The bias-variance trade-off is a central concept in supervised learning. In classical statistics, increasing the complexity of a model (e.g., number of parameters) reduces bias but also increases variance. Until recently, it was commonly believed that optimal performance is achieved at intermediate model complexities which strike a balance between bias and variance. Modern Deep Learning methods flout this dogma, achieving state-of-the-art performance using "over-parameterized models" where the number of fit parameters is large enough to perfectly fit the training data. As a result, understanding bias and variance in over-parameterized models has emerged as a fundamental problem in machine learning. Here, we use methods from statistical physics to derive analytic expressions for bias and variance in three minimal models for over-parameterization (linear regression and two-layer neural networks with linear and nonlinear activation functions), allowing us to disentangle properties stemming from the model architecture and random sampling of data. All three models exhibit a phase transition to an interpolation regime where the training error is zero, with linear neural-networks possessing an additional phase transition between regimes with zero and nonzero bias. The test error diverges at the interpolation transition for all three models. However, beyond the transition, it decreases again for the neural network models due to a decrease in both bias and variance with model complexity. We also show that over-parameterized models can overfit even in the absence of noise. We synthesize these results to construct a holistic understanding of generalization error and the bias-variance trade-off in over-parameterized models.

Saliency methods are used extensively to highlight the importance of input features in model predictions. These methods are mostly used in vision and language tasks, and their applications to time series data is relatively unexplored. In this paper, we set out to extensively compare the performance of various saliency-based interpretability methods across diverse neural architectures, including Recurrent Neural Network, Temporal Convolutional Networks, and Transformers in a new benchmark of synthetic time series data. We propose and report multiple metrics to empirically evaluate the performance of saliency methods for detecting feature importance over time using both precision (i.e., whether identified features contain meaningful signals) and recall (i.e., the number of features with signal identified as important). Through several experiments, we show that (i) in general, network architectures and saliency methods fail to reliably and accurately identify feature importance over time in time series data, (ii) this failure is mainly due to the conflation of time and feature domains, and (iii) the quality of saliency maps can be improved substantially by using our proposed two-step temporal saliency rescaling (TSR) approach that first calculates the importance of each time step before calculating the importance of each feature at a time step.

Large-scale numerical simulations are used across many scientific disciplines to facilitate experimental development and provide insights into underlying physical processes, but they come with a significant computational cost. Deep neural networks (DNNs) can serve as highly-accurate surrogate models, with the capacity to handle diverse datatypes, offering tremendous speed-ups for prediction and many other downstream tasks. An important use-case for these surrogates is the comparison between simulations and experiments; prediction uncertainty estimates are crucial for making such comparisons meaningful, yet standard DNNs do not provide them. In this work we define the fundamental requirements for a DNN to be useful for scientific applications, and demonstrate a general variational inference approach to equip predictions of scalar and image data from a DNN surrogate model trained on inertial confinement fusion simulations with calibrated Bayesian uncertainties. Critically, these uncertainties are interpretable, meaningful and preserve physics-correlations in the predicted quantities.

Hyperparameter optimization of neural networks can be elegantly formulated as a bilevel optimization problem. While research on bilevel optimization of neural networks has been dominated by implicit differentiation and unrolling, hypernetworks such as Self-Tuning Networks (STNs) have recently gained traction due to their ability to amortize the optimization of the inner objective. In this paper, we diagnose several subtle pathologies in the training of STNs. Based on these observations, we propose the $\Delta$-STN, an improved hypernetwork architecture which stabilizes training and optimizes hyperparameters much more efficiently than STNs. The key idea is to focus on accurately approximating the best-response Jacobian rather than the full best-response function; we achieve this by reparameterizing the hypernetwork and linearizing the network around the current parameters. We demonstrate empirically that our $\Delta$-STN can tune regularization hyperparameters (e.g. weight decay, dropout, number of cutout holes) with higher accuracy, faster convergence, and improved stability compared to existing approaches.

We introduce implicit Bayesian neural networks, a simple and scalable approach for uncertainty representation in deep learning. Standard Bayesian approach to deep learning requires the impractical inference of the posterior distribution over millions of parameters. Instead, we propose to induce a distribution that captures the uncertainty over neural networks by augmenting each layer's inputs with latent variables. We present appropriate input distributions and demonstrate state-of-the-art performance in terms of calibration, robustness and uncertainty characterisation over large-scale, multi-million parameter image classification tasks.

Decision making for autonomous driving in urban environments is challenging due to the complexity of the road structure and the uncertainty in the behavior of diverse road users. Traditional methods consist of manually designed rules as the driving policy, which require expert domain knowledge, are difficult to generalize and might give sub-optimal results as the environment gets complex. Whereas, using reinforcement learning, optimal driving policy could be learned and improved automatically through several interactions with the environment. However, current research in the field of reinforcement learning for autonomous driving is mainly focused on highway setup with little to no emphasis on urban environments. In this work, a deep reinforcement learning based decision-making approach for high-level driving behavior is proposed for urban environments in the presence of pedestrians. For this, the use of Deep Recurrent Q-Network (DRQN) is explored, a method combining state-of-the art Deep Q-Network (DQN) with a long term short term memory (LSTM) layer helping the agent gain a memory of the environment. A 3-D state representation is designed as the input combined with a well defined reward function to train the agent for learning an appropriate behavior policy in a real-world like urban simulator. The proposed method is evaluated for dense urban scenarios and compared with a rule-based approach and results show that the proposed DRQN based driving behavior decision maker outperforms the rule-based approach.

