-
Simple Text Detoxification by Identifying a Linear Toxic Subspace in Language Model Embeddings
Authors:
Andrew Wang,
Mohit Sudhakar,
Yangfeng Ji
Abstract:
Large pre-trained language models are often trained on large volumes of internet data, some of which may contain toxic or abusive language. Consequently, language models encode toxic information, which makes the real-world usage of these language models limited. Current methods aim to prevent toxic features from appearing generated text. We hypothesize the existence of a low-dimensional toxic subs…
▽ More
Large pre-trained language models are often trained on large volumes of internet data, some of which may contain toxic or abusive language. Consequently, language models encode toxic information, which makes the real-world usage of these language models limited. Current methods aim to prevent toxic features from appearing generated text. We hypothesize the existence of a low-dimensional toxic subspace in the latent space of pre-trained language models, the existence of which suggests that toxic features follow some underlying pattern and are thus removable. To construct this toxic subspace, we propose a method to generalize toxic directions in the latent space. We also provide a methodology for constructing parallel datasets using a context based word masking system. Through our experiments, we show that when the toxic subspace is removed from a set of sentence representations, almost no toxic representations remain in the result. We demonstrate empirically that the subspace found using our method generalizes to multiple toxicity corpora, indicating the existence of a low-dimensional toxic subspace.
△ Less
Submitted 15 December, 2021;
originally announced December 2021.
-
Integrated Grad-CAM: Sensitivity-Aware Visual Explanation of Deep Convolutional Networks via Integrated Gradient-Based Scoring
Authors:
Sam Sattarzadeh,
Mahesh Sudhakar,
Konstantinos N. Plataniotis,
Jongseong Jang,
Yeonjeong Jeong,
Hyunwoo Kim
Abstract:
Visualizing the features captured by Convolutional Neural Networks (CNNs) is one of the conventional approaches to interpret the predictions made by these models in numerous image recognition applications. Grad-CAM is a popular solution that provides such a visualization by combining the activation maps obtained from the model. However, the average gradient-based terms deployed in this method unde…
▽ More
Visualizing the features captured by Convolutional Neural Networks (CNNs) is one of the conventional approaches to interpret the predictions made by these models in numerous image recognition applications. Grad-CAM is a popular solution that provides such a visualization by combining the activation maps obtained from the model. However, the average gradient-based terms deployed in this method underestimates the contribution of the representations discovered by the model to its predictions. Addressing this problem, we introduce a solution to tackle this issue by computing the path integral of the gradient-based terms in Grad-CAM. We conduct a thorough analysis to demonstrate the improvement achieved by our method in measuring the importance of the extracted representations for the CNN's predictions, which yields to our method's administration in object localization and model interpretation.
△ Less
Submitted 15 February, 2021;
originally announced February 2021.
-
Ada-SISE: Adaptive Semantic Input Sampling for Efficient Explanation of Convolutional Neural Networks
Authors:
Mahesh Sudhakar,
Sam Sattarzadeh,
Konstantinos N. Plataniotis,
Jongseong Jang,
Yeonjeong Jeong,
Hyunwoo Kim
Abstract:
Explainable AI (XAI) is an active research area to interpret a neural network's decision by ensuring transparency and trust in the task-specified learned models. Recently, perturbation-based model analysis has shown better interpretation, but backpropagation techniques are still prevailing because of their computational efficiency. In this work, we combine both approaches as a hybrid visual explan…
▽ More
Explainable AI (XAI) is an active research area to interpret a neural network's decision by ensuring transparency and trust in the task-specified learned models. Recently, perturbation-based model analysis has shown better interpretation, but backpropagation techniques are still prevailing because of their computational efficiency. In this work, we combine both approaches as a hybrid visual explanation algorithm and propose an efficient interpretation method for convolutional neural networks. Our method adaptively selects the most critical features that mainly contribute towards a prediction to probe the model by finding the activated features. Experimental results show that the proposed method can reduce the execution time up to 30% while enhancing competitive interpretability without compromising the quality of explanation generated.
△ Less
Submitted 15 February, 2021;
originally announced February 2021.
