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.

EXPRESS: Automatically binding relevant and irrelevant features in visual working memory

It is generally assumed that, in order to save storage space, features are stored as integrated objects in visual working memory (VWM). Although such an object-based account does not always hold because features can be processed in parallel, a previous study has shown that different features can be automatically bound with their locations (task-irrelevant feature) into an integrated unit, resulting in improved memory performance (Wang, Cao, Theeuwes, Olivers, & Wang, 2016). The present study was designed to further explore this phenomenon by investigating whether other features, that are not spatial in origin, can act as the binding cue to form such automatical binding. To test this, we used three different features as binding cues (i.e., color, spatial frequency, and shape) over multiple separate experiments. The results consistently showed that when two features shared the same binding cue, memory performance was better relative to when each of those features had their own binding cue. We conclude, that any task-irrelevant feature can act as a binding cue to automatically bind with task-relevant features even across different objects, resulting in memory enhancement.

EXPRESS: Object-based Encoding in Visual Working Memory: A Critical Revisit

Visual working memory (VWM) is responsible for the temporal retention and manipulation of visual information. It has been suggested that VWM employs an object-based encoding (OBE) manner to extract highly-discriminable information from visual perception: Whenever one feature dimension of the objects is selected for entry into VWM, the other task-irrelevant highly-discriminable dimension is also extracted into VWM involuntarily. However, the task-irrelevant feature in OBE studies might reflect a high capacity fragile VWM trace (FVWM for short) that stores maskable sensory representations. To directly compare the VWM storage hypothesis and the FVWM storage hypothesis, we used a unique characteristic of FVWM that the representations in FVWM could be erased by backward masks presented at the original locations of the memory array. We required participants to memorize the orientations of three colored bars while ignoring their colors, and presented backward masks during the VWM maintenance interval. In four experiments, we consistently observed that the OBE occurs regardless of the presentation of the backward masks, except when even the task-relevant features in VWM were significantly interrupted by immediate backward masks, suggesting that the task-irrelevant features of objects are stored in VWM rather than in FVWM.

Are Irrelevant Items Actively Deleted from Visual Working Memory?: No Evidence from Repulsion Effects in Dual-Retrocue Tasks

In addition to processes associated with maintaining, manipulating, and updating to-be-remembered information for ongoing cognition, some theories suggest that working memory (WM) also involves the active deletion of irrelevant information, including items that were retained in WM, but are no longer relevant for ongoing cognition. Considerable evidence provides support for an active deletion mechanism, particularly for categorical representations (Rose et al., 2016; Fulvio & Postle, 2020; but see Bae & Luck, 2017 for contradictory evidence with line orientations). On each trial of the current task, healthy young adults maintained two line orientations in visual WM, switched attention to maintaining and recalling the orientation cued first, and then switched to recall the item cued second, at which point the uncued orientation was no-longer-relevant on the trial. The results showed that the no-longer-relevant items exerted the strongest “repulsive” bias on participants’ recall of to-be-remembered items, directly contradicting the active deletion hypothesis. We suggest that visual WM binds features like line orientations into ensemble representations, and an irrelevant feature of a bound object cannot be actively deleted--it biases recall of the target feature via repulsion. Models of WM will need to be updated to explain this dynamic phenomenon.

Noise correlations for faster and more robust learning

Distributed population codes are ubiquitous in the brain and pose a challenge to downstream neurons that must learn an appropriate readout. Here we explore the possibility that this learning problem is simplified through inductive biases implemented by stimulus-independent noise correlations that constrain learning to task-relevant dimensions. We test this idea in a set of neural networks that learn to perform a perceptual discrimination task. Correlations among similarly tuned units were manipulated independently of overall population signal-to-noise ratio in order to test how the format of stored information affects learning. Higher noise correlations among similarly tuned units led to faster and more robust learning, favoring homogenous weights assigned to neurons within a functionally similar pool, and could emerge through Hebbian learning. When multiple discriminations were learned simultaneously, noise correlations across relevant feature dimensions sped learning whereas those across irrelevant feature dimensions slowed it. Our results complement existing theory on noise correlations by demonstrating that when such correlations are produced without degradation of signal-to-noise ratio, they can improve readout learning by constraining it to appropriate dimensions.

