Florin Pop

Vlad-Constantin Lungu-Stan, Dumitru-Clementin Cercel, Florin Pop

Skin cancer is a treatable disease if discovered early. We provide a production-specific solution to the skin cancer classification problem that matches human performance in melanoma identification by training a vision transformer on melanoma medical images annotated by experts. Since inference cost, both time and memory wise is important in practice, we employ knowledge distillation to obtain a model that retains 98.33% of the teacher's balanced multi-class accuracy, at a fraction of the cost. Memory-wise, our model is 49.60% smaller than the teacher. Time-wise, our solution is 69.25% faster on GPU and 97.96% faster on CPU. By adding classification heads at each level of the transformer and employing a cascading distillation process, we improve the balanced multi-class accuracy of the base model by 2.1%, while creating a range of models of various sizes but comparable performance. We provide the code at https://github.com/Longman-Stan/SkinDistilVit.

Răzvan-Alexandru Smădu, George-Eduard Zaharia, Andrei-Marius Avram et al.

Keyphrase identification and classification is a Natural Language Processing and Information Retrieval task that involves extracting relevant groups of words from a given text related to the main topic. In this work, we focus on extracting keyphrases from scientific documents. We introduce TA-DA, a Topic-Aware Domain Adaptation framework for keyphrase extraction that integrates Multi-Task Learning with Adversarial Training and Domain Adaptation. Our approach improves performance over baseline models by up to 5% in the exact match of the F1-score.

Sebastian-Vasile Echim, Răzvan-Alexandru Smădu, Andrei-Marius Avram et al.

Satire detection and sentiment analysis are intensively explored natural language processing (NLP) tasks that study the identification of the satirical tone from texts and extracting sentiments in relationship with their targets. In languages with fewer research resources, an alternative is to produce artificial examples based on character-level adversarial processes to overcome dataset size limitations. Such samples are proven to act as a regularization method, thus improving the robustness of models. In this work, we improve the well-known NLP models (i.e., Convolutional Neural Networks, Long Short-Term Memory (LSTM), Bidirectional LSTM, Gated Recurrent Units (GRUs), and Bidirectional GRUs) with adversarial training and capsule networks. The fine-tuned models are used for satire detection and sentiment analysis tasks in the Romanian language. The proposed framework outperforms the existing methods for the two tasks, achieving up to 99.08% accuracy, thus confirming the improvements added by the capsule layers and the adversarial training in NLP approaches.

Iulian-Marius Tăiatu, Andrei-Marius Avram, Dumitru-Clementin Cercel et al.

Developing natural language processing (NLP) systems for social media analysis remains an important topic in artificial intelligence research. This article introduces RoBERTweet, the first Transformer architecture trained on Romanian tweets. Our RoBERTweet comes in two versions, following the base and large architectures of BERT. The corpus used for pre-training the models represents a novelty for the Romanian NLP community and consists of all tweets collected from 2008 to 2022. Experiments show that RoBERTweet models outperform the previous general-domain Romanian and multilingual language models on three NLP tasks with tweet inputs: emotion detection, sexist language identification, and named entity recognition. We make our models and the newly created corpus of Romanian tweets freely available.

Răzvan-Alexandru Smădu, Sebastian-Vasile Echim, Dumitru-Clementin Cercel et al.

Unsupervised Domain Adaptation (UDA) is a popular technique that aims to reduce the domain shift between two data distributions. It was successfully applied in computer vision and natural language processing. In the current work, we explore the effects of various unsupervised domain adaptation techniques between two text classification tasks: fake and hyperpartisan news detection. We investigate the knowledge transfer from fake to hyperpartisan news detection without involving target labels during training. Thus, we evaluate UDA, cluster alignment with a teacher, and cross-domain contrastive learning. Extensive experiments show that these techniques improve performance, while including data augmentation further enhances the results. In addition, we combine clustering and topic modeling algorithms with UDA, resulting in improved performances compared to the initial UDA setup.

Elena-Beatrice Nicola, Dumitru-Clementin Cercel, Florin Pop

Offensive language detection is a crucial task in today's digital landscape, where online platforms grapple with maintaining a respectful and inclusive environment. However, building robust offensive language detection models requires large amounts of labeled data, which can be expensive and time-consuming to obtain. Semi-supervised learning offers a feasible solution by utilizing labeled and unlabeled data to create more accurate and robust models. In this paper, we explore a few different semi-supervised methods, as well as data augmentation techniques. Concretely, we implemented eight semi-supervised methods and ran experiments for them using only the available data in the RO-Offense dataset and applying five augmentation techniques before feeding the data to the models. Experimental results demonstrate that some of them benefit more from augmentations than others.

