More Like this

1. Compos Mentis at SemEval2024 Task6: A Multi-Faceted Role-based Large Language Model Ensemble to ; Detect Hallucination

   Das, Souvik; Srihari, Rohini K (June 2024, ACL Anthology)

   Hallucinations in large language models (LLMs), where they generate fluent but factually incorrect outputs, pose challenges for applications requiring strict truthfulness. This work proposes a multi-faceted approach to detect such hallucinations across various language tasks. We leverage automatic data annotation using a proprietary LLM, fine-tuning of the Mistral-7B-instruct-v0.2 model on annotated and benchmark data, role-based and rationale-based prompting strategies, and an ensemble method combining different model outputs through majority voting. This comprehensive framework aims to improve the robustness and reliability of hallucination detection for LLM generations. Code and data1 1 Introduction The modern natural language generation (NLG) (OpenAI et al., 2023; Touvron et al., 2023) landscape faces two interconnected challenges: firstly, current neural models have a tendency to produce f luent yet inaccurate outputs, and secondly, our evaluation metrics are better suited for assessing f luency rather than correctness(Bang et al., 2023; Guerreiro et al., 2023). This phenomenon, known as "hallucination," (Ji et al., 2023) where neural networks generate plausible-sounding but factually incorrect outputs, is a significant hurdle, especially for NLG applications that require strict adherence to correctness. For instance, in machine translation(Lee et al., 2019), producing a fluent translation that deviates from the source text’s meaning renders the entire translation pipeline unreliable. This issue may arise as LLMs are trained on vast amounts of data from the internet, which can contain inaccuracies, biases, and false information. Also, it may arise due improper representations learned during training even if good quality data is 1https://github.com/souvikdgp16/shroom_compos_mentis used. As a result, LLMs can sometimes hallucinate or fabricate details, especially when prompted to discuss topics outside their training data or make inferences beyond their capabilities. Hallucination detection (Liu et al., 2022), also known as factual verification or truthfulness evaluation, identifies and mitigates these hallucinations in the outputs of LLMs. This is an active area of research and development, as it is crucial for ensuring the reliability and trustworthiness of LLMgenerated content, particularly in high-stakes domains such as healthcare, finance, and legal applications. In this task, the primary focus will be to classify whether a generation is hallucinated. This work proposes a multi-faceted approach to detecting hallucinations in large language models. 

   more » « less
2. Low-Depth Algebraic Circuit Lower Bounds over Any Field

   https://doi.org/10.4230/LIPIcs.CCC.2024.31  

   Forbes, Michael A (July 2024, Schloss Dagstuhl – Leibniz-Zentrum für Informatik)

   Santhanam, Rahul (Ed.)

   The recent breakthrough of Limaye, Srinivasan and Tavenas [Limaye et al., 2022] (LST) gave the first super-polynomial lower bounds against low-depth algebraic circuits, for any field of zero (or sufficiently large) characteristic. It was an open question to extend this result to small-characteristic ([Limaye et al., 2022; Govindasamy et al., 2022; Fournier et al., 2023]), which in particular is relevant for an approach to prove superpolynomial AC⁰[p]-Frege lower bounds ([Govindasamy et al., 2022]). In this work, we prove super-polynomial algebraic circuit lower bounds against low-depth algebraic circuits over any field, with the same parameters as LST (or even matching the improved parameters of Bhargav, Dutta, and Saxena [Bhargav et al., 2022]). We give two proofs. The first is logical, showing that even though the proof of LST naively fails in small characteristic, the proof is sufficiently algebraic that generic transfer results imply the result over characteristic zero implies the result over all fields. Motivated by this indirect proof, we then proceed to give a second constructive proof, replacing the field-dependent set-multilinearization result of LST with a set-multilinearization that works over any field, by using the Binet-Minc identity [Minc, 1979]. 

