Susan McKeever

Muzhaffar Hazman, Susan McKeever, Josephine Griffith

Internet memes are characterised by the interspersing of text amongst visual elements. State-of-the-art multimodal meme classifiers do not account for the relative positions of these elements across the two modalities, despite the latent meaning associated with where text and visual elements are placed. Against two meme sentiment classification datasets, we systematically show performance gains from incorporating the spatial position of visual objects, faces, and text clusters extracted from memes. In addition, we also present facial embedding as an impactful enhancement to image representation in a multimodal meme classifier. Finally, we show that incorporating this spatial information allows our fully automated approaches to outperform their corresponding baselines that rely on additional human validation of OCR-extracted text.

Muzhaffar Hazman, Susan McKeever, Josephine Griffith

Internet Memes remain a challenging form of user-generated content for automated sentiment classification. The availability of labelled memes is a barrier to developing sentiment classifiers of multimodal memes. To address the shortage of labelled memes, we propose to supplement the training of a multimodal meme classifier with unimodal (image-only and text-only) data. In this work, we present a novel variant of supervised intermediate training that uses relatively abundant sentiment-labelled unimodal data. Our results show a statistically significant performance improvement from the incorporation of unimodal text data. Furthermore, we show that the training set of labelled memes can be reduced by 40% without reducing the performance of the downstream model.

Muzhaffar Hazman, Susan McKeever, Josephine Griffith

Warning: This paper contains memes that may be offensive to some readers. Multimodal Internet Memes are now a ubiquitous fixture in online discourse. One strand of meme-based research is the classification of memes according to various affects, such as sentiment and hate, supported by manually compiled meme datasets. Understanding the unique characteristics of memes is crucial for meme classification. Unlike other user-generated content, memes spread via memetics, i.e. the process by which memes are imitated and transformed into symbols used to create new memes. In effect, there exists an ever-evolving pool of visual and linguistic symbols that underpin meme culture and are crucial to interpreting the meaning of individual memes. The current approach of training supervised learning models on static datasets, without taking memetics into account, limits the depth and accuracy of meme interpretation. We argue that meme datasets must contain genuine memes, as defined via memetics, so that effective meme classifiers can be built. In this work, we develop a meme identification protocol which distinguishes meme from non-memetic content by recognising the memetics within it. We apply our protocol to random samplings of the leading 7 meme classification datasets and observe that more than half (50. 4\%) of the evaluated samples were found to contain no signs of memetics. Our work also provides a meme typology grounded in memetics, providing the basis for more effective approaches to the interpretation of memes and the creation of meme datasets.

Muzhaffar Hazman, Susan McKeever, Josephine Griffith

Internet memes, now a staple of digital communication, play a pivotal role in how users engage within online communities and allow researchers to gain insight into contemporary digital culture. These engaging user-generated content are characterised by their reuse of visual elements also found in other memes. Matching instances of memes via these shared visual elements, called Meme Matching, is the basis of a wealth of meme analysis approaches. However, most existing methods assume that every meme consists of a shared visual background, called a Template, with some overlaid text, thereby limiting meme matching to comparing the background image alone. Current approaches exclude the many memes that are not template-based and limit the effectiveness of automated meme analysis and would not be effective at linking memes to contemporary web-based meme dictionaries. In this work, we introduce a broader formulation of meme matching that extends beyond template matching. We show that conventional similarity measures, including a novel segment-wise computation of the similarity measures, excel at matching template-based memes but fall short when applied to non-template-based meme formats. However, the segment-wise approach was found to consistently outperform the whole-image measures on matching non-template-based memes. Finally, we explore a prompting-based approach using a pretrained Multimodal Large Language Model for meme matching. Our results highlight that accurately matching memes via shared visual elements, not just background templates, remains an open challenge that requires more sophisticated matching techniques.

Allassan Tchangmena A Nken, Susan Mckeever, Peter Corcoran et al.

Considerable effort has been made in privacy-preserving video human activity recognition (HAR). Two primary approaches to ensure privacy preservation in Video HAR are differential privacy (DP) and visual privacy. Techniques enforcing DP during training provide strong theoretical privacy guarantees but offer limited capabilities for visual privacy assessment. Conversely methods, such as low-resolution transformations, data obfuscation and adversarial networks, emphasize visual privacy but lack clear theoretical privacy assurances. In this work, we focus on two main objectives: (1) leveraging DP properties to develop a model-free approach for visual privacy in videos and (2) evaluating our proposed technique using both differential privacy and visual privacy assessments on HAR tasks. To achieve goal (1), we introduce Video-DPRP: a Video-sample-wise Differentially Private Random Projection framework for privacy-preserved video reconstruction for HAR. By using random projections, noise matrices and right singular vectors derived from the singular value decomposition of videos, Video-DPRP reconstructs DP videos using privacy parameters ($ε,δ$) while enabling visual privacy assessment. For goal (2), using UCF101 and HMDB51 datasets, we compare Video-DPRP's performance on activity recognition with traditional DP methods, and state-of-the-art (SOTA) visual privacy-preserving techniques. Additionally, we assess its effectiveness in preserving privacy-related attributes such as facial features, gender, and skin color, using the PA-HMDB and VISPR datasets. Video-DPRP combines privacy-preservation from both a DP and visual privacy perspective unlike SOTA methods that typically address only one of these aspects.