Emilian-Claudiu Mănescu, Răzvan-Alexandru Smădu, Andrei-Marius Avram et al.

Lip reading or visual speech recognition has gained significant attention in recent years, particularly because of hardware development and innovations in computer vision. While considerable progress has been obtained, most models have only been tested on a few large-scale datasets. This work addresses this shortcoming by analyzing several architectures and optimizations on the underrepresented, short-scale Romanian language dataset called Wild LRRo. Most notably, we compare different backend modules, demonstrating the effectiveness of adding ample regularization methods. We obtain state-of-the-art results using our proposed method, namely cross-lingual domain adaptation and unlabeled videos from English and German datasets to help the model learn language-invariant features. Lastly, we assess the performance of adding a layer inspired by the neural inhibition mechanism.

Răzvan-Alexandru Smădu, David-Gabriel Ion, Dumitru-Clementin Cercel et al.

Complex Word Identification (CWI) is an essential step in the lexical simplification task and has recently become a task on its own. Some variations of this binary classification task have emerged, such as lexical complexity prediction (LCP) and complexity evaluation of multi-word expressions (MWE). Large language models (LLMs) recently became popular in the Natural Language Processing community because of their versatility and capability to solve unseen tasks in zero/few-shot settings. Our work investigates LLM usage, specifically open-source models such as Llama 2, Llama 3, and Vicuna v1.5, and closed-source, such as ChatGPT-3.5-turbo and GPT-4o, in the CWI, LCP, and MWE settings. We evaluate zero-shot, few-shot, and fine-tuning settings and show that LLMs struggle in certain conditions or achieve comparable results against existing methods. In addition, we provide some views on meta-learning combined with prompt learning. In the end, we conclude that the current state of LLMs cannot or barely outperform existing methods, which are usually much smaller.

Vickram N. Premakumar, Michael Vaiana, Florin Pop et al.

Self-models have been a topic of great interest for decades in studies of human cognition and more recently in machine learning. Yet what benefits do self-models confer? Here we show that when artificial networks learn to predict their internal states as an auxiliary task, they change in a fundamental way. To better perform the self-model task, the network learns to make itself simpler, more regularized, more parameter-efficient, and therefore more amenable to being predictively modeled. To test the hypothesis of self-regularizing through self-modeling, we used a range of network architectures performing three classification tasks across two modalities. In all cases, adding self-modeling caused a significant reduction in network complexity. The reduction was observed in two ways. First, the distribution of weights was narrower when self-modeling was present. Second, a measure of network complexity, the real log canonical threshold (RLCT), was smaller when self-modeling was present. Not only were measures of complexity reduced, but the reduction became more pronounced as greater training weight was placed on the auxiliary task of self-modeling. These results strongly support the hypothesis that self-modeling is more than simply a network learning to predict itself. The learning has a restructuring effect, reducing complexity and increasing parameter efficiency. This self-regularization may help explain some of the benefits of self-models reported in recent machine learning literature, as well as the adaptive value of self-models to biological systems. In particular, these findings may shed light on the possible interaction between the ability to model oneself and the ability to be more easily modeled by others in a social or cooperative context.

Andrei-Marius Avram, Mircea Timpuriu, Andreea Iuga et al.

Using supervised automatic summarisation methods requires sufficient corpora that include pairs of documents and their summaries. Similarly to many tasks in natural language processing, most of the datasets available for summarization are in English, posing challenges for developing summarization models in other languages. Thus, in this work, we introduce RoLargeSum, a novel large-scale summarization dataset for the Romanian language crawled from various publicly available news websites from Romania and the Republic of Moldova that were thoroughly cleaned to ensure a high-quality standard. RoLargeSum contains more than 615K news articles, together with their summaries, as well as their headlines, keywords, dialect, and other metadata that we found on the targeted websites. We further evaluated the performance of several BART variants and open-source large language models on RoLargeSum for benchmarking purposes. We manually evaluated the results of the best-performing system to gain insight into the potential pitfalls of this data set and future development.