-
Explaining Convolutional Neural Networks through Attribution-Based Input Sampling and Block-Wise Feature Aggregation
Authors:
Sam Sattarzadeh,
Mahesh Sudhakar,
Anthony Lem,
Shervin Mehryar,
K. N. Plataniotis,
Jongseong Jang,
Hyunwoo Kim,
Yeonjeong Jeong,
Sangmin Lee,
Kyunghoon Bae
Abstract:
As an emerging field in Machine Learning, Explainable AI (XAI) has been offering remarkable performance in interpreting the decisions made by Convolutional Neural Networks (CNNs). To achieve visual explanations for CNNs, methods based on class activation map** and randomized input sampling have gained great popularity. However, the attribution methods based on these techniques provide lower reso…
▽ More
As an emerging field in Machine Learning, Explainable AI (XAI) has been offering remarkable performance in interpreting the decisions made by Convolutional Neural Networks (CNNs). To achieve visual explanations for CNNs, methods based on class activation map** and randomized input sampling have gained great popularity. However, the attribution methods based on these techniques provide lower resolution and blurry explanation maps that limit their explanation power. To circumvent this issue, visualization based on various layers is sought. In this work, we collect visualization maps from multiple layers of the model based on an attribution-based input sampling technique and aggregate them to reach a fine-grained and complete explanation. We also propose a layer selection strategy that applies to the whole family of CNN-based models, based on which our extraction framework is applied to visualize the last layers of each convolutional block of the model. Moreover, we perform an empirical analysis of the efficacy of derived lower-level information to enhance the represented attributions. Comprehensive experiments conducted on shallow and deep models trained on natural and industrial datasets, using both ground-truth and model-truth based evaluation metrics validate our proposed algorithm by meeting or outperforming the state-of-the-art methods in terms of explanation ability and visual quality, demonstrating that our method shows stability regardless of the size of objects or instances to be explained.
△ Less
Submitted 24 December, 2020; v1 submitted 1 October, 2020;
originally announced October 2020.
-
Experimental validation of XRF inversion code for Chandrayaan-1
Authors:
P. S. Athiray,
M. Sudhakar,
M. K. Tiwari,
S. Narendranath,
G. S. Lodha,
S. K. Deb,
P. Sreekumar,
S. K. Dash
Abstract:
We have developed an algorithm (x2abundance) to derive the lunar surface chemistry from X-ray fluorescence (XRF) data for the Chandrayaan-1 X-ray Spectrometer (C1XS) experiment. The algorithm converts the observed XRF line fluxes to elemental abundances with uncertainties. We validated the algorithm in the laboratory using high Z elements (20 < Z < 30) published in Athiray et al. (2013). In this p…
▽ More
We have developed an algorithm (x2abundance) to derive the lunar surface chemistry from X-ray fluorescence (XRF) data for the Chandrayaan-1 X-ray Spectrometer (C1XS) experiment. The algorithm converts the observed XRF line fluxes to elemental abundances with uncertainties. We validated the algorithm in the laboratory using high Z elements (20 < Z < 30) published in Athiray et al. (2013). In this paper, we complete the exercise of validation using samples containing low Z elements, which are also analogous to the lunar surface composition (ie., contains major elements between 11 < Z < 30). The paper summarizes results from XRF experiments performed on Lunar simulant (JSC-1A) and anorthosite using a synchrotron beam excitation. We also discuss results from the validation of x2abundance using Monte Carlo simulation (GEANT4 XRF simulation).
△ Less
Submitted 4 October, 2013;
originally announced October 2013.