Recent diarization technologies can be categorized into two approaches, i.e., clustering and end-to-end neural approaches, which have different pros and cons. The clustering-based approaches assign speaker labels to speech regions by clustering speaker embeddings such as x-vectors. While it can be seen as a current state-of-the-art approach that works for various challenging data with reasonable robustness and accuracy, it has a critical disadvantage that it cannot handle overlapped speech that is inevitable in natural conversational data. In contrast, the end-to-end neural diarization (EEND), which directly predicts diarization labels using a neural network, was devised to handle the overlapped speech. While the EEND, which can easily incorporate emerging deep-learning technologies, has started outperforming the x-vector clustering approach in some realistic database, it is difficult to make it work for `long' recordings (e.g., recordings longer than 10 minutes) because of, e.g., its huge memory consumption. Block-wise independent processing is also difficult because it poses an inter-block label permutation problem, i.e., an ambiguity of the speaker label assignments between blocks. In this paper, we propose a simple but effective hybrid diarization framework that works with overlapped speech and for long recordings containing an arbitrary number of speakers. It modifies the conventional EEND framework to simultaneously output global speaker embeddings so that speaker clustering can be performed across blocks to solve the permutation problem. With experiments based on simulated noisy reverberant 2-speaker meeting-like data, we show that the proposed framework works significantly better than the original EEND especially when the input data is long.

Deep unfolding is a promising deep-learning technique, whose network architecture is based on expanding the recursive structure of existing iterative algorithms. Although convergence acceleration is a remarkable advantage of deep unfolding, its theoretical aspects have not been revealed yet. The first half of this study details the theoretical analysis of the convergence acceleration in deep-unfolded gradient descent (DUGD) whose trainable parameters are step sizes. We propose a plausible interpretation of the learned step-size parameters in DUGD by introducing the principle of Chebyshev steps derived from Chebyshev polynomials. The use of Chebyshev steps in gradient descent (GD) enables us to bound the spectral radius of a matrix governing the convergence speed of GD, leading to a tight upper bound on the convergence rate. The convergence rate of GD using Chebyshev steps is shown to be asymptotically optimal, although it has no momentum terms. We also show that Chebyshev steps numerically explain the learned step-size parameters in DUGD well. In the second half of the study, %we apply the theory of Chebyshev steps and Chebyshev-periodical successive over-relaxation (Chebyshev-PSOR) is proposed for accelerating linear/nonlinear fixed-point iterations. Theoretical analysis and numerical experiments indicate that Chebyshev-PSOR exhibits significantly faster convergence for various examples such as Jacobi method and proximal gradient methods.

Irregularly sampled time series (ISTS) data has irregular temporal intervals between observations and different sampling rates between sequences. ISTS commonly appears in healthcare, economics, and geoscience. Especially in the medical environment, the widely used Electronic Health Records (EHRs) have abundant typical irregularly sampled medical time series (ISMTS) data. Developing deep learning methods on EHRs data is critical for personalized treatment, precise diagnosis and medical management. However, it is challenging to directly use deep learning models for ISMTS data. On the one hand, ISMTS data has the intra-series and inter-series relations. Both the local and global structures should be considered. On the other hand, methods should consider the trade-off between task accuracy and model complexity and remain generality and interpretability. So far, many existing works have tried to solve the above problems and have achieved good results. In this paper, we review these deep learning methods from the perspectives of technology and task. Under the technology-driven perspective, we summarize them into two categories - missing data-based methods and raw data-based methods. Under the task-driven perspective, we also summarize them into two categories - data imputation-oriented and downstream task-oriented. For each of them, we point out their advantages and disadvantages. Moreover, we implement some representative methods and compare them on four medical datasets with two tasks. Finally, we discuss the challenges and opportunities in this area.

Outcomes with a natural order commonly occur in prediction tasks and oftentimes the available input data are a mixture of complex data, like images, and tabular predictors. Although deep Learning (DL) methods have shown outstanding performance on image classification, most models treat ordered outcomes as unordered and lack interpretability. In contrast, classical ordinal regression models yield interpretable predictor effects but are limited to tabular input data. Here, we present the highly modular class of ordinal neural network transformation models (ONTRAMs). Transformation models use a parametric transformation function and a simple distribution to trade off flexibility and interpretability of individual model components. In ONTRAMs, this trade-off is achieved by additively decomposing the transformation function into terms for the tabular and image data using a set of jointly trained neural networks. We show that the most flexible ONTRAMs achieve on-par performance with DL classifiers while outperforming them in training speed. We discuss how to interpret components of ONTRAMs in general and in the case of correlated tabular and image data. Taken together, ONTRAMs join benefits of DL and distributional regression to create interpretable prediction models for ordinal outcomes.