   more » « less
3. Enhancing Real Hurricane Forecasts by Rotation-based Adjustments of Vertical Diffusion Parameterizations in the Planetary Boundary Layer

   Khondaker, MMH; Momen, Mostafa (January 2025, American Meteorological Society)

   Forecasting hurricanes is critically important for mitigating their devastating impacts caused by wind damage, storm surges, and flooding. Despite remarkable advancements in numerical weather prediction (NWP) models, such as the Weather Research and Forecasting (WRF) model, accurate hurricane forecasts remain challenging likely due to inaccurate physical parameterizations of complex dynamics of these storms. One major issue of these models is related to their Planetary Boundary Layer (PBL) schemes, which are not typically designed for hurricane flows with strong rotation. Previous studies have shown that the existing PBL schemes of hurricane simulations are often overly dissipative, leading to underestimations of the storm intensity (Matak and Momen 2023; Romdhani et al. 2022). Our recent research (Khondaker and Momen 2024) demonstrated that reducing diffusion in these models improved the hurricane’s intensity and size forecasts by more than ~30% on average in four considered major hurricanes. This reduced diffusion is due to the strong rotational nature of hurricanes, which suppresses turbulence and produces smaller eddies compared to regular PBLs (Momen et al. 2021). While prior studies showed that decreasing the vertical diffusion significantly improves major hurricane intensity forecasts, they mostly relied on simplified and often invariable adjustments of vertical diffusion such as multiplying it by a constant coefficient. The objective of this study is to address this issue by introducing a rotation-based variable adjustment of diffusion to account for the strong rotational nature of tropical cyclone (TC) dynamics. To this end, we will present multiple real strong and weak hurricane simulations using the Advanced Research WRF (ARW) model in the US. We modified the vertical eddy diffusivity based on the relative vorticity to accommodate the rotational dynamics of TCs in PBL schemes. While the default model significantly underpredicts hurricane intensity, our new adjustments outperform the default schemes for these strong hurricanes (see, e.g., attached fig. a), with notable improvements in track and minimum sea level pressure accuracy. This modification also remarkably increases the inflow in hurricanes compared to default models and leads to the intensification of the TC vortex (see, e.g., attached fig. b,c). Our newly adjusted model matched more closely with dropsonde, and satellite observations compared to the default WRF’s PBL schemes. These modifications to the PBL schemes make them more physics-based adjustments compared to previous treatments, offering valuable insights for improving hurricane forecasts in NWP models. References: Khondaker, M. H., and M. Momen, 2024: Improving hurricane intensity and streamflow forecasts in coupled hydro-meteorological simulations by analyzing precipitation and boundary layer schemes. J Hydrometeorol, https://doi.org/10.1175/JHM-D-23-0153.1. Matak, L., and M. Momen, 2023: The Role of Vertical Diffusion Parameterizations in the Dynamics and Accuracy of Simulated Intensifying Hurricanes. Boundary Layer Meteorology, https://doi.org/10.1007/s10546-023-00818-w. Momen, M., M. B. Parlange, and M. G. Giometto, 2021: Scrambling and Reorientation of Classical Atmospheric Boundary Layer Turbulence in Hurricane Winds. Geophysical Research Letters, 48, https://doi.org/10.1029/2020GL091695. Romdhani, O., J. A. Zhang, and M. Momen, 2022: Characterizing the Impacts of Turbulence Closures on Real Hurricane Forecasts: A Comprehensive Joint Assessment of Grid Resolution, Horizontal Turbulence Models, and Horizontal Mixing Length. Journal of Advanced Modeling Earth System, 14, https://doi.org/10.1029/2021ms002796. 

   more » « less
4. On Image Quality Metadata, FAIR in ML, AI-Readiness and Reproducibility: Fish-AIR example

   https://doi.org/10.3897/biss.7.112178  

   Bakış, Yasin; Wang, Xiaojun; Altıntaş, Bahadır; Jebbia, Dom; Bart_Jr, Henry (September 2023, Biodiversity Information Science and Standards)

