Training calibration-based counterfactual explainers for deep learning models in medical image analysis

The rapid adoption of artificial intelligence methods in healthcare is coupled with the critical need for techniques to rigorously introspect models and thereby ensure that they behave reliably. This has led to the design of explainable AI techniques that uncover the relationships between discernible data signatures and model predictions. In this context, counterfactual explanations that synthesize small, interpretable changes to a given query while producing desired changes in model predictions have become popular. This under-constrained, inverse problem is vulnerable to introducing irrelevant feature manipulations, particularly when the model’s predictions are not well-calibrated. Hence, in this paper, we propose the TraCE (training calibration-based explainers) technique, which utilizes a novel uncertainty-based interval calibration strategy for reliably synthesizing counterfactuals. Given the wide-spread adoption of machine-learned solutions in radiology, our study focuses on deep models used for identifying anomalies in chest X-ray images. Using rigorous empirical studies, we demonstrate the superiority of TraCE explanations over several state-of-the-art baseline approaches, in terms of several widely adopted evaluation metrics. Our findings show that TraCE can be used to obtain a holistic understanding of deep models by enabling progressive exploration of decision boundaries, to detect shortcuts, and to infer relationships between patient attributes and disease severity.

Cryptocurrency Price Estimation Using Hyperparameterized Oscillatory Activation Functions in LSTM Networks

Activation functions are critical components of neural networks, helping the model learn highly-intricate dependencies, trends, and patterns. Non-linear activation functions allow the model to behave as a functional approximator, learning complex decision boundaries and multi-dimensional patterns in the data. Activation functions can be combined with one another to learn better representations with the objective of improving gradient flow, performance metrics reducing training time and computational cost. Recent work on oscillatory activation functions\cite{noel2021growing}\cite{noel2021biologically} showcased their ability to perform competitively on image classification tasks using a compact architecture. Our work proposes the utilization of these oscillatory activation functions for predicting the volume-weighted average of Bitcoin on the G-Research Cryptocurrency Dataset. We utilize a popular LSTM architecture for this task achieving competitive results when compared to popular activation functions formally used.

The time-course of distractor-based activation modulates effects of speed-accuracy tradeoffs in conflict tasks

The cognitive processes underlying the ability of human performers to trade speed for accuracy is often conceptualized within evidence accumulation models, but it is not yet clear whether and how these models can account for decision-making in the presence of various sources of conflicting information. In the present study, we provide evidence that speed-accuracy tradeoffs (SATs) can have opposing effects on performance across two different conflict tasks. Specifically, in a single preregistered experiment, the mean reaction time (RT) congruency effect in the Simon task increased, whereas the mean RT congruency effect in the Eriksen task decreased, when the focus was put on response speed versus accuracy. Critically, distributional RT analyses revealed distinct delta plot patterns across tasks, thus indicating that the unfolding of distractor-based response activation in time is sufficient to explain the opposing pattern of congruency effects. In addition, a recent evidence accumulation model with the notion of time-varying conflicting information was successfully fitted to the experimental data. These fits revealed task-specific time-courses of distractor-based activation and suggested that time pressure substantially decreases decision boundaries in addition to reducing the duration of non-decision processes and the rate of evidence accumulation. Overall, the present results suggest that time pressure can have multiple effects in decision-making under conflict, but that strategic adjustments of decision boundaries in conjunction with different time-courses of distractor-based activation can produce counteracting effects on task performance with different types of distracting sources of information.

Auditing and Debugging Deep Learning Models via Flip Points: Individual-Level and Group-Level Analysis

Deep learning models have been criticized for their lack of easy interpretation, which undermines confidence in their use for important applications. Nevertheless, they are consistently utilized in many applications, consequential to humans’ lives, usually because of their better performance. Therefore, there is a great need for computational methods that can explain, audit, and debug such models. Here, we use flip points to accomplish these goals for deep learning classifiers used in social applications. A trained deep learning classifier is a mathematical function that maps inputs to classes. By way of training, the function partitions its domain and assigns a class to each of the partitions. Partitions are defined by the decision boundaries which are expected to be geometrically complex. This complexity is usually what makes deep learning models powerful classifiers. Flip points are points on those boundaries and, therefore, the key to understanding and changing the functional behavior of models. We use advanced numerical optimization techniques and state-of-the-art methods in numerical linear algebra, such as rank determination and reduced-order models to compute and analyze them. The resulting insight into the decision boundaries of a deep model can clearly explain the model’s output on the individual level, via an explanation report that is understandable by non-experts. We also develop a procedure to understand and audit model behavior towards groups of people. We show that examining decision boundaries of models in certain subspaces can reveal hidden biases that are not easily detectable. Flip points can also be used as synthetic data to alter the decision boundaries of a model and improve their functional behaviors. We demonstrate our methods by investigating several models trained on standard datasets used in social applications of machine learning. We also identify the features that are most responsible for particular classifications and misclassifications. Finally, we discuss the implications of our auditing procedure in the public policy domain.