Adrian Gheorghiu, Iulian-Marius Tăiatu, Dumitru-Clementin Cercel et al.

The detection and classification of diseases in Robusta coffee leaves are essential to ensure that plants are healthy and the crop yield is kept high. However, this job requires extensive botanical knowledge and much wasted time. Therefore, this task and others similar to it have been extensively researched subjects in image classification. Regarding leaf disease classification, most approaches have used the more popular PlantVillage dataset while completely disregarding other datasets, like the Robusta Coffee Leaf (RoCoLe) dataset. As the RoCoLe dataset is imbalanced and does not have many samples, fine-tuning of pre-trained models and multiple augmentation techniques need to be used. The current paper uses the RoCoLe dataset and approaches based on deep learning for classifying coffee leaf diseases from images, incorporating the pix2pix model for segmentation and cycle-generative adversarial network (CycleGAN) for augmentation. Our study demonstrates the effectiveness of Transformer-based models, online augmentations, and CycleGAN augmentation in improving leaf disease classification. While synthetic data has limitations, it complements real data, enhancing model performance. These findings contribute to developing robust techniques for plant disease detection and classification.

Sebastian-Vasile Echim, Iulian-Marius Tăiatu, Dumitru-Clementin Cercel et al.

This work focuses on plant leaf disease classification and explores three crucial aspects: adversarial training, model explainability, and model compression. The models' robustness against adversarial attacks is enhanced through adversarial training, ensuring accurate classification even in the presence of threats. Leveraging explainability techniques, we gain insights into the model's decision-making process, improving trust and transparency. Additionally, we explore model compression techniques to optimize computational efficiency while maintaining classification performance. Through our experiments, we determine that on a benchmark dataset, the robustness can be the price of the classification accuracy with performance reductions of 3%-20% for regular tests and gains of 50%-70% for adversarial attack tests. We also demonstrate that a student model can be 15-25 times more computationally efficient for a slight performance reduction, distilling the knowledge of more complex models.

Keenan Pepper, Alex McKenzie, Florin Pop et al.

Self-interpretation methods prompt language models to describe their own internal states, but remain unreliable due to hyperparameter sensitivity. We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation across tasks and model families. A scalar affine adapter with just $d_\text{model}+1$ parameters suffices: trained adapters generate sparse autoencoder feature labels that outperform the training labels themselves (71% vs 63% generation scoring at 70B scale), identify topics with 94% recall@1 versus 1% for untrained baselines, and decode bridge entities in multi-hop reasoning that appear in neither prompt nor response, surfacing implicit reasoning without chain-of-thought. The learned bias vector alone accounts for 85% of improvement, and simpler adapters generalize better than more expressive alternatives. Controlling for model knowledge via prompted descriptions, we find self-interpretation gains outpace capability gains from 7B to 72B parameters. Our results demonstrate that self-interpretation improves with scale, without modifying the model being interpreted.

Adrian-Dinu Urse, Dumitru-Clementin Cercel, Florin Pop

Natural disaster assessment relies on accurate and rapid access to information, with social media emerging as a valuable real-time source. However, existing datasets suffer from class imbalance and limited samples, making effective model development a challenging task. This paper explores augmentation techniques to address these issues on the CrisisMMD multimodal dataset. For visual data, we apply diffusion-based methods, namely Real Guidance and DiffuseMix. For text data, we explore back-translation, paraphrasing with transformers, and image caption-based augmentation. We evaluated these across unimodal, multimodal, and multi-view learning setups. Results show that selected augmentations improve classification performance, particularly for underrepresented classes, while multi-view learning introduces potential but requires further refinement. This study highlights effective augmentation strategies for building more robust disaster assessment systems.

Sebastian-Vasile Echim, Andrei-Alexandru Preda, Dumitru-Clementin Cercel et al.