   A new science discipline has emerged within the last decade at the intersection of informatics, computer science and biology:Imageomics. Like most other -omics fields, Imageomics also uses emerging technologies to analyze biological data but from the images. One of the most applied data analysis methods for image datasets is Machine Learning (ML). In 2019, we started working on a United States National Science Foundation (NSF) funded project, known as Biology Guided Neural Networks (BGNN) with the purpose of extracting information about biology by using neural networks and biological guidance such as species descriptions, identifications, phylogenetic trees and morphological annotations (Bart et al. 2021). Even though the variety and abundance of biological data is satisfactory for some ML analysis and the data are openly accessible, researchers still spend up to 80% of their time preparing data into a usable, AI-ready format, leaving only 20% for exploration and modeling (Long and Romanoff 2023). For this reason, we have built a dataset composed of digitized fish specimens, taken either directly from collections or from specialized repositories. The range of digital representations we cover is broad and growing, from photographs and radiographs, to CT scans, and even illustrations. We have added new groups of vocabularies to the dataset management system including image quality metadata, extended image metadata and batch metadata. With the image quality metadata and extended image metadata, we aimed to extract information from the digital objects that can possibly help ML scientists in their research with filtering, image processing and object recognition routines. Image quality metadata provides information about objects contained in the image, features and condition of the specimen, and some basic visual properties of the image, while extended image metadata provides information about technical properties of the digital file and the digital multimedia object (Bakış et al. 2021, Karnani et al. 2022, Leipzig et al. 2021, Pepper et al. 2021, Wang et al. 2021) (see details on Fish-AIR vocabulary web page). Batch metadata is used for separating different datasets and facilitates downloading and uploading data in batches with additional batch information and supplementary files. Additional flexibility, built into the database infrastructure using an RDF framework, will enable the system to host different taxonomic groups, which might require new metadata features (Jebbia et al. 2023). By the combination of these features, along with FAIR (Findable, Accessable, Interoperable, Reusable) principles, and reproducibility, we provide Artificial Intelligence Readiness (AIR; Long and Romanoff 2023) to the dataset. Fish-AIR provides an easy-to-access, filtered, annotated and cleaned biological dataset for researchers from different backgrounds and facilitates the integration of biological knowledge based on digitized preserved specimens into ML pipelines. Because of the flexible database infrastructure and addition of new datasets, researchers will also be able to access additional types of data—such as landmarks, specimen outlines, annotated parts, and quality scores—in the near future. Already, the dataset is the largest and most detailed AI-ready fish image dataset with integrated Image Quality Management System (Jebbia et al. 2023, Wang et al. 2021). 

   more » « less
5. Optimal Eye Surgeon: Finding image priors through sparse generators at initialization

   Ghosh, Avrajit; Zhang, Xitong; Sun, Kenneth K; Qu, Qing; Ravishankar, Saiprasad; Wang, Rongrong (June 2024, International Conference on Machine Learning)

   We introduce Optimal Eye Surgeon (OES), a framework for pruning and training deep image generator networks. Typically, untrained deep convolutional networks, which include image sampling operations, serve as effective image priors (Ulyanov et al., 2018). However, they tend to overfit to noise in image restoration tasks due to being overparameterized. OES addresses this by adaptively pruning networks at random initialization to a level of underparameterization. This process effectively captures low-frequency image components even without training, by just masking. When trained to fit noisy images, these pruned subnetworks, which we term Sparse-DIP, resist overfitting to noise. This benefit arises from underparameterization and the regularization effect of masking, constraining them in the manifold of image priors. We demonstrate that subnetworks pruned through OES surpass other leading pruning methods, such as the Lottery Ticket Hypothesis, which is known to be suboptimal for image recovery tasks (Wu et al., 2023). Our extensive experiments demonstrate the transferability of OES-masks and the characteristics of sparse-subnetworks for image generation. 

   more » « less