Ankan Mullick, Sombit Bose, Rounak Saha, Ayan Kumar Bhowmick, Aditya Vempaty, Prasenjit Dey, Ravi Kokku, Pawan Goyal, Niloy Ganguly. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

The United Nations Development Programme (UNDP) helps countries implement the United Nations (UN) Sustainable Development Goals (SDGs), an agenda for tackling major societal issues such as poverty, hunger, and environmental degradation by the year 2030. A key service provided by UNDP to countries that seek it is a review of national development plans and sector strategies by policy experts to assess alignment of national targets with one or more of the 169 targets of the 17 SDGs. Known as the Rapid Integrated Assessment (RIA), this process involves manual review of hundreds, if not thousands, of pages of documents and takes weeks to complete. In this work, we develop a natural language processing-based methodology to accelerate the workflow of policy experts. Specifically we use paragraph embedding techniques to find paragraphs in the documents that match the semantic concepts of each of the SDG targets. One novel technical contribution of our work is in our use of historical RIAs from other countries as a form of neighborhood-based supervision for matches in the country under study. We have successfully piloted the algorithm to perform the RIA for Papua New Guinea’s national plan, with the UNDP estimating it will help reduce their completion time from an estimated 3-4 weeks to 3 days.

The emerging paradigm of Human-Machine Inference Networks (HuMaINs) combines complementary cognitive strengths of humans and machines in an intelligent manner to tackle various inference tasks and achieves higher performance than either humans or machines by themselves. While inference performance optimization techniques for human-only or sensor-only networks are quite mature, HuMaINs require novel signal processing and machine learning solutions. In this paper, we present an overview of the HuMaINs architecture with a focus on three main issues that include architecture design, inference algorithms including security/privacy challenges, and application areas/use cases.

In this paper, an overview of recent developments in received-signal-strength (RSS)-based localization in wireless sensor networks is presented. Several important practical issues and their solutions are discussed. To save communication bandwidth and sensor energy, a maximum-likelihood estimator based on quantized data is presented along with its corresponding Cramér-Rao lower bound (CRLB) and optimal quantizer design schemes. For further system resource savings, an iterative sensor selection approach is presented to activate only the most informative sensors, by maximizing the mutual information or minimizing the posterior CRLB at each iteration. For a resource constrained WSN with imperfect wireless channels, channel-aware target localization is described, where the channel model is incorporated into the localization scheme itself, thereby improving performance without increasing communication overhead. Another practical issue involving the presence of malicious sensors called Byzantines is discussed and mitigation schemes are provided. A recent coding-theorybased approach which is both computationally inexpensive and robust to such malicious attacks is also discussed.

Present day distributed inference systems consist of sensors with different modalities working as a system to perform specific tasks. With multiple sensors sensing heterogeneous data over multiple time instants, diversity is an inherent aspect of such systems. In this work, we take the first step to characterize the diversity of a general heterogeneous sensing system performing inference tasks. We provide a unified definition for diversity which can be customized for the system in use. The use of the definition is illustrated by applying it to a specific detection system where the sensors collect data over heterogeneous sensing channels. We assume the data to be both temporally and spatially correlated and analyze the effect of dependence on the diversity of the detection system.

In this work, we use crowds for acoustic classification of animal species in supervised and unsupervised manners. We demonstrate the effectiveness of the proposed triplet based crowdsourcing systems via actual experiments. Moreover, we propose a generalized 1-bit RPCA algorithm to further improve classification performance. The unique marriage of crowdsourcing and generalized 1-bit RPCA algorithm is shown to yield excellent performance for acoustic data classification.

Automatic forecasting of time series data is a challenging problem in many industries. Current forecast models adopted by businesses do not provide adequate means for including data representing external factors that may have a significant impact on the time series, such as weather, national events, local events, social media trends, promotions, etc. This paper introduces a novel neural network attention mechanism that naturally incorporates data from multiple external sources without the feature engineering needed to get other techniques to work. We demonstrate empirically that the proposed model achieves superior performance for predicting the demand of 20 commodities across 107 stores of one of America’s largest retailers when compared to other baseline models, including neural networks, linear models, certain kernel methods, Bayesian regression, and decision trees. Our method ultimately accounts for a 23.9% relative improvement as a result of the incorporation of external data sources, and provides an unprecedented level of descriptive ability for a neural network forecasting model.

We consider the CEO problem for belief sharing. Multiple subordinates observe independently corrupted versions of uniformly distributed data and transmit coded versions over rate-limited links to a CEO who then estimates the underlying data. Agents are not allowed to convene before transmitting their observations. This formulation is motivated by the practical problem of a firm's CEO estimating uniformly distributed beliefs about a sequence of events, before acting on them. Agents' observations are modeled as jointly distributed with the underlying data through a given conditional probability density function. We study the asymptotic behavior of the minimum achievable mean squared error distortion at the CEO in the limit when the number of agents L and the sum rate R tend to infinity. We establish a 1/R2convergence of the distortion, an intermediate regime of performance between the exponential behavior in discrete CEO problems [Berger, Zhang, and Viswanathan (1996)], and the 1/R behavior in Gaussian CEO problems [Viswanathan and Berger (1997)]. Achievability is proved by a layered architecture with scalar quantization, distributed entropy coding, and midrange estimation. The converse is proved using the Bayesian Chazan-Zakai-Ziv bound.

