-
Identifying and Mitigating Spurious Correlations for Improving Robustness in NLP Models
Authors:
Tianlu Wang,
Rohit Sridhar,
Diyi Yang,
Xuezhi Wang
Abstract:
Recently, NLP models have achieved remarkable progress across a variety of tasks; however, they have also been criticized for being not robust. Many robustness problems can be attributed to models exploiting spurious correlations, or shortcuts between the training data and the task labels. Most existing work identifies a limited set of task-specific shortcuts via human priors or error analyses, wh…
▽ More
Recently, NLP models have achieved remarkable progress across a variety of tasks; however, they have also been criticized for being not robust. Many robustness problems can be attributed to models exploiting spurious correlations, or shortcuts between the training data and the task labels. Most existing work identifies a limited set of task-specific shortcuts via human priors or error analyses, which requires extensive expertise and efforts. In this paper, we aim to automatically identify such spurious correlations in NLP models at scale. We first leverage existing interpretability methods to extract tokens that significantly affect model's decision process from the input text. We then distinguish "genuine" tokens and "spurious" tokens by analyzing model predictions across multiple corpora and further verify them through knowledge-aware perturbations. We show that our proposed method can effectively and efficiently identify a scalable set of "shortcuts", and mitigating these leads to more robust models in multiple applications.
△ Less
Submitted 25 May, 2022; v1 submitted 14 October, 2021;
originally announced October 2021.
-
Unsupervised Dimension Selection using a Blue Noise Spectrum
Authors:
Jayaraman J. Thiagarajan,
Rushil Anirudh,
Rahul Sridhar,
Peer-Timo Bremer
Abstract:
Unsupervised dimension selection is an important problem that seeks to reduce dimensionality of data, while preserving the most useful characteristics. While dimensionality reduction is commonly utilized to construct low-dimensional embeddings, they produce feature spaces that are hard to interpret. Further, in applications such as sensor design, one needs to perform reduction directly in the inpu…
▽ More
Unsupervised dimension selection is an important problem that seeks to reduce dimensionality of data, while preserving the most useful characteristics. While dimensionality reduction is commonly utilized to construct low-dimensional embeddings, they produce feature spaces that are hard to interpret. Further, in applications such as sensor design, one needs to perform reduction directly in the input domain, instead of constructing transformed spaces. Consequently, dimension selection (DS) aims to solve the combinatorial problem of identifying the top-$k$ dimensions, which is required for effective experiment design, reducing data while kee** it interpretable, and designing better sensing mechanisms. In this paper, we develop a novel approach for DS based on graph signal analysis to measure feature influence. By analyzing synthetic graph signals with a blue noise spectrum, we show that we can measure the importance of each dimension. Using experiments in supervised learning and image masking, we demonstrate the superiority of the proposed approach over existing techniques in capturing crucial characteristics of high dimensional spaces, using only a small subset of the original features.
△ Less
Submitted 31 October, 2018;
originally announced October 2018.
-
MARGIN: Uncovering Deep Neural Networks using Graph Signal Analysis
Authors:
Rushil Anirudh,
Jayaraman J. Thiagarajan,
Rahul Sridhar,
Peer-Timo Bremer
Abstract:
Interpretability has emerged as a crucial aspect of building trust in machine learning systems, aimed at providing insights into the working of complex neural networks that are otherwise opaque to a user. There are a plethora of existing solutions addressing various aspects of interpretability ranging from identifying prototypical samples in a dataset to explaining image predictions or explaining…
▽ More
Interpretability has emerged as a crucial aspect of building trust in machine learning systems, aimed at providing insights into the working of complex neural networks that are otherwise opaque to a user. There are a plethora of existing solutions addressing various aspects of interpretability ranging from identifying prototypical samples in a dataset to explaining image predictions or explaining mis-classifications. While all of these diverse techniques address seemingly different aspects of interpretability, we hypothesize that a large family of interepretability tasks are variants of the same central problem which is identifying \emph{relative} change in a model's prediction. This paper introduces MARGIN, a simple yet general approach to address a large set of interpretability tasks MARGIN exploits ideas rooted in graph signal analysis to determine influential nodes in a graph, which are defined as those nodes that maximally describe a function defined on the graph. By carefully defining task-specific graphs and functions, we demonstrate that MARGIN outperforms existing approaches in a number of disparate interpretability challenges.