Land Cover Dynamics on the Lower Ganges–Brahmaputra Delta: Agriculture–Aquaculture Transitions, 1972–2017

Aquaculture in tropical and subtropical developing countries has expanded in recent years. This practice is controversial due to its potential for serious economic, food security, and environmental impacts—especially for intensive operations in and near mangrove ecosystems, where many shrimp species spawn. While considerable effort has been directed toward understanding aquaculture impacts, maps of spatial extent and multi-decade spatiotemporal dynamics remain sparse. This is in part because aquaculture ponds (ghers) can be challenging to distinguish from other shallow water targets on the basis of water-leaving radiance alone. Here, we focus on the Lower Ganges–Brahmaputra Delta (GBD), one of the most expansive areas of recent aquaculture growth on Earth and adjacent to the Sundarbans mangrove forest, a biodiversity hotspot. We use a combination of MODIS 16-day EVI composites and 45 years (1972–2017) of Landsat observations to characterize dominant spatiotemporal patterns in the vegetation phenology of the area, identify consistent seasonal optical differences between flooded ghers and other land uses, and quantify the multi-decade expansion of standing water bodies. Considerable non-uniqueness exists in the spectral signature of ghers on the GBD, propagating into uncertainty in estimates of spatial extent. We implement three progressive decision boundaries to explicitly quantify this uncertainty and provide liberal, moderate, and conservative estimates of flooded gher extent on three different spatial scales. Using multiple extents and multiple thresholds, we quantify the size distribution of contiguous regions of flooded gher extent at ten-year intervals. The moderate threshold shows standing water area within Bangladeshi polders to have expanded from less than 300 km2 in 1990 to over 1400 km2 in 2015. At all three scales investigated, the size distribution of standing water bodies is increasingly dominated by larger, more interconnected networks of flooded areas associated with aquaculture. Much of this expansion has occurred in immediate proximity to the Bangladeshi Sundarbans.

Adversarial Domain Adaptation of Asymmetric Mapping with Coral Alignment for Intelligent Fault Diagnosis

Rolling bearings play a vital role in the overall operation of rotating machineries. In practical diagnosis, many learning methods for variable speed fault diagnosis ignore task-specific decision boundaries, which make it very difficult to match feature distributions between different domains completely. Therefore, an adversarial domain adaptation of asymmetric mapping with coral alignment (ADA-AMCA) is presented to dispose this problem. By using the asymmetric mapping feature extractor, more features of specific domain with obvious distinction can be extracted. Meanwhile, combining the maximum classifier discrepancy of deep transfer to form an adversarial approach, and the task-specific decision boundary is taken into account, the class-level alignment between the features of source domain and target domain is attempted. For the sake of preventing degenerate learning which is possibly caused by asymmetric mapping and adversarial learning, the model is constrained by deep coral to extract more domain invariant features. Experimental results show that the proposed method can solve the variable speed fault diagnosis problem well, with high transfer accuracy and strong generalization.

Public Perceptions Concerning Responsibility for Climate Change Adaptation

For successful climate change adaptation, the distribution of responsibility within society is an important question. While the literature highlights the need for involving both public and private actors, little is still known of how citizens perceive their own and others’ responsibility, let alone the moral groundings for such perceptions. In this paper, we report the results of a survey regarding people’s attitudes towards different ways of distributing responsibility for climate change adaptation. The survey was distributed to citizens in six Swedish municipalities and completed by 510 respondents. A large number of respondents wanted to assign responsibility for making decisions about and implementing adaptation measures to local governments, but also to property owners, whereas the national government was raised as responsible for setting decision boundaries and for financial support. The most preferred principles for a fair distribution of responsibility among the respondents were desert, ability, efficiency and need, while the principle of equal shares found less support. All principles received some support, indicating that it is necessary to consider several principles when distributing responsibility for climate change adaptation. Compared to earlier studies, this study shows more nuanced perceptions on who should be responsible and on what moral grounds.

Context aware semantic adaptation network for cross domain implicit sentiment classification

Cross-domain sentiment classification could be attributed to two steps. The first step is used to extract the text representation, and the other is to reduce domain discrepancy. Existing methods mostly focus on learning the domain-invariant information, rarely consider using the domain-specific semantic information, which could help cross-domain sentiment classification; traditional adversarial-based models merely focus on aligning the global distribution ignore maximizing the class-specific decision boundaries. To solve these problems, we propose a context-aware semantic adaptation (CASA) network for cross-domain implicit sentiment classification (ISC). CASA can provide more semantic relationships and an accurate understanding of the emotion-changing process for ISC tasks lacking explicit emotion words. (1) To obtain inter- and intrasentence semantic associations, our model builds a context-aware heterogeneous graph (CAHG), which can aggregate the intrasentence dependency information and the intersentence node interaction information, followed by an attention mechanism that remains high-level domain-specific features. (2) Moreover, we conduct a new multigrain discriminator (MGD) to effectively reduce the interdomain distribution discrepancy and improve intradomain class discrimination. Experimental results demonstrate the effectiveness of different modules compared with existing models on the Chinese implicit emotion dataset and four public explicit datasets.