-
Conceptual challenges and computational progress in X-ray simulation
Authors:
Maria Grazia Pia,
Mauro Augelli,
Marcia Begalli,
Chan-Hyeung Kim,
Lina Quintieri,
Paolo Saracco,
Hee Seo,
Manju Sudhakar,
Georg Weidenspointner,
Andreas Zoglauer
Abstract:
Recent developments and validation tests related to the simulation of X-ray fluorescence and PIXE with Geant4 are reviewed. They concern new models for PIXE, which has enabled the first Geant4-based simulation of PIXE in a concrete experimental application, and the experimental validation of the content of the EADL data library relevant to the simulation of X-ray fluorescence. Achievements and ope…
▽ More
Recent developments and validation tests related to the simulation of X-ray fluorescence and PIXE with Geant4 are reviewed. They concern new models for PIXE, which has enabled the first Geant4-based simulation of PIXE in a concrete experimental application, and the experimental validation of the content of the EADL data library relevant to the simulation of X-ray fluorescence. Achievements and open issues in this domain are discussed.
△ Less
Submitted 15 December, 2010;
originally announced December 2010.
-
New techniques in Monte Carlo simulation: experience with a prototype of generic programming application to Geant4 physics processes
Authors:
Maria Grazia Pia,
Mauro Augelli,
Marcia Begalli,
Lina Quintieri,
Paolo Saracco,
Manju Sudhakar,
Georg Weidenspointner,
Andreas Zoglauer
Abstract:
An investigation is in progress to evaluate extensively and quantitatively the possible benefits and drawbacks of new programming paradigms in a Monte Carlo simulation environment, namely in the domain of physics modeling. The prototype design and extensive benchmarks, including a variety of rigorous quantitative metrics, are presented. The results of this research project allow the evaluation of…
▽ More
An investigation is in progress to evaluate extensively and quantitatively the possible benefits and drawbacks of new programming paradigms in a Monte Carlo simulation environment, namely in the domain of physics modeling. The prototype design and extensive benchmarks, including a variety of rigorous quantitative metrics, are presented. The results of this research project allow the evaluation of new software techniques for their possible adoption in Monte Carlo simulation on objective, quantitative ground.
△ Less
Submitted 15 December, 2010;
originally announced December 2010.
-
Data libraries as a collaborative tool across Monte Carlo codes
Authors:
Mauro Augelli,
Marcia Begalli,
Mincheol Han,
Steffen Hauf,
Chan-Hyeung Kim,
Markus Kuster,
Maria Grazia Pia,
Lina Quintieri,
Paolo Saracco,
Hee Seo,
Manju Sudhakar,
Georg Eidenspointner,
Andreas Zoglauer
Abstract:
The role of data libraries in Monte Carlo simulation is discussed. A number of data libraries currently in preparation are reviewed; their data are critically examined with respect to the state-of-the-art in the respective fields. Extensive tests with respect to experimental data have been performed for the validation of their content.
The role of data libraries in Monte Carlo simulation is discussed. A number of data libraries currently in preparation are reviewed; their data are critically examined with respect to the state-of-the-art in the respective fields. Extensive tests with respect to experimental data have been performed for the validation of their content.
△ Less
Submitted 3 December, 2010;
originally announced December 2010.
-
New models for PIXE simulation with Geant4
Authors:
M. G. Pia,
G. Weidenspointner,
M. Augelli,
L. Quintieri,
P. Saracco,
M. Sudhakar,
A. Zoglauer
Abstract:
Particle induced X-ray emission (PIXE) is a physical effect that is not yet adequately modelled in Geant4. The current status as in Geant4 9.2 release is reviewed and new developments are described. The capabilities of the software prototype are illustrated in application to the shielding of the X-ray detectors of the eROSITA telescope on the upcoming Spectrum-X-Gamma space mission.
Particle induced X-ray emission (PIXE) is a physical effect that is not yet adequately modelled in Geant4. The current status as in Geant4 9.2 release is reviewed and new developments are described. The capabilities of the software prototype are illustrated in application to the shielding of the X-ray detectors of the eROSITA telescope on the upcoming Spectrum-X-Gamma space mission.
△ Less
Submitted 15 January, 2010;
originally announced January 2010.
