Machine learning (ML) models are becoming increasingly common in the atmospheric science community with a wide range of applications. To enable users to understand what an ML model has learned, ML explainability has become a field of active research. In Part I of this two-part study, we described several explainability methods and demonstrated that feature rankings from different methods can substantially disagree with each other. It is unclear, though, whether the disagreement is overinflated due to some methods being less faithful in assigning importance. Herein, "faithfulness" or "fidelity" refer to the correspondence between the assigned feature importance and the contribution of the feature to model performance. In the present study, we evaluate the faithfulness of feature ranking methods using multiple methods. Given the sensitivity of explanation methods to feature correlations, we also quantify how much explainability faithfulness improves after correlated features are limited. Before dimensionality reduction, the feature relevance methods [e.g., SHAP, LIME, ALE variance, and logistic regression (LR) coefficients] were generally more faithful than the permutation importance methods due to the negative impact of correlated features. Once correlated features were reduced, traditional permutation importance became the most faithful method. In addition, the ranking uncertainty (i.e., the spread in rank assigned to a feature by the different ranking methods) was reduced by a factor of 2-10, and excluding less faithful feature ranking methods reduces it further. This study is one of the first to quantify the improvement in explainability from limiting correlated features and knowing the relative fidelity of different explainability methods.

The stochastic gradient Langevin Dynamics is one of the most fundamental algorithms to solve sampling problems and non-convex optimization appearing in several machine learning applications. Especially, its variance reduced versions have nowadays gained particular attention. In this paper, we study two variants of this kind, namely, the Stochastic Variance Reduced Gradient Langevin Dynamics and the Stochastic Recursive Gradient Langevin Dynamics. We prove their convergence to the objective distribution in terms of KL-divergence under the sole assumptions of smoothness and Log-Sobolev inequality which are weaker conditions than those used in prior works for these algorithms. With the batch size and the inner loop length set to $\sqrt{n}$, the gradient complexity to achieve an $\epsilon$-precision is $\tilde{O}((n+dn^{1/2}\epsilon^{-1})\gamma^2 L^2\alpha^{-2})$, which is an improvement from any previous analyses. We also show some essential applications of our result to non-convex optimization.

Speech Emotion Recognition (SER) is one of the essential perceptual methods of humans in understanding the situation and how to interact with others, therefore, in recent years, it has been tried to add the ability to recognize emotions to human-machine communication systems. Since the SER process relies on labeled data, databases are essential for it. Incomplete, low-quality or defective data may lead to inaccurate predictions. In this paper, we fixed the inconsistencies in Sharif Emotional Speech Database (ShEMO), as a Persian database, by using an Automatic Speech Recognition (ASR) system and investigating the effect of Farsi language models obtained from accessible Persian text corpora. We also introduced a Persian/Farsi ASR-based SER system that uses linguistic features of the ASR outputs and Deep Learning-based models.

Invariant learning methods try to find an invariant predictor across several environments and have become popular in OOD generalization. However, in situations where environments do not naturally exist in the data, they have to be decided by practitioners manually. Environment partitioning, which splits the whole training dataset into environments by algorithms, will significantly influence the performance of invariant learning and has been left undiscussed. A good environment partitioning method can bring invariant learning to applications with more general settings and improve its performance. We propose to split the dataset into several environments by finding low-correlated data subsets. Theoretical interpretations and algorithm details are both introduced in the paper. Through experiments on both synthetic and real data, we show that our Decorr method can achieve outstanding performance, while some other partitioning methods may lead to bad, even below-ERM results using the same training scheme of IRM. Machine learning methods achieve great successes in image classification, speech recognition, and many other areas. However, these methods rely on the assumption that training and testing data are independently and identically distributed.

Online learning (OL) from data streams is an emerging area of research that encompasses numerous challenges from stream processing, machine learning, and networking. Recent extensions of stream-processing platforms, such as Apache Kafka and Flink, already provide basic extensions for the training of neural networks in a stream-processing pipeline. However, these extensions are not scalable and flexible enough for many real-world use-cases, since they do not integrate the neural-network libraries as a first-class citizen into their architectures. In this paper, we present TensAIR, which provides an end-to-end dataflow engine for OL from data streams via a protocol to which we refer as asynchronous iterative routing. TensAIR supports the common dataflow operators, such as Map, Reduce, Join, and has been augmented by the data-parallel OL functions train and predict. These belong to the new Model operator, in which an initial TensorFlow model (either freshly initialized or pre-trained) is replicated among multiple decentralized worker nodes. Our decentralized architecture allows TensAIR to efficiently shard incoming data batches across the distributed model replicas, which in turn trigger the model updates via asynchronous stochastic gradient descent. We empirically demonstrate that TensAIR achieves a nearly linear scale-out in terms of (1) the number of worker nodes deployed in the network, and (2) the throughput at which the data batches arrive at the dataflow operators. We exemplify the versatility of TensAIR by investigating both sparse (Word2Vec) and dense (CIFAR-10) use-cases, for which we are able to demonstrate very significant performance improvements in comparison to Kafka, Flink, and Horovod. We also demonstrate the magnitude of these improvements by depicting the possibility of real-time concept drift adaptation of a sentiment analysis model trained over a Twitter stream.

