the interaction of drugs and proteins. We demonstrate that using FTL to learn stepwise, across the label confidence distribution, results in higher performance compared to deep neural network models trained on a single confidence range. We anticipate that this approach will enable the machine learning community to benefit from large datasets with uncertain labels in fields such as biology and medicine.

Neural ordinary differential equations (NODEs) have recently attracted increasing attention; however, their empirical performance on benchmark tasks (e.g. image classification) are significantly inferior to discrete-layer models. We demonstrate an explanation for their poorer performance is the inaccuracy of existing gradient estimation methods: the adjoint method has numerical errors in reverse-mode integration; the naive method directly back-propagates through ODE solvers, but suffers from a redundantly deep computation graph when searching for the optimal stepsize. We propose the Adaptive Checkpoint Adjoint (ACA) method: in automatic differentiation, ACA applies a trajectory checkpoint strategy which records the forward-mode trajectory as the reverse-mode trajectory to guarantee accuracy; ACA deletes redundant components for shallow computation graphs; and ACA supports adaptive solvers. On image classification tasks, compared with the adjoint and naive method, ACA achieves half the error rate in half the training time; NODE trained with ACA outperforms ResNet in both accuracy and test-retest reliability. On time-series modeling, ACA outperforms competing methods. Finally, in an example of the three-body problem, we show NODE with ACA can incorporate physical knowledge to achieve better accuracy. We provide the PyTorch implementation of ACA: \url{https://github.com/juntang-zhuang/torch-ACA}.

We propose a loss function for generative adversarial networks (GANs) using R\'{e}nyi information measures with parameter $\alpha$. More specifically, we formulate GAN's generator loss function in terms of R\'{e}nyi cross-entropy functionals. We demonstrate that for any $\alpha$, this generalized loss function preserves the equilibrium point satisfied by the original GAN loss based on the Jensen-Renyi divergence, a natural extension of the Jensen-Shannon divergence. We also prove that the R\'{e}nyi-centric loss function reduces to the original GAN loss function as $\alpha \to 1$. We show empirically that the proposed loss function, when implemented on both DCGAN (with $L_1$ normalization) and StyleGAN architectures, confers performance benefits by virtue of the extra degree of freedom provided by the parameter $\alpha$. More specifically, we show improvements with regard to: (a) the quality of the generated images as measured via the Fr\'echet Inception Distance (FID) score (e.g., best FID=8.33 for RenyiStyleGAN vs 9.7 for StyleGAN when evaluated over 64$\times$64 CelebA images) and (b) training stability. While it was applied to GANs in this study, the proposed approach is generic and can be used in other applications of information theory to deep learning, e.g., AI bias or privacy.

Random ordinary differential equations (RODEs), i.e. ODEs with random parameters, are often used to model complex dynamics. Most existing methods to identify unknown governing RODEs from observed data often rely on strong prior knowledge. Extracting the governing equations from data with less prior knowledge remains a great challenge. In this paper, we propose a deep neural network, called RODE-Net, to tackle such challenge by fitting a symbolic expression of the differential equation and the distribution of parameters simultaneously. To train the RODE-Net, we first estimate the parameters of the unknown RODE using the symbolic networks \cite{long2019pde} by solving a set of deterministic inverse problems based on the measured data, and use a generative adversarial network (GAN) to estimate the true distribution of the RODE's parameters. Then, we use the trained GAN as a regularization to further improve the estimation of the ODE's parameters. The two steps are operated alternatively. Numerical results show that the proposed RODE-Net can well estimate the distribution of model parameters using simulated data and can make reliable predictions. It is worth noting that, GAN serves as a data driven regularization in RODE-Net and is more effective than the $\ell_1$ based regularization that is often used in system identifications.

While many approaches exist in the literature to learn representations for data collections in multiple modalities, the generalizability of the learnt representations to previously unseen data is a largely overlooked subject. In this work, we first present a theoretical analysis of learning multi-modal nonlinear embeddings in a supervised setting. Our performance bounds indicate that for successful generalization in multi-modal classification and retrieval problems, the regularity of the interpolation functions extending the embedding to the whole data space is as important as the between-class separation and cross-modal alignment criteria. We then propose a multi-modal nonlinear representation learning algorithm that is motivated by these theoretical findings, where the embeddings of the training samples are optimized jointly with the Lipschitz regularity of the interpolators. Experimental comparison to recent multi-modal and single-modal learning algorithms suggests that the proposed method yields promising performance in multi-modal image classification and cross-modal image-text retrieval applications.

