Despite the impressive capability of large language models (LLMs), knowing when to trust their generations remains an open challenge. The recent literature on uncertainty quantification of natural language generation (NLG) utilizes a conventional natural language inference (NLI) classifier to measure the semantic dispersion of LLMs responses. These studies employ logits of NLI classifier for semantic clustering to estimate uncertainty. However, logits represent the probability of the predicted class and barely contain feature information for potential clustering. Alternatively, CLIP (Contrastive Language-Image Pre-training) performs impressively in extracting image-text pair features and measuring their similarity. To extend its usability, we propose Contrastive Semantic Similarity, the CLIP-based feature extraction module to obtain similarity features for measuring uncertainty for text pairs. We apply this method to selective NLG, which detects and rejects unreliable generations for better trustworthiness of LLMs. We conduct extensive experiments with three LLMs on several benchmark question-answering datasets with comprehensive evaluation metrics. Results show that our proposed method performs better in estimating reliable responses of LLMs than comparable baselines. The code are available at https://github.com/AoShuang92/css_uq_llms.

We investigate the potential of a new citizen science paradigm that facilitates collaborative learning between humans and artificial intelligence (AI). Recognising the potential of AI to support and empower rather than replace human participation, we explore the integration of image recognition as a ‘dialogic AI partner’ in citizen science (CS) projects, interacting with participants in real time. We study this in the context of a biodiversity monitoring project that relies on volunteers to identify biological species from images taken in the wild. Guided by the idea of Bakhtin’s dialogism and Bayesian inference principles, we developed a web interface that integrated an image recognition model, fine-tuned for classifying 22 UK bumblebee species, into an interactive interface based on visual feature keys to enable real-time dialogue between humans and AI. We report a significant improvement in identification accuracy for both humans and AI when they engage in such dialogue and retain the ability to reach independent conclusions rather than achieve consensus. Given the inherent need for convergence in decision-making within scientific processes such as species identification tasks, we augmented the dialogic process with a Bayesian model that unifies potentially divergent human and AI perspectives post collaboration to achieve a more accurate consensus decision than that achieved by either AI or citizens. Our work provides new understandings around the design of a dialogic space for CS practice that effectively builds on the complementary strengths of human and AI visual recognition approaches.

Failure detection (FD) in AI systems is a crucial safeguard for the deployment for safety-critical tasks. The common evaluation method of FD performance is the Risk-coverage (RC) curve, which reveals the trade-off between the data coverage rate and the performance on accepted data. One common way to quantify the RC curve by calculating the area under the RC curve. However, this metric does not inform on how suited any method is for FD, or what the optimal coverage rate should be. As FD aims to achieve higher performance with fewer data discarded, evaluating with partial coverage excluding the most uncertain samples is more intuitive and meaningful than full coverage. In addition, there is an optimal point in the coverage where the model could achieve ideal performance theoretically. We propose the Excess Area Under the Optimal RC Curve (E-AUoptRC), with the area in coverage from the optimal point to the full coverage. Further, the model performance at this optimal point can represent both model learning ability and calibration. We propose it as the Trust Index (TI), a complementary evaluation metric to the overall model accuracy. We report extensive experiments on three benchmark image datasets with ten variants of transformer and CNN models. Our results show that our proposed methods can better reflect the model trustworthiness than existing evaluation metrics. We further observe that the model with high overall accuracy does not always yield the high TI, which indicates the necessity of the proposed Trust Index as a complementary metric to the model overall accuracy. The code are available at \url{https://github.com/AoShuang92/optimal_risk}.

Proper confidence calibration of deep neural networks is essential for reliable predictions in safety-critical tasks. Miscalibration can lead to model over-confidence and/or under-confidence; i.e., the model's confidence in its prediction can be greater or less than the model's accuracy. Recent studies have highlighted the over-confidence issue by introducing calibration techniques and demonstrated success on various tasks. However, miscalibration through under-confidence has not yet to receive much attention. In this paper, we address the necessity of paying attention to the under-confidence issue. We first introduce a novel metric, a miscalibration score, to identify the overall and class-wise calibration status, including being over or under-confident. Our proposed metric reveals the pitfalls of existing calibration techniques, where they often overly calibrate the model and worsen under-confident predictions. Then we utilize the class-wise miscalibration score as a proxy to design a calibration technique that can tackle both over and under-confidence. We report extensive experiments that show our proposed methods substantially outperforming existing calibration techniques. We also validate our proposed calibration technique on an automatic failure detection task with a risk-coverage curve, reporting that our methods improve failure detection as well as trustworthiness of the model. The code are available at \url{https://github.com/AoShuang92/miscalibration_TS}.