△ Less
Submitted 15 January, 2021; v1 submitted 14 November, 2017;
originally announced November 2017.
-
How to Bypass Verified Boot Security in Chromium OS
Authors:
Mohammad Iftekhar Husain,
Lokesh Mandvekar,
Chunming Qiao,
Ramalingam Sridhar
Abstract:
Verified boot is an interesting feature of Chromium OS that supposedly can detect any modification in the root file system (rootfs) by a dedicated adversary. However, by exploiting a design flaw in verified boot, we show that an adversary can replace the original rootfs by a malicious rootfs containing exploits such as a spyware or keylogger and still pass the verified boot process. The exploit is…
▽ More
Verified boot is an interesting feature of Chromium OS that supposedly can detect any modification in the root file system (rootfs) by a dedicated adversary. However, by exploiting a design flaw in verified boot, we show that an adversary can replace the original rootfs by a malicious rootfs containing exploits such as a spyware or keylogger and still pass the verified boot process. The exploit is based on the fact that a dedicated adversary can replace the rootfs and the corresponding verification information in the bootloader. We experimentally demonstrate an attack using both the base and developer version of Chromium OS in which the adversary installs a spyware in the target system to send cached user data to the attacker machine in plain text which are otherwise encrypted, and thus inaccessible. We also demonstrate techniques to mitigate this vulnerability.
△ Less
Submitted 2 June, 2012; v1 submitted 23 February, 2012;
originally announced February 2012.
-
Robust seed selection algorithm for k-means type algorithms
Authors:
K. Karteeka Pavan,
Allam Appa Rao,
A. V. Dattatreya Rao,
G. R. Sridhar
Abstract:
Selection of initial seeds greatly affects the quality of the clusters and in k-means type algorithms. Most of the seed selection methods result different results in different independent runs. We propose a single, optimal, outlier insensitive seed selection algorithm for k-means type algorithms as extension to k-means++. The experimental results on synthetic, real and on microarray data sets demo…
▽ More
Selection of initial seeds greatly affects the quality of the clusters and in k-means type algorithms. Most of the seed selection methods result different results in different independent runs. We propose a single, optimal, outlier insensitive seed selection algorithm for k-means type algorithms as extension to k-means++. The experimental results on synthetic, real and on microarray data sets demonstrated that effectiveness of the new algorithm in producing the clustering results
△ Less
Submitted 7 February, 2012;
originally announced February 2012.
-
CD-PHY: Physical Layer Security in Wireless Networks through Constellation Diversity
Authors:
Mohammad Iftekhar Husain,
Suyash Mahant,
Ramalingam Sridhar
Abstract:
A common approach for introducing security at the physical layer is to rely on the channel variations of the wireless environment. This type of approach is not always suitable for wireless networks where the channel remains static for most of the network lifetime. For these scenarios, a channel independent physical layer security measure is more appropriate which will rely on a secret known to the…
▽ More
A common approach for introducing security at the physical layer is to rely on the channel variations of the wireless environment. This type of approach is not always suitable for wireless networks where the channel remains static for most of the network lifetime. For these scenarios, a channel independent physical layer security measure is more appropriate which will rely on a secret known to the sender and the receiver but not to the eavesdropper. In this paper, we propose CD-PHY, a physical layer security technique that exploits the constellation diversity of wireless networks which is independent of the channel variations. The sender and the receiver use a custom bit sequence to constellation symbol map** to secure the physical layer communication which is not known a priori to the eavesdropper. Through theoretical modeling and experimental simulation, we show that this information theoretic construct can achieve Shannon secrecy and any brute force attack from the eavesdropper incurs high overhead and minuscule probability of success. Our results also show that the high bit error rate also makes decoding practically infeasible for the eavesdropper, thus securing the communication between the sender and receiver.