This paper introduces a network architecture, called dynoNet, utilizing linear dynamical operators as elementary building blocks. Owing to the dynamical nature of these blocks, dynoNet networks are tailored for sequence modeling and system identification purposes. The back-propagation behavior of the linear dynamical operator with respect to both its parameters and its input sequence is defined. This enables end-to-end training of structured networks containing linear dynamical operators and other differentiable units, exploiting existing deep learning software. Examples show the effectiveness of the proposed approach on well-known system identification benchmarks. Examples show the effectiveness of the proposed approach against well-known system identification benchmarks.

Deep learning methods for graphs have seen rapid progress in recent years with much focus awarded to generalising Convolutional Neural Networks (CNN) to graph data. CNNs are typically realised by alternating convolutional and pooling layers where the pooling layers subsample the grid and exchange spatial or temporal resolution for increased feature dimensionality. Whereas the generalised convolution operator for graphs has been studied extensively and proven useful, hierarchical coarsening of graphs is still challenging since nodes in graphs have no spatial locality and no natural order. This paper proposes two main contributions, the first is a differential module calculating structural similarity features based on the adjacency matrix. These structural similarity features may be used with various algorithms however in this paper the focus and the second main contribution is on integrating these features with a revisited pooling layer DiffPool arXiv:1806.08804 to propose a pooling layer referred to as SimPool. This is achieved by linking the concept of network reduction by means of structural similarity in graphs with the concept of hierarchical localised pooling. Experimental results demonstrate that as part of an end-to-end Graph Neural Network architecture SimPool calculates node cluster assignments that functionally resemble more to the locality preserving pooling operations used by CNNs that operate on local receptive fields in the standard grid. Furthermore the experimental results demonstrate that these features are useful in inductive graph classification tasks with no increase to the number of parameters.

Deep learning techniques play an increasingly important role in industrial and research environments due to their outstanding results. However, the large number of hyper-parameters to be set may lead to errors if they are set manually. The state-of-the-art hyper-parameters tuning methods are grid search, random search, and Bayesian Optimization. The first two methods are expensive because they try, respectively, all possible combinations and random combinations of hyper-parameters. Bayesian Optimization, instead, builds a surrogate model of the objective function, quantifies the uncertainty in the surrogate using Gaussian Process Regression and uses an acquisition function to decide where to sample the new set of hyper-parameters. This work faces the field of Hyper-Parameters Optimization (HPO). The aim is to improve Bayesian Optimization applied to Deep Neural Networks. For this goal, we build a new algorithm for evaluating and analyzing the results of the network on the training and validation sets and use a set of tuning rules to add new hyper-parameters and/or to reduce the hyper-parameter search space to select a better combination.

This paper presents fast non-sampling based methods to assess the risk of trajectories for autonomous vehicles when probabilistic predictions of other agents' futures are generated by deep neural networks (DNNs). The presented methods address a wide range of representations for uncertain predictions including both Gaussian and non-Gaussian mixture models for predictions of both agent positions and controls. We show that the problem of risk assessment when Gaussian mixture models (GMMs) of agent positions are learned can be solved rapidly to arbitrary levels of accuracy with existing numerical methods. To address the problem of risk assessment for non-Gaussian mixture models of agent position, we propose finding upper bounds on risk using Chebyshev's Inequality and sums-of-squares (SOS) programming; they are both of interest as the former is much faster while the latter can be arbitrarily tight. These approaches only require statistical moments of agent positions to determine upper bounds on risk. To perform risk assessment when models are learned for agent controls as opposed to positions, we develop TreeRing, an algorithm analogous to tree search over the ring of polynomials that can be used to exactly propagate moments of control distributions into position distributions through nonlinear dynamics. The presented methods are demonstrated on realistic predictions from DNNs trained on the Argoverse and CARLA datasets and are shown to be effective for rapidly assessing the probability of low probability events.

Short-term traffic speed prediction has been an important research topic in the past decade, and many approaches have been introduced. However, providing fine-grained, accurate, and efficient traffic-speed prediction for large-scale transportation networks where numerous traffic detectors are deployed has not been well studied. In this paper, we propose DistPre, which is a distributed fine-grained traffic speed prediction scheme for large-scale transportation networks. To achieve fine-grained and accurate traffic-speed prediction, DistPre customizes a Long Short-Term Memory (LSTM) model with an appropriate hyperparameter configuration for a detector. To make such customization process efficient and applicable for large-scale transportation networks, DistPre conducts LSTM customization on a cluster of computation nodes and allows any trained LSTM model to be shared between different detectors. If a detector observes a similar traffic pattern to another one, DistPre directly shares the existing LSTM model between the two detectors rather than customizing an LSTM model per detector. Experiments based on traffic data collected from freeway I5-N in California are conducted to evaluate the performance of DistPre. The results demonstrate that DistPre provides time-efficient LSTM customization and accurate fine-grained traffic-speed prediction for large-scale transportation networks.