We consider the CEO problem for non-regular source distributions (such as uniform or truncated Gaussian). A group of agents observe independently corrupted versions of data and transmit coded versions over rate-limited links to a CEO. The CEO then estimates the underlying data based on the received coded observations. Agents are not allowed to convene before transmitting their observations. This formulation is motivated by the practical problem of a firm's CEO estimating (non-regular) beliefs about a sequence of events, before acting on them. Agents' observations are modeled as jointly distributed with the underlying data through a given conditional probability density function. We study the asymptotic behavior of the minimum achievable mean squared error distortion at the CEO in the limit when the number of agents L and the sum rate R tend to infinity. We establish a 1/R2convergence of the distortion, an intermediate regime of performance between the exponential behavior in discrete CEO problems [Berger, Zhang, and Viswanathan (1996)], and the 1/R behavior in Gaussian CEO problems [Viswanathan and Berger (1997)]. Achievability is proved by a layered architecture with scalar quantization, distributed entropy coding, and midrange estimation. The converse is proved using the Bayesian Chazan-Zakai-Ziv bound.

In this letter, a new sparsity-promoting penalty function is introduced for sensor selection problems in field reconstruction, which has the property of avoiding scenarios where the same sensors are successively selected. Using a reweighted ℓ1relaxation of the ℓ0norm, the sensor selection problem is reformulated as a convex quadratic program. In order to handle large-scale problems, we also present two fast algorithms: accelerated proximal gradient method and alternating direction method of multipliers. Numerical results are provided to demonstrate the effectiveness of our approaches.

In this paper, we consider the task of target localization using quantized data in wireless sensor networks. We propose a computationally efficient localization scheme by modeling it as an iterative classification problem. We design coding theory based iterative approaches for target localization where at every iteration, the fusion center (FC) solves an M-ary hypothesis testing problem and decides the region of interest for the next iteration. The coding theory based iterative approach works well even in the presence of Byzantine (malicious) sensors in the network. We further consider the effect of non-ideal channels. We suggest the use of soft-decision decoding to compensate for the loss due to the presence of fading channels between the local sensors and FC. We evaluate the performance of the proposed schemes in terms of the Byzantine fault tolerance capability and probability of detection of the target region. We also present performance bounds, which help us in designing the system. We provide asymptotic analysis of the proposed schemes and show that the schemes achieve perfect region detection irrespective of the noise variance when the number of sensors tends to infinity. Our numerical results show that the proposed schemes provide a similar performance in terms of mean square error as compared with the traditional maximum likelihood estimation but are computationally much more efficient and are resilient to errors due to Byzantines and non-ideal channels.

In this paper, we consider the problem of quantizer design for distributed estimation under the Bayesian criterion. We derive general optimality conditions under the assumption of conditionally independent observations at the local sensors and show that for a conditionally unbiased and efficient estimator at the Fusion Center, identical quantizers are optimal when local observations have identical distributions. This results in an N-fold reduction in complexity where N is the number of sensors. We illustrate our approach by applying it to the location parameter estimation problem.

We consider the problem of target localization using quantized data in Wireless Sensor Networks in the presence of Byzantines (malicious sensors). Since the effect of Byzantines can be treated as errors in the transmitted data, we propose the use of error correcting codes for the task of target localization. We design coding based iterative schemes for target localization where, at every iteration, the Fusion Center performs an M-ary hypothesis test and decides the Region of Interest for the next iteration. Simulation results show that our proposed schemes provide a better performance as compared to the traditional Maximum Likelihood Estimation and are also computationally much more efficient.

We consider the use of error-control codes and decoding algorithms to perform reliable classification using unreliable and anonymous human crowd workers by adapting coding-theoretic techniques for the specific crowdsourcing application. We develop an ordering principle for the quality of crowds and describe how system perfor-mance changes with the quality of the crowd. We demonstrate the effectiveness of the proposed coding scheme using both simulated data and real datasets from Amazon Mechanical Turk, a crowd-sourcing microtask platform. Results suggest that good codes may improve the performance of the crowdsourcing task over typical majority-vote approaches. Index Terms — crowdsourcing, classification, error-control codes

This paper considers the problem of Byzantine attacks on cooperative spectrum sensing in cognitive radio networks. Our major contribution is a technique to learn about the cognitive radio (CR) potential malicious behavior over time and thereby identifies the Byzantines and then estimates their probabilities of false alarm (Pfa) and detection (PD). We show that for a given set of data over time, the Byzantines can be identified for any a (percentage of Byzantines). It has also been shown that these estimates of Pfaand Pn of the Byzantines are asymptotically unbiased and converge to their true values at the rate of O(T-1/2). We then use these probabilities to adaptively design the fusion rule. We calculate the Probability of error (Qe) and compare it with the minimum probability of error possible.