Using model explanations to guide deep learning models towards consistent explanations for EHR data.

It has been shown that identical deep learning (DL) architectures will produce distinct explanations when trained with different hyperparameters that are orthogonal to the task (e.g. random seed, training set order). In domains such as healthcare and finance, where transparency and explainability is paramount, this can be a significant barrier to DL adoption. In this study we present a further analysis of explanation (in)consistency on 6 tabular datasets/tasks, with a focus on Electronic Health Records data. We propose a novel deep learning ensemble architecture that trains its sub-models to produce consistent explanations, improving explanation consistency by as much as 315% (e.g. from 0.02433 to 0.1011 on MIMIC-IV), and on average by 124% (e.g. from 0.12282 to 0.4450 on the BCW dataset). We evaluate the effectiveness of our proposed technique and discuss the implications our results have for both industrial applications of DL and explainability as well as future methodological work.

Beyond the topics: how deep learning can improve the discriminability of probabilistic topic modelling.

The article presents a discriminative approach to complement the unsupervised probabilistic nature of topic modelling. The framework transforms the probabilities of the topics per document into class-dependent deep learning models that extract highly discriminatory features suitable for classification. The framework is then used for sentiment analysis with minimum feature engineering. The approach transforms the sentiment analysis problem from the word/document domain to the topics domain making it more robust to noise and incorporating complex contextual information that are not represented otherwise. A stacked denoising autoencoder (SDA) is then used to model the complex relationship among the topics per sentiment with minimum assumptions. To achieve this, a distinct topic model and SDA per sentiment polarity is built with an additional decision layer for classification. The framework is tested on a comprehensive collection of benchmark datasets that vary in sample size, class bias and classification task. A significant improvement to the state of the art is achieved without the need for a sentiment lexica or over-engineered features. A further analysis is carried out to explain the observed improvement in accuracy.

It doesn't matter what you say: FMRI correlates of voice learning and recognition independent of speech content.

Listeners can recognize newly learned voices from previously unheard utterances, suggesting the acquisition of high-level speech-invariant voice representations during learning. Using functional magnetic resonance imaging (fMRI) we investigated the anatomical basis underlying the acquisition of voice representations for unfamiliar speakers independent of speech, and their subsequent recognition among novel voices. Specifically, listeners studied voices of unfamiliar speakers uttering short sentences and subsequently classified studied and novel voices as "old" or "new" in a recognition test. To investigate "pure" voice learning, i.e., independent of sentence meaning, we presented German sentence stimuli to non-German speaking listeners. To disentangle stimulus-invariant and stimulus-dependent learning, during the test phase we contrasted a "same sentence" condition in which listeners heard speakers repeating the sentences from the preceding study phase, with a "different sentence" condition. Voice recognition performance was above chance in both conditions although, as expected, performance was higher for same than for different sentences. During study phases activity in the left inferior frontal gyrus (IFG) was related to subsequent voice recognition performance and same versus different sentence condition, suggesting an involvement of the left IFG in the interactive processing of speaker and speech information during learning. Importantly, at test reduced activation for voices correctly classified as "old" compared to "new" emerged in a network of brain areas including temporal voice areas (TVAs) of the right posterior superior temporal gyrus (pSTG), as well as the right inferior/middle frontal gyrus (IFG/MFG), the right medial frontal gyrus, and the left caudate. This effect of voice novelty did not interact with sentence condition, suggesting a role of temporal voice-selective areas and extra-temporal areas in the explicit recognition of learned voice identity, independent of speech content.

"Hearing faces and seeing voices": Amodal coding of person identity in the human brain.

Recognizing familiar individuals is achieved by the brain by combining cues from several sensory modalities, including the face of a person and her voice. Here we used functional magnetic resonance (fMRI) and a whole-brain, searchlight multi-voxel pattern analysis (MVPA) to search for areas in which local fMRI patterns could result in identity classification as a function of sensory modality. We found several areas supporting face or voice stimulus classification based on fMRI responses, consistent with previous reports; the classification maps overlapped across modalities in a single area of right posterior superior temporal sulcus (pSTS). Remarkably, we also found several cortical areas, mostly located along the middle temporal gyrus, in which local fMRI patterns resulted in identity "cross-classification": vocal identity could be classified based on fMRI responses to the faces, or the reverse, or both. These findings are suggestive of a series of cortical identity representations increasingly abstracted from the input modality.

Gender differences in the temporal voice areas.

There is not only evidence for behavioral differences in voice perception between female and male listeners, but also recent suggestions for differences in neural correlates between genders. The fMRI functional voice localizer (comprising a univariate analysis contrasting stimulation with vocal vs. non-vocal sounds) is known to give robust estimates of the temporal voice areas (TVAs). However, there is growing interest in employing multivariate analysis approaches to fMRI data (e.g., multivariate pattern analysis; MVPA). The aim of the current study was to localize voice-related areas in both female and male listeners and to investigate whether brain maps may differ depending on the gender of the listener. After a univariate analysis, a random effects analysis was performed on female (n = 149) and male (n = 123) listeners and contrasts between them were computed. In addition, MVPA with a whole-brain searchlight approach was implemented and classification maps were entered into a second-level permutation based random effects models using statistical non-parametric mapping (SnPM; Nichols and Holmes, 2002). Gender differences were found only in the MVPA. Identified regions were located in the middle part of the middle temporal gyrus (bilateral) and the middle superior temporal gyrus (right hemisphere). Our results suggest differences in classifier performance between genders in response to the voice localizer with higher classification accuracy from local BOLD signal patterns in several temporal-lobe regions in female listeners.

Estimation of internal noise using double passes: does it matter how the second pass is delivered?

Human sensory processing is inherently noisy: if a participant is presented with the same set of stimuli multiple times and is asked to perform a task related to some property of the stimulus by pressing one of two buttons, the set of responses generated by the participant will differ on different presentations even though the set of stimuli remained the same. This response variability can be used to estimate the amount of internal noise (i.e. noise that is not present in the stimulus but in the participant's decision making process). The procedure by which the same set of stimuli is presented twice is referred to as double-pass (DP) methodology. This procedure is well-established, but there is no accepted recipe for how the repeated trials may be delivered (e.g. in the same order as they were originally presented, or in a different order); more importantly, it is not known whether the choice of delivery matters to the resulting estimates. Our results show that this factor (as well as feedback) has no measurable impact. We conclude that, for the purpose of estimating internal noise using the DP method, the system can be assumed to have no inter-trial memory.

Unsupervised movement onset detection from EEG recorded during self-paced real hand movement.

This article presents an unsupervised method for movement onset detection from electroencephalography (EEG) signals recorded during self-paced real hand movement. A Gaussian Mixture Model (GMM) is used to model the movement and idle-related EEG data. The GMM built along with appropriate classification and post processing methods are used to detect movement onsets using self-paced EEG signals recorded from five subjects, achieving True-False rate difference between 63 and 98%. The results show significant performance enhancement using the proposed unsupervised method, both in the sample-by-sample classification accuracy and the event-by-event performance, in comparison with the state-of-the-art supervised methods. The effectiveness of the proposed method suggests its potential application in self-paced Brain-Computer Interfaces (BCI).