As the citizen science (CS) community flourishes, there is an opportunity to reflect on how practitioners can widen participation and work with participants as co-researchers to investigate and take action around global challenges. Through the lens of one CS case study, the X-Polli:Nation project, we report on how technologists, ecologists, and education specialists repurposed older projects by cross-pollinating ideas with children and teachers in the UK and in Italy to create Artificial Intelligence–enhanced tools appropriate for teaching sustainability in schools. Taking part in an actionable CS cycle, children learn about pollinating insects, record scientific data, create flowering habitats, and communicate their importance. Through this process, X-Polli:Nation demonstrates relevance across a number of Sustainable Development Goals (e.g., SDG 4, Quality Education; SDG 10, Reducing Inequality; and SDG 15, Life on Land), and applies the underlying SDG principle “leave no one behind.” We go on to investigate if, and how, young people would like to deepen their engagement with the SDGs, and we report that taking action and communicating the importance of the SDGs were of paramount interest. The challenge of building sustainability into an already crowded curriculum can be alleviated by understanding its value, considering the audience, and adapting to new contexts. The considerable benefits include raising awareness about global sustainability issues and giving children the confidence to become passionate environmental stewards, all the while extending the life of older projects and thus making CS methods sustainable too.

Despite the great success of state-of-the-art deep neural networks, several studies have reported models to be over-confident in predictions, indicating miscalibration. Label Smoothing has been proposed as a solution to the over-confidence problem and works by softening hard targets during training, typically by distributing part of the probability mass from a 'one-hot' label uniformly to all other labels. However, neither model nor human confidence in a label are likely to be uniformly distributed in this manner, with some labels more likely to be confused than others. In this paper we integrate notions of model confidence and human confidence with label smoothing, respectively Model Confidence LS and Human Confidence LS, to achieve better model calibration and generalization. To enhance model generalization, we show how our model and human confidence scores can be successfully applied to curriculum learning, a training strategy inspired by learning of 'easier to harder' tasks. A higher model or human confidence score indicates a more recognisable and therefore easier sample, and can therefore be used as a scoring function to rank samples in curriculum learning. We evaluate our proposed methods with four state-of-the-art architectures for image and text classification task, using datasets with multi-rater label annotations by humans. We report that integrating model or human confidence information in label smoothing and curriculum learning improves both model performance and model calibration. The code are available at https://github.com/AoShuang92/Confidence Calibration CL.

Commercial camera traps are usually triggered by a Passive Infra-Red (PIR) motion sensor necessitating a delay between triggering and the image being captured. This often seriously limits the ability to record images of small and fast moving animals. It also results in many “empty” images, e.g., owing to moving foliage against a background of different temperature. In this paper we detail a new triggering mechanism based solely on the camera sensor. This is intended for use by citizen scientists and for deployment on an affordable, compact, low-power Raspberry Pi computer (RPi). Our system introduces a video frame filtering pipeline consisting of movement and image-based processing. This makes use of Machine Learning (ML) feasible on a live camera stream on an RPi. We describe our free and open-source software implementation of the system; introduce a suitable ecology efficiency measure that mediates between specificity and recall; provide ground-truth for a video clip collection from camera traps; and evaluate the effectiveness of our system thoroughly. Overall, our video camera trap turns out to be robust and effective.

We argue that expert finding is sensitive to multiple document features in an organization, and therefore, can benefit from the incorporation of these document features. We propose a unified language model, which integrates multiple document features, namely, multiple levels of associations, PageRank, indegree, internal document structure, and URL length. Our experiments on two TREC Enterprise Track collections, i.e., the W3C and CSIRO datasets, demonstrate that the natures of the two organizational intranets and two types of expert finding tasks, i.e., key contact finding for CSIRO and knowledgeable person finding for W3C, influence the effectiveness of different document features. Our work provides insights into which document features work for certain types of expert finding tasks, and helps design expert finding strategies that are effective for different scenarios.

This paper outlines a system for searching and browsing 1.14 million images from the World Wide Web (WWW) based on their visual content. At the heart of the system lies an automatically constructed network of images that can be navigated quickly by following its edges. The browsing experience is enhanced in a number of ways including multidimensional scaling of the graph neighbourhood for display purposes, Markov clustering of the image network to provide summaries of its content, and automated annotation of the images to allow users to access the network through text queries.

