Ignacio Castro

Peyman Hosseini, Mehran Hosseini, Sana Sabah Al-Azzawi et al.

We study the influence of different activation functions in the output layer of deep neural network models for soft and hard label prediction in the learning with disagreement task. In this task, the goal is to quantify the amount of disagreement via predicting soft labels. To predict the soft labels, we use BERT-based preprocessors and encoders and vary the activation function used in the output layer, while keeping other parameters constant. The soft labels are then used for the hard label prediction. The activation functions considered are sigmoid as well as a step-function that is added to the model post-training and a sinusoidal activation function, which is introduced for the first time in this paper.

Waleed Iqbal, Vahid Ghafouri, Gareth Tyson et al.

From health to education, income impacts a huge range of life choices. Earlier research has leveraged data from online social networks to study precisely this impact. In this paper, we ask the opposite question: do different levels of income result in different online behaviors? We demonstrate it does. We present the first large-scale study of Nextdoor, a popular location-based social network. We collect 2.6 Million posts from 64,283 neighborhoods in the United States and 3,325 neighborhoods in the United Kingdom, to examine whether online discourse reflects the income and income inequality of a neighborhood. We show that posts from neighborhoods with different incomes indeed differ, e.g. richer neighborhoods have a more positive sentiment and discuss crimes more, even though their actual crime rates are much lower. We then show that user-generated content can predict both income and inequality. We train multiple machine learning models and predict both income (R-squared=0.841) and inequality (R-squared=0.77).

Peyman Hosseini, Ignacio Castro, Iacopo Ghinassi et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in comprehending and analyzing lengthy sequential inputs, owing to their extensive context windows that allow processing millions of tokens in a single forward pass. However, this paper uncovers a surprising limitation: LLMs fall short when handling long input sequences. We investigate this issue using three datasets and two tasks (sentiment analysis and news categorization) across various LLMs, including Claude 3, Gemini Pro, GPT 3.5 Turbo, Llama 3 Instruct, and Mistral Instruct models. To address this limitation, we propose and evaluate ad-hoc solutions that substantially enhance LLMs' performance on long input sequences by up to 50%, while reducing API cost and latency by up to 93% and 50%, respectively.

Peyman Hosseini, Ondrej Bohdal, Taha Ceritli et al.

Test-time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test-time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, in-context learning has been widely explored at inference time and demonstrated the ability to enhance model performance without weight updates. However, TTRL's two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose context-guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7% relative accuracy improvement over TTRL), while boosting efficiency by obtaining strong performance after only a few steps of test-time training (e.g. 8% relative improvement rather than 1% over TTRL after 3 steps).

Waleed Iqbal, Gareth Tyson, Ignacio Castro

How similar are politicians to those who vote for them? This is a critical question at the heart of democratic representation and particularly relevant at times when political dissatisfaction and populism are on the rise. To answer this question we compare the online discourse of elected politicians and their constituents. We collect a two and a half years (September 2020 - February 2023) constituency-level dataset for USA and UK that includes: (i) the Twitter timelines (5.6 Million tweets) of elected political representatives (595 UK Members of Parliament and 433 USA Representatives), (ii) the Nextdoor posts (21.8 Million posts) of the constituency (98.4% USA and 91.5% UK constituencies). We find that elected politicians tend to be equally similar to their constituents in terms of content and style regardless of whether a constituency elects a right or left-wing politician. The size of the electoral victory and the level of income of a constituency shows a nuanced picture. The narrower the electoral victory, the more similar the style and the more dissimilar the content is. The lower the income of a constituency, the more similar the content is. In terms of style, poorer constituencies tend to have a more similar sentiment and more dissimilar psychological text traits (i.e. measured with LIWC categories).

Vibhor Agarwal, Aravindh Raman, Nishanth Sastry et al.

