Incorporating Uncertainty into Deep Learning for Spoken Language Assessment

Malinin, Andrey; Ragni, Anton; Knill, Kate; Gales, Mark J. F.

doi:10.18653/v1/p17-2008

Cited by 61 publications

(126 citation statements)

References 6 publications

Supporting

Mentioning

123

Contrasting

Order By: Relevance

“…In this work all graders were constructed to predict scores for each of the five sections, then the scores averaged to yield the final score. Two feature-based graders were built; one GPbased [35] (GPtxt) and the other DNN-based [36] (DNNtxt). The features for these systems were the text features described in [35].…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Universal Adversarial Attacks on Spoken Language Assessment Systems

Vyas¹,

Gales²,

Knill³

2020

Interspeech 2020

View full text Add to dashboard Cite

There is an increasing demand for automated spoken language assessment (SLA) systems, partly driven by the performance improvements that have come from deep learning based approaches. One aspect of deep learning systems is that they do not require expert derived features, operating directly on the original signal such as a speech recognition (ASR) transcript. This, however, increases their potential susceptibility to adversarial attacks as a form of candidate malpractice. In this paper the sensitivity of SLA systems to a universal black-box attack on the ASR text output is explored. The aim is to obtain a single, universal phrase to maximally increase any candidate's score. Four approaches to detect such adversarial attacks are also described. All the systems, and associated detection approaches, are evaluated on a free (spontaneous) speaking section from a Business English test. It is shown that on deep learning based SLA systems the average candidate score can be increased by almost one grade level using a single six word phrase appended to the end of the response hypothesis. Although these large gains can be obtained, they can be easily detected based on detection shifts from the scores of a "traditional" Gaussian Process based grader.

show abstract

Section: Methodsmentioning

confidence: 99%

“…For the neural assessment system (Neurtxt) 5 , BERT was used to extract the word-embeddings, followed by a multi-headself attention mechanism [37]. The output of this process was then fed to the same DNN configuration as [36]. For the neural systems an ensemble of 10 models were built and the predictions averaged to yield the final score.…”

Section: Methodsmentioning

confidence: 99%

Universal Adversarial Attacks on Spoken Language Assessment Systems

Vyas¹,

Gales²,

Knill³

2020

Interspeech 2020

View full text Add to dashboard Cite

show abstract

“…The models are not appropriate to describe the data. These failures can be named uncertainty, and Malinin [27] has categorized them into two major groups:…”

Section: Uncertainty Quantification In Deep Learningmentioning

confidence: 99%

Automated Detection of Presymptomatic Conditions in Spinocerebellar Ataxia Type 2 Using Monte Carlo Dropout and Deep Neural Network Techniques with Electrooculogram Signals

Stoean

Atencia

et al. 2020

Sensors

View full text Add to dashboard Cite

Application of deep learning (DL) to the field of healthcare is aiding clinicians to make an accurate diagnosis. DL provides reliable results for image processing and sensor interpretation problems most of the time. However, model uncertainty should also be thoroughly quantified. This paper therefore addresses the employment of Monte Carlo dropout within the DL structure to automatically discriminate presymptomatic signs of spinocerebellar ataxia type 2 in saccadic samples obtained from electrooculograms. The current work goes beyond the common incorporation of this special type of dropout into deep neural networks and uses the uncertainty derived from the validation samples to construct a decision tree at the register level of the patients. The decision tree built from the uncertainty estimates obtained a classification accuracy of 81.18% in automatically discriminating control, presymptomatic and sick classes. This paper proposes a novel method to address both uncertainty quantification and explainability to develop reliable healthcare support systems.

show abstract

“…6(c)). The epistemic uncertainty is caused by an insufficient amount of data, and thus it appears far from data distribution and can detect out-of-distribution samples that are potentially mis-classified or with excessive errors [65,66,75]. With significantly more data samples, the decision of the DNN becomes stable, and the epistemic uncertainty decreases.…”

Section: Uncertainty Quantificationmentioning

confidence: 99%

Bayesian deep learning: A model-based interpretable approach

Matsubara

2020

NOLTA

View full text Add to dashboard Cite

Deep learning is considered to be a model-free, end-to-end, and black-box approach. It requires numerous data samples instead of expert knowledge on the target domain. Hence, it does not specify the mechanism and reasons for its decision making. This aspect is considered a critical limitation of deep learning. This paper introduces another viewpoint, namely Bayesian deep learning. Deep learning can be installed in any framework, such as Bayesian networks and reinforcement learning. Subsequently, an expert can implement the knowledge as the graph structure, accelerate learning, and obtain new knowledge on the target domain. The framework is termed as the deep generative model. Conversely, we can directly introduce the Bayesian modeling approach to deep learning. Subsequently, it is possible to explore deep learning with respect to the confidence of its decision making via uncertainty quantification of the output and detect wrong decision-making or anomalous inputs. Given the aforementioned approaches, it is possible to adjust the "brightness" of deep learning.

show abstract

Incorporating Uncertainty into Deep Learning for Spoken Language Assessment

Abstract: There is a growing demand for automatic assessment of spoken English proficiency.

Cited by 61 publications

References 6 publications

Universal Adversarial Attacks on Spoken Language Assessment Systems

Universal Adversarial Attacks on Spoken Language Assessment Systems

Automated Detection of Presymptomatic Conditions in Spinocerebellar Ataxia Type 2 Using Monte Carlo Dropout and Deep Neural Network Techniques with Electrooculogram Signals

Bayesian deep learning: A model-based interpretable approach

Contact Info

Product

Resources

About