△ Less
Submitted 25 August, 2011;
originally announced August 2011.
-
Comparison Of Modified Dual Ternary Indexing And Multi-Key Hashing Algorithms For Music Information Retrieval
Authors:
Rajeswari Sridhar,
A. Amudha,
S. Karthiga,
Geetha T V
Abstract:
In this work we have compared two indexing algorithms that have been used to index and retrieve Carnatic music songs. We have compared a modified algorithm of the Dual ternary indexing algorithm for music indexing and retrieval with the multi-key hashing indexing algorithm proposed by us. The modification in the dual ternary algorithm was essential to handle variable length query phrase and to acc…
▽ More
In this work we have compared two indexing algorithms that have been used to index and retrieve Carnatic music songs. We have compared a modified algorithm of the Dual ternary indexing algorithm for music indexing and retrieval with the multi-key hashing indexing algorithm proposed by us. The modification in the dual ternary algorithm was essential to handle variable length query phrase and to accommodate features specific to Carnatic music. The dual ternary indexing algorithm is adapted for Carnatic music by segmenting using the segmentation technique for Carnatic music. The dual ternary algorithm is compared with the multi-key hashing algorithm designed by us for indexing and retrieval in which features like MFCC, spectral flux, melody string and spectral centroid are used as features for indexing data into a hash table. The way in which collision resolution was handled by this hash table is different than the normal hash table approaches. It was observed that multi-key hashing based retrieval had a lesser time complexity than dual-ternary based indexing The algorithms were also compared for their precision and recall in which multi-key hashing had a better recall than modified dual ternary indexing for the sample data considered.
△ Less
Submitted 29 July, 2010;
originally announced July 2010.
-
(p,q)-Rogers-Szego polynomial and the (p,q)-oscillator
Authors:
R. Jagannathan,
R. Sridhar
Abstract:
A (p,q)-analogue of the classical Rogers-Szego polynomial is defined by replacing the q-binomial coefficient in it by the (p,q)-binomial coefficient. Exactly like the Rogers-Szego polynomial is associated with the q-oscillator algebra it is found that the (p,q)-Rogers-Szego polynomial is associated with the (p,q)-oscillator algebra.
A (p,q)-analogue of the classical Rogers-Szego polynomial is defined by replacing the q-binomial coefficient in it by the (p,q)-binomial coefficient. Exactly like the Rogers-Szego polynomial is associated with the q-oscillator algebra it is found that the (p,q)-Rogers-Szego polynomial is associated with the (p,q)-oscillator algebra.
△ Less
Submitted 24 May, 2010;
originally announced May 2010.
-
A q-Generalization of Product Densities and Janossy Functions in Stochastic Point Processes
Authors:
R. Parthasarathy,
R. Sridhar
Abstract:
A q-generalization of the product densities in stochastic point processes is developed. The properties of these functions are studied and a q-generalization of the usual $C^r_s$ coefficients is obtained. This for fixed q-number of particles coincides with the q-Stirling numbers of the second kind. The q-product densities are investigated using q-Poisson distribution and this shows that the stoch…
▽ More
A q-generalization of the product densities in stochastic point processes is developed. The properties of these functions are studied and a q-generalization of the usual $C^r_s$ coefficients is obtained. This for fixed q-number of particles coincides with the q-Stirling numbers of the second kind. The q-product densities are investigated using q-Poisson distribution and this shows that the stochastic point processes involving consistent q-generalization are inherently correlated. A closely related function to q-product densities is a q-generalized Janossy function and a relation between the two is established.
△ Less
Submitted 30 March, 2004;
originally announced March 2004.