Xuehao Liu, Sarah Jane Delany, Susan McKeever

Deep learning models suffer from opaqueness. For Convolutional Neural Networks (CNNs), current research strategies for explaining models focus on the target classes within the associated training dataset. As a result, the understanding of hidden feature map activations is limited by the discriminative knowledge gleaned during training. The aim of our work is to explain and expand CNNs models via the mirroring or alignment of CNN to an external knowledge base. This will allow us to give a semantic context or label for each visual feature. We can match CNN feature activations to nodes in our external knowledge base. This supports knowledge-based interpretation of the features associated with model decisions. To demonstrate our approach, we build two separate graphs. We use an entity alignment method to align the feature nodes in a CNN with the nodes in a ConceptNet based knowledge graph. We then measure the proximity of CNN graph nodes to semantically meaningful knowledge base nodes. Our results show that in the aligned embedding space, nodes from the knowledge graph are close to the CNN feature nodes that have similar meanings, indicating that nodes from an external knowledge base can act as explanatory semantic references for features in the model. We analyse a variety of graph building methods in order to improve the results from our embedding space. We further demonstrate that by using hierarchical relationships from our external knowledge base, we can locate new unseen classes outside the CNN training set in our embeddings space, based on visual feature activations. This suggests that we can adapt our approach to identify unseen classes based on CNN feature activations. Our demonstrated approach of aligning a CNN with an external knowledge base paves the way to reason about and beyond the trained model, with future adaptations to explainable models and zero-shot learning.

Trevor Doherty, Susan McKeever, Nebras Al-Attar et al.

The diagnosis of prostate cancer is challenging due to the heterogeneity of its presentations, leading to the over diagnosis and treatment of non-clinically important disease. Accurate diagnosis can directly benefit a patient's quality of life and prognosis. Towards addressing this issue, we present a learning model for the automatic identification of prostate cancer. While many prostate cancer studies have adopted Raman spectroscopy approaches, none have utilised the combination of Raman Chemical Imaging (RCI) and other imaging modalities. This study uses multimodal images formed from stained Digital Histopathology (DP) and unstained RCI. The approach was developed and tested on a set of 178 clinical samples from 32 patients, containing a range of non-cancerous, Gleason grade 3 (G3) and grade 4 (G4) tissue microarray samples. For each histological sample, there is a pathologist labelled DP - RCI image pair. The hypothesis tested was whether multimodal image models can outperform single modality baseline models in terms of diagnostic accuracy. Binary non-cancer/cancer models and the more challenging G3/G4 differentiation were investigated. Regarding G3/G4 classification, the multimodal approach achieved a sensitivity of 73.8% and specificity of 88.1% while the baseline DP model showed a sensitivity and specificity of 54.1% and 84.7% respectively. The multimodal approach demonstrated a statistically significant 12.7% AUC advantage over the baseline with a value of 85.8% compared to 73.1%, also outperforming models based solely on RCI and median Raman spectra. Feature fusion of DP and RCI does not improve the more trivial task of tumour identification but does deliver an observed advantage in G3/G4 discrimination. Building on these promising findings, future work could include the acquisition of larger datasets for enhanced model generalization.

hsan Ullah, Andre Rios, Vaibhav Gala et al.

Trust and credibility in machine learning models is bolstered by the ability of a model to explain itsdecisions. While explainability of deep learning models is a well-known challenge, a further chal-lenge is clarity of the explanation itself, which must be interpreted by downstream users. Layer-wiseRelevance Propagation (LRP), an established explainability technique developed for deep models incomputer vision, provides intuitive human-readable heat maps of input images. We present the novelapplication of LRP for the first time with structured datasets using a deep neural network (1D-CNN),for Credit Card Fraud detection and Telecom Customer Churn prediction datasets. We show how LRPis more effective than traditional explainability concepts of Local Interpretable Model-agnostic Ex-planations (LIME) and Shapley Additive Explanations (SHAP) for explainability. This effectivenessis both local to a sample level and holistic over the whole testing set. We also discuss the significantcomputational time advantage of LRP (1-2s) over LIME (22s) and SHAP (108s), and thus its poten-tial for real time application scenarios. In addition, our validation of LRP has highlighted features forenhancing model performance, thus opening up a new area of research of using XAI as an approachfor feature subset selection