We study the problem of detecting sentences describing adverse drug reactions (ADRs) and frame the problem as binary classification. We investigate different neural network (NN) architectures for ADR classification. In particular, we propose two new neural network models, Convolutional Recurrent Neural Network (CRNN) by concatenating convolutional neural networks with recurrent neural networks, and Convolutional Neural Network with Attention (CNNA) by adding attention weights into convolutional neural networks. We evaluate various NN architectures on a Twitter dataset containing informal language and an Adverse Drug Effects (ADE) dataset constructed by sampling from MEDLINE case reports. Experimental results show that all the NN architectures outperform the traditional maximum entropy classifiers trained from n-grams with different weighting strategies considerably on both datasets. On the Twitter dataset, all the NN architectures perform similarly. But on the ADE dataset, CNN performs better than other more complex CNN variants. Nevertheless, CNNA allows the visualisation of attention weights of words when making classification decisions and hence is more appropriate for the extraction of word subsequences describing ADRs.

We study the problem of detecting sentences describing adverse drug reactions (ADRs) and frame the problem as binary classification. We investigate different neural network (NN) architectures for ADR classification. In particular, we propose two new neural network models, Convolutional Recurrent Neural Network (CRNN) by concatenating convolutional neural networks with recurrent neural networks, and Convolutional Neural Network with Attention (CNNA) by adding attention weights into convolutional neural networks. We evaluate various NN architectures on a Twitter dataset containing informal language and an Adverse Drug Effects (ADE) dataset constructed by sampling from MEDLINE case reports. Experimental results show that all the NN architectures outperform the traditional maximum entropy classifiers trained from n-grams with different weighting strategies considerably on both datasets. On the Twitter dataset, all the NN architectures perform similarly. But on the ADE dataset, CNN performs better than other more complex CNN variants. Nevertheless, CNNA allows the visualisation of attention weights of words when making classification decisions and hence is more appropriate for the extraction of word subsequences describing ADRs.

In this paper we formulate image retrieval by text query as a vector space classification problem. This is achieved by creating a high-dimensional visual vocabulary that represents the image documents in great detail. We show how the representation of these image documents enables the application of well known text retrieval techniques such as Rocchio tf-idf and naïve Bayes to the semantic image retrieval problem. We tested these methods on a Corel images subset and achieve state-of-the-art retrieval performance using the proposed methods.

ICMR was initially started as a workshop on challenges in image retrieval (in Newcastle in 1998 ) and later transformed into the Conference on Image and Video Retrieval (CIVR) series. In 2011 the CIVR and the ACM Workshop on Multimedia Information Retrieval were combined into a single conference that now forms the ICMR series. The 4th ACM International Conference on Multimedia Retrieval took place in Glasgow, Scotland, from 1 – 4 April 2014. This was the largest edition of ICMR to date with approximately 170 attendees from 25 different countries. ICMR is one of the premier scientific conference for multimedia retrieval held worldwide, with the stated mission “to illuminate the state of the art in multimedia retrieval by bringing together researchers and practitioners in the field of multimedia retrieval .” According to the Chinese Computing Federation Conference Ranking (2013), ACM ICMR is the number one multimedia retrieval conference worldwide and the number four conference in the category of multimedia and graphics. Although ICMR is about multimedia retrieval, in a wider sense, it is also about automated multimedia understanding. Much of the work in that area involves the analysis of media on a pixel, voxel, and wavelet level, but it also involves innovative retrieval, visualisation and interaction paradigms utilising the nature of the multimedia — be it video, images, speech, or more abstract (sensor) data. The conference aims to promote intellectual exchanges and interactions among scientists, engineers, students, and multimedia researchers in academia as well as industry through various events, including a keynote talk, oral, special and poster sessions focused on re search challenges and solutions, technical and industrial demonstrations of prototypes, tutorials, research, and an industrial panel. In the remainder of this report we will summarise the events that took place at the 4th ACM ICMR conference.

In recent years, learning word vector representations has attracted much interest in Natural Language Processing. Word representations or embeddings learned using unsupervised methods help addressing the problem of traditional bag-of-word approaches which fail to capture contextual semantics. In this paper we go beyond the vector representations at the word level and propose a novel framework that learns higher-level feature representations of n-grams, phrases and sentences using a deep neural network built from stacked Convolutional Restricted Boltzmann Machines (CRBMs). These representations have been shown to map syntactically and semantically related n-grams to closeby locations in the hidden feature space. We have experimented to additionally incorporate these higher-level features into supervised classifier training for two sentiment analysis tasks: subjectivity classification and sentiment classification. Our results have demonstrated the success of our proposed framework with 4

This article provides a recap of the 4th ACM International Conference on Multimedia Retrieval, which took place in Glasgow, Scotland, from 1-4 April 2014.

We present an analysis of existing methods to automatic classification of photos according to aesthetics. We review different components of the classification process: existing evaluation datasets, their properties, most commonly-used image features, qualitative and quantitative, and classification results where comparable. We argue there are methodology gaps in the existing approaches to evaluating the classification results. We introduce the results of our experiments with Random Forest classification applied to image aesthetics classification and compare them to AdaBoost and SVM approaches.