-
On the q-analogues of the Zassenhaus formula for dientangling exponential operators
Authors:
R. Sridhar,
R. Jagannathan
Abstract:
Katriel, Rasetti and Solomon introduced a $q$-analogue of the Zassenhaus formula written as $e_q^{(A+B)}$ $=$ $e_q^Ae_q^Be_q^{c_2}e_q^{c_3}e_q^{c_4}e_q^{c_5}...$, where $A$ and $B$ are two generally noncommuting operators and $e_q^z$ is the Jackson $q$-exponential, and derived the expressions for $c_2$, $c_3$ and $c_4$. It is shown that one can also write $e_q^{(A+B)}$ $=$…
▽ More
Katriel, Rasetti and Solomon introduced a $q$-analogue of the Zassenhaus formula written as $e_q^{(A+B)}$ $=$ $e_q^Ae_q^Be_q^{c_2}e_q^{c_3}e_q^{c_4}e_q^{c_5}...$, where $A$ and $B$ are two generally noncommuting operators and $e_q^z$ is the Jackson $q$-exponential, and derived the expressions for $c_2$, $c_3$ and $c_4$. It is shown that one can also write $e_q^{(A+B)}$ $=$ $e_q^Ae_q^Be_{q^2}^{\C_2}e_{q^3}^{\C_3}e_{q^4}^{\C_4}e_{q^5}^{\C_5}...$. Explicit expressions for $\C_2$, $\C_3$ and $\C_4$ are given.
△ Less
Submitted 28 April, 2003; v1 submitted 24 December, 2002;
originally announced December 2002.
-
A Diagonal Representation of Quantum Density Matrix Using q-Boson Oscillator Coherent States
Authors:
R. Parthasarathy,
R. Sridhar
Abstract:
A q-analogue of Sudarshan's diagonal representation of the Quantum Mechanical density matrix is obtained using q-boson coherent states. Earlier result of Mehta and Sudarshan on the self reproducing property of rho(z',z) is also generalized and a self-consistent self-reproducing kernel {K-tilde}(z',z) is constructed.
A q-analogue of Sudarshan's diagonal representation of the Quantum Mechanical density matrix is obtained using q-boson coherent states. Earlier result of Mehta and Sudarshan on the self reproducing property of rho(z',z) is also generalized and a self-consistent self-reproducing kernel {K-tilde}(z',z) is constructed.
△ Less
Submitted 16 August, 2002;
originally announced August 2002.
-
Supersymmetry in Thermo Field Dynamics
Authors:
R. Parthasarathy,
R. Sridhar
Abstract:
By considering the enlarged thermal system including the heat bath, it is shown that this system has supersymmetry which is not broken at finite temperature. The super algebra is constructed and the Hamiltonian is expressed as the anti-commutator of two kinds of super charges. With this Hamiltonian and the thermal vacuum $\mid 0(β)>$, this supersymmetry is found to be preserved.
By considering the enlarged thermal system including the heat bath, it is shown that this system has supersymmetry which is not broken at finite temperature. The super algebra is constructed and the Hamiltonian is expressed as the anti-commutator of two kinds of super charges. With this Hamiltonian and the thermal vacuum $\mid 0(β)>$, this supersymmetry is found to be preserved.
△ Less
Submitted 21 June, 2000;
originally announced June 2000.
-
The N-Chain Hubbard model in the Composite Operator Method
Authors:
A. Avella,
F. Mancini,
M. M. Sanchez,
R. Sridhar
Abstract:
We propose a theoretical framework to describe the ladder systems. The N-chain Hubbard model has been studied within the Composite Operator Method. In this scheme of calculations the single-particle Green's function for any number of coupled chains is obtained by solving self-consistently a system of integral equations.
We propose a theoretical framework to describe the ladder systems. The N-chain Hubbard model has been studied within the Composite Operator Method. In this scheme of calculations the single-particle Green's function for any number of coupled chains is obtained by solving self-consistently a system of integral equations.
△ Less
Submitted 27 July, 1998;
originally announced July 1998.