-
Design and performance evaluations of generic programming techniques in a R&D prototype of Geant4 physics
Authors:
M. G. Pia,
P. Saracco,
M. Sudhakar,
A. Zoglauer,
M. Augelli,
E. Gargioni,
C. H. Kim,
L. Quintieri,
P. P. de Queiroz Filho,
D. de Souza Santos,
G. Weidenspointner,
M. Begalli
Abstract:
A R&D project has been recently launched to investigate Geant4 architectural design in view of addressing new experimental issues in HEP and other related physics disciplines. In the context of this project the use of generic programming techniques besides the conventional object oriented is investigated. Software design features and preliminary results from a new prototype implementation of Gea…
▽ More
A R&D project has been recently launched to investigate Geant4 architectural design in view of addressing new experimental issues in HEP and other related physics disciplines. In the context of this project the use of generic programming techniques besides the conventional object oriented is investigated. Software design features and preliminary results from a new prototype implementation of Geant4 electromagnetic physics are illustrated. Performance evaluations are presented. Issues related to quality assurance in Geant4 physics modelling are discussed.
△ Less
Submitted 15 January, 2010;
originally announced January 2010.
-
R&D on co-working transport schemes in Geant4
Authors:
M. G. Pia,
P. Saracco,
M. Sudhakar,
A. Zoglauer,
M. Augelli,
E. Gargioni,
C. H. Kim,
L. Quintieri,
P. P. de Queiroz Filho,
D. de Souza Santos,
G. Weidenspointner,
M. Begalli
Abstract:
A research and development (R&D) project related to the extension of the Geant4 toolkit has been recently launched to address fundamental methods in radiation transport simulation. The project focuses on simulation at different scales in the same experimental environment; this problem requires new methods across the current boundaries of condensed-random-walk and discrete transport schemes. The…
▽ More
A research and development (R&D) project related to the extension of the Geant4 toolkit has been recently launched to address fundamental methods in radiation transport simulation. The project focuses on simulation at different scales in the same experimental environment; this problem requires new methods across the current boundaries of condensed-random-walk and discrete transport schemes. The new developments have been motivated by experimental requirements in various domains, including nanodosimetry, astronomy and detector developments for high energy physics applications.
△ Less
Submitted 15 January, 2010;
originally announced January 2010.
-
Inter-Comparison and Validation of Geant4 Photon Interaction Models
Authors:
M. Augelli,
M. Begalli,
M. G. Pia,
P. P. Queiroz,
L. Quintieri,
D. Souza-Santos,
M. Sudhakar,
P. Saracco,
G. Weidenspointner,
A. Zoglauer
Abstract:
A R&D project, named Nano5, has been recently launched to study an architectural design in view of addressing new experimental issues related to particle transport in high energy physics and other related physics disciplines with Geant4. In this frame, the first step has involved the redesign of the photon interaction models currently available in Geant4; this task has motivated a thorough inves…
▽ More
A R&D project, named Nano5, has been recently launched to study an architectural design in view of addressing new experimental issues related to particle transport in high energy physics and other related physics disciplines with Geant4. In this frame, the first step has involved the redesign of the photon interaction models currently available in Geant4; this task has motivated a thorough investigation of the physics and computational features of these models, whose first results are presented here.
△ Less
Submitted 9 December, 2009;
originally announced December 2009.
-
Recent Developments on PIXE Simulation with Geant4
Authors:
M. G. Pia,
G. Weidenspointner,
M. Augelli,
L. Quintieri,
P. Saracco,
M. Sudhakar,
A. Zoglauer
Abstract:
Particle induced X-ray emission (PIXE) is an important physical effect that is not yet adequately modelled in Geant4. This paper provides a critical analysis of the problem domain associated with PIXE simulation and describes a set of software developments to improve PIXE simulation with Geant4. The capabilities of the developed software prototype are illustrated and applied to a study of the pa…
▽ More
Particle induced X-ray emission (PIXE) is an important physical effect that is not yet adequately modelled in Geant4. This paper provides a critical analysis of the problem domain associated with PIXE simulation and describes a set of software developments to improve PIXE simulation with Geant4. The capabilities of the developed software prototype are illustrated and applied to a study of the passive shielding of the X-ray detectors of the German eROSITA telescope on the upcoming Russian Spectrum-X-Gamma space mission.
△ Less
Submitted 9 December, 2009;
originally announced December 2009.