The recent development of decentralised and interoperable social networks (such as the "fediverse") creates new challenges for content moderators. This is because millions of posts generated on one server can easily "spread" to another, even if the recipient server has very different moderation policies. An obvious solution would be to leverage moderation tools to automatically tag (and filter) posts that contravene moderation policies, e.g. related to toxic speech. Recent work has exploited the conversational context of a post to improve this automatic tagging, e.g. using the replies to a post to help classify if it contains toxic speech. This has shown particular potential in environments with large training sets that contain complete conversations. This, however, creates challenges in a decentralised context, as a single conversation may be fragmented across multiple servers. Thus, each server only has a partial view of an entire conversation because conversations are often federated across servers in a non-synchronized fashion. To address this, we propose a decentralised conversation-aware content moderation approach suitable for the fediverse. Our approach employs a graph deep learning model (GraphNLI) trained locally on each server. The model exploits local data to train a model that combines post and conversational information captured through random walks to detect toxicity. We evaluate our approach with data from Pleroma, a major decentralised and interoperable micro-blogging network containing 2 million conversations. Our model effectively detects toxicity on larger instances, exclusively trained using their local post information (0.8837 macro-F1). Our approach has considerable scope to improve moderation in decentralised and interoperable social networks such as Pleroma or Mastodon.

Yimeng Gu, Zhao Tong, Ignacio Castro et al.

Multimodal out-of-context news is a type of misinformation in which the image is used outside of its original context. Many existing works have leveraged multimodal large language models (MLLMs) for detecting out-of-context news. However, observing the limited zero-shot performance of smaller MLLMs, they generally require label-rich fine-tuning and/or expensive API calls to GPT models to improve the performance, which is impractical in low-resource scenarios. In contrast, we aim to improve the performance of small MLLMs in a more label-efficient and cost-effective manner. To this end, we first prompt multiple teacher MLLMs to generate both label predictions and corresponding rationales, which collectively serve as the teachers' knowledge. We then introduce a two-stage knowledge distillation framework to transfer this knowledge to a student MLLM. In Stage 1, we apply LoRA fine-tuning to the student model using all training data. In Stage 2, we further fine-tune the student model using both LoRA fine-tuning and DPO on the data points where teachers' predictions conflict. This two-stage strategy reduces annotation costs and helps the student model uncover subtle patterns in more challenging cases. Experimental results demonstrate that our approach achieves state-of-the-art performance using less than 10% labeled data.

Haris Bin Zia, Aravindh Raman, Ignacio Castro et al.

The Fediverse, a group of interconnected servers providing a variety of interoperable services (e.g. micro-blogging in Mastodon) has gained rapid popularity. This sudden growth, partly driven by Elon Musk's acquisition of Twitter, has created challenges for administrators though. This paper focuses on one particular challenge: content moderation, e.g. the need to remove spam or hate speech. While centralized platforms like Facebook and Twitter rely on automated tools for moderation, their dependence on massive labeled datasets and specialized infrastructure renders them impractical for decentralized, low-resource settings like the Fediverse. In this work, we design and evaluate FedMod, a collaborative content moderation system based on federated learning. Our system enables servers to exchange parameters of partially trained local content moderation models with similar servers, creating a federated model shared among collaborating servers. FedMod demonstrates robust performance on three different content moderation tasks: harmful content detection, bot content detection, and content warning assignment, achieving average per-server macro-F1 scores of 0.71, 0.73, and 0.58, respectively.

Peyman Hosseini, Mehran Hosseini, Ignacio Castro et al.

From natural language processing to vision, Scaled Dot Product Attention (SDPA) is the backbone of most modern deep learning applications. Unfortunately, its memory and computational requirements can be prohibitive in low-resource settings. In this paper, we improve its efficiency without sacrificing its versatility. We propose three attention variants where we remove consecutive linear transformations or add a novel one, and evaluate them on a range of standard NLP and vision tasks. Our proposed models are substantially lighter than standard SDPA (and have 25-50% fewer parameters). We show that the performance cost of these changes is negligible relative to size reduction and that in one case (Super Attention) we succeed in outperforming SDPA by up to 10% while improving its speed and reducing its parameters by 25%.

Yimeng Gu, Mengqi Zhang, Ignacio Castro et al.

Out-of-context news is a common type of misinformation on online media platforms. This involves posting a caption, alongside a mismatched news image. Existing out-of-context news detection models only consider the scenario where pre-labeled data is available for each domain, failing to address the out-of-context news detection on unlabeled domains (e.g. news topics or agencies). In this work, we therefore focus on domain adaptive out-of-context news detection. In order to effectively adapt the detection model to unlabeled news topics or agencies, we propose ConDA-TTA (Contrastive Domain Adaptation with Test-Time Adaptation) which applies contrastive learning and maximum mean discrepancy (MMD) to learn domain-invariant features. In addition, we leverage test-time target domain statistics to further assist domain adaptation. Experimental results show that our approach outperforms baselines in most domain adaptation settings on two public datasets, by as much as 2.93% in F1 and 2.08% in accuracy.