Involuntary Orienting and Conflict Resolution during Auditory Attention: The Role of Ventral and Dorsal Streams

Selective attention to sound object features such as pitch and location is associated with enhanced brain activity in ventral and dorsal streams, respectively. We examined the role of these pathways in involuntary orienting and conflict resolution using fMRI. Participants were presented with two tones that may, or may not, share the same nonspatial (frequency) or spatial (location) auditory features. In separate blocks of trials, participants were asked to attend to sound frequency or sound location and ignore the change in the task-irrelevant feature. In both attend-frequency and attend-location tasks, RTs were slower when the task-irrelevant feature changed than when it stayed the same (involuntary orienting). This behavioral cost coincided with enhanced activity in the pFC and superior temporal gyrus. Conflict resolution was examined by comparing situations where the change in stimulus features was congruent (both features changed) and incongruent (only one feature changed). Participants were slower and less accurate for incongruent than congruent sound features. This congruency effect was associated with enhanced activity in the pFC and was greater in the right superior temporal gyrus and medial frontal cortex during the attend-location task than during the attend-frequency task. Together, these findings do not support a strict division of “labor” into ventral and dorsal streams but rather suggest interactions between these pathways in situations involving changes in task-irrelevant sound feature and conflict resolution. These findings also validate the Test of Attention in Listening task by revealing distinct neural correlates for involuntary orienting and conflict resolution.

The missing-colour effect: The attentional beam captures reading-relevant and reading-irrelevant information

According to many models, reading is driven by an attentional beam. In two experiments, we investigated the specificity of the beam by testing its sensitivity to a reading-irrelevant feature: colour. More specifically, participants were asked to read either a black-and-white version or a multi-colour version of the text in which each letter was printed in a different colour. In addition, while reading for comprehension, participants either searched for a target letter ( t or d) or for a colour (pink or black). In Experiment 1, we used the Nelson–Denny reading test and in Experiment 2, we used an experimental text. In both the experiments, the typical missing-letter effect was observed with letters: Participants missed more letters in function than in content words. Most importantly, although the effect was smaller, this pattern of results was also observed when participants searched for a colour (e.g., pink or black letters in a multi-coloured passage). Our results suggest that the attentional beam involved in reading is sensitive to both reading-relevant and reading-irrelevant information.

Building Extraction Based on U-Net with an Attention Block and Multiple Losses

Semantic segmentation of high-resolution remote sensing images plays an important role in applications for building extraction. However, the current algorithms have some semantic information extraction limitations, and these can lead to poor segmentation results. To extract buildings with high accuracy, we propose a multiloss neural network based on attention. The designed network, based on U-Net, can improve the sensitivity of the model by the attention block and suppress the background influence of irrelevant feature areas. To improve the ability of the model, a multiloss approach is proposed during training the network. The experimental results show that the proposed model offers great improvement over other state-of-the-art methods. For the public Inria Aerial Image Labeling dataset, the F1 score reached 76.96% and showed good performance on the Aerial Imagery for Roof Segmentation dataset.

Involuntary orienting and conflict resolution during auditory attention: The role of ventral and dorsal streams

Selective attention to sound object features such as pitch and location is associated with enhanced brain activity in ventral and dorsal streams, respectively. We examined the role of these pathways in involuntary orienting and conflict resolution using functional magnetic resonance imaging (fMRI). Participants were presented with two tones that may share, or not, the same non-spatial (frequency) or spatial (location) auditory features. In separate blocks of trials, participants were asked to attend to sound frequency or sound location and ignore the change in the task-irrelevant feature. In both attend-frequency and attend-location tasks, response times were slower when the task-irrelevant feature changed than when it stayed the same (involuntary orienting). This behavioural cost coincided with enhanced activity in the prefrontal cortex and superior temporal gyrus (STG). Conflict resolution was examined by comparing situations where the change in stimulus features was congruent (both features changed) and incongruent (only one feature changed). Participants were slower and less accurate for incongruent than congruent sound features. This congruency effect was associated with enhanced activity in the prefrontal cortex, and was greater in the right STG and medial frontal cortex during the attend-location than during the attend-frequency task. Together, these findings do not support a strict division of ‘labour’ into ventral and dorsal streams, but rather suggest interactions between these pathways in situations involving changes in task-irrelevant sound feature and conflict resolution. These findings also validate the Test of Attention in Listening task by revealing distinct neural correlates for involuntary orienting and conflict resolution.