ML models often operate within the context of a larger system that can adapt its response when the ML model is uncertain, such as falling back on safe defaults or a human in the loop. This commonly encountered operational context calls for principled techniques for training ML models with the option to abstain from predicting when uncertain. Selective neural networks are trained with an integrated option to abstain, allowing them to learn to recognize and optimize for the subset of the data distribution for which confident predictions can be made. However, optimizing selective networks is challenging due to the non-differentiability of the binary selection function (the discrete decision of whether to predict or abstain). This paper presents a general method for training selective networks that leverages the Gumbel-softmax reparameterization trick to enable selection within an end-to-end differentiable training framework. Experiments on public datasets demonstrate the potential of Gumbel-softmax selective networks for selective regression and classification.

In this work we leverage a weakly-labeled dataset of spectral data from NASAs IRIS satellite for the prediction of solar flares using the Multiple Instance Learning (MIL) paradigm. While standard supervised learning models expect a label for every instance, MIL relaxes this and only considers bags of instances to be labeled. This is ideally suited for flare prediction with IRIS data that consists of time series of bags of UV spectra measured along the instrument slit. In particular, we consider the readout window around the Mg II h&k lines that encodes information on the dynamics of the solar chromosphere. Our MIL models are not only able to predict whether flares occur within the next $\sim$25 minutes with accuracies of around 90%, but are also able to explain which spectral profiles were particularly important for their bag-level prediction. This information can be used to highlight regions of interest in ongoing IRIS observations in real-time and to identify candidates for typical flare precursor spectral profiles. We use k-means clustering to extract groups of spectral profiles that appear relevant for flare prediction. The recovered groups show high intensity, triplet red wing emission and single-peaked h and k lines, as found by previous works. They seem to be related to small-scale explosive events that have been reported to occur tens of minutes before a flare.

We propose a general approach for training survival analysis models that minimizes a worst-case error across all subpopulations that are large enough (occurring with at least a user-specified minimum probability). This approach uses a training loss function that does not know any demographic information to treat as sensitive. Despite this, we demonstrate that our proposed approach often scores better on recently established fairness metrics (without a significant drop in prediction accuracy) compared to various baselines, including ones which directly use sensitive demographic information in their training loss. Our code is available at: https://github.com/discovershu/DRO_COX

This paper studies the decentralized nonconvex optimization problem $\min_{x\in{\mathbb R}^d} f(x)\triangleq \frac{1}{m}\sum_{i=1}^m f_i(x)$, where $f_i(x)\triangleq \frac{1}{n}\sum_{j=1}^n f_{i,j}(x)$ is the local function on the $i$-th agent of the network. We propose a novel stochastic algorithm called DEcentralized probAbilistic Recursive gradiEnt deScenT (\DEAREST), which integrates the techniques of variance reduction, gradient tracking and multi-consensus. We construct a Lyapunov function that simultaneously characterizes the function value, the gradient estimation error and the consensus error for the convergence analysis. Based on this measure, we provide a concise proof to show DEAREST requires at most ${\mathcal O}(mn+\sqrt{mn}L\varepsilon^{-2})$ incremental first-order oracle (IFO) calls and ${\mathcal O}({L\varepsilon^{-2}}/{\sqrt{1-\lambda_2(W)}}\,)$ communication rounds to find an $\varepsilon$-stationary point in expectation, where $L$ is the smoothness parameter and $\lambda_2(W)$ is the second-largest eigenvalue of the gossip matrix $W$. We can verify both of the IFO complexity and communication complexity match the lower bounds. To the best of our knowledge, DEAREST is the first optimal algorithm for decentralized nonconvex finite-sum optimization.

In urban cities, visual information on and along roadways is likely to distract drivers and lead to missing traffic signs and other accident-prone (AP) features. To avoid accidents due to missing these visual cues, this paper proposes a visual notification of AP-features to drivers based on real-time images obtained via dashcam. For this purpose, Google Street View images around accident hotspots (areas of dense accident occurrence) identified by a real-accident dataset are used to train a novel attention module to classify a given urban scene into an accident hotspot or a non-hotspot (area of sparse accident occurrence). The proposed module leverages channel, point, and spatial-wise attention learning on top of different CNN backbones. This leads to better classification results and more certain AP-features with better contextual knowledge when compared with CNN backbones alone. Our proposed module achieves up to 92% classification accuracy. The capability of detecting AP-features by the proposed model were analyzed by a comparative study of three different class activation map (CAM) methods, which were used to inspect specific AP-features causing the classification decision. Outputs of CAM methods were processed by an image processing pipeline to extract only the AP-features that are explainable to drivers and notified using a visual notification system. Range of experiments was performed to prove the efficacy and AP-features of the system. Ablation of the AP-features taking 9.61%, on average, of the total area in each image increased the chance of a given area to be classified as a non-hotspot by up to 21.8%.