Deep neural networks currently dominate many fields of the artificial intelligence landscape, achieving state-of-the-art results on numerous tasks while remaining hard to understand and exhibiting surprising weaknesses. An active area of research focuses on adversarial attacks, which aim to generate inputs that uncover these weaknesses. However, this proves challenging, especially in the black-box scenario where model details are inaccessible. This paper explores in detail the impact of such adversarial algorithms on ResNet-18, DenseNet-121, Swin Transformer V2, and Vision Transformer network architectures. Leveraging the Tiny ImageNet, Caltech-256, and Food-101 datasets, we benchmark two novel black-box iterative adversarial algorithms based on affine transformations and genetic algorithms: 1) Affine Transformation Attack (ATA), an iterative algorithm maximizing our attack score function using random affine transformations, and 2) Affine Genetic Attack (AGA), a genetic algorithm that involves random noise and affine transformations. We evaluate the performance of the models in the algorithm parameter variation, data augmentation, and global and targeted attack configurations. We also compare our algorithms with two black-box adversarial algorithms, Pixle and Square Attack. Our experiments yield better results on the image classification task than similar methods in the literature, achieving an accuracy improvement of up to 8.82%. We provide noteworthy insights into successful adversarial defenses and attacks at both global and targeted levels, and demonstrate adversarial robustness through algorithm parameter variation.

Răzvan-Alexandru Smădu, Andreea Iuga, Dumitru-Clementin Cercel et al.

Satire, irony, and sarcasm are techniques typically used to express humor and critique, rather than deceive; however, they can occasionally be mistaken for factual reporting, akin to fake news. These techniques can be applied at a more granular level, allowing satirical information to be incorporated into news articles. In this paper, we introduce the first sentence-level dataset for Romanian satire detection for news articles, called SeLeRoSa. The dataset comprises 13,873 manually annotated sentences spanning various domains, including social issues, IT, science, and movies. With the rise and recent progress of large language models (LLMs) in the natural language processing literature, LLMs have demonstrated enhanced capabilities to tackle various tasks in zero-shot settings. We evaluate multiple baseline models based on LLMs in both zero-shot and fine-tuning settings, as well as baseline transformer-based models. Our findings reveal the current limitations of these models in the sentence-level satire detection task, paving the way for new research directions.

Andrei Vlad Man, Răzvan-Alexandru Smădu, Cristian-George Craciun et al.

The intersection of AI and legal systems presents a growing need for tools that support legal education, particularly in under-resourced languages such as Romanian. In this work, we aim to evaluate the capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs) in understanding and reasoning about Romanian driving law through textual and visual question-answering tasks. To facilitate this, we introduce RoD-TAL, a novel multimodal dataset comprising Romanian driving test questions, text-based and image-based, alongside annotated legal references and human explanations. We implement and assess retrieval-augmented generation (RAG) pipelines, dense retrievers, and reasoning-optimized models across tasks including Information Retrieval (IR), Question Answering (QA), Visual IR, and Visual QA. Our experiments demonstrate that domain-specific fine-tuning significantly enhances retrieval performance. At the same time, chain-of-thought prompting and specialized reasoning models improve QA accuracy, surpassing the minimum grades required to pass driving exams. However, visual reasoning remains challenging, highlighting the potential and the limitations of applying LLMs and VLMs to legal education.

Mihnea-Alexandru Vîrlan, Răzvan-Alexandru Smădu, Dumitru-Clementin Cercel et al.

The primary goal of a news headline is to summarize an event in as few words as possible. Depending on the media outlet, a headline can serve as a means to objectively deliver a summary or improve its visibility. For the latter, specific publications may employ stylistic approaches that incorporate the use of sarcasm, irony, and exaggeration, key elements of a satirical approach. As such, even the headline must reflect the tone of the satirical main content. Current approaches for the Romanian language tend to detect the non-conventional tone (i.e., satire and clickbait) of the news content by combining both the main article and the headline. Because we consider a headline to be merely a brief summary of the main article, we investigate in this paper the presence of satirical tone in headlines alone, testing multiple baselines ranging from standard machine learning algorithms to deep learning models. Our experiments show that Bidirectional Transformer models outperform both standard machine-learning approaches and Large Language Models (LLMs), particularly when the meta-learning Reptile approach is employed.

Florin Pop, Judd Rosenblatt, Diogo Schwerz de Lucena et al.

In this paper, we investigate the degree to which fine-tuning in Large Language Models (LLMs) effectively mitigates versus merely conceals undesirable behavior. Through the lens of semi-realistic role-playing exercises designed to elicit such behaviors, we explore the response dynamics of LLMs post fine-tuning interventions. Our methodology involves prompting models for Chain-of-Thought (CoT) reasoning and analyzing the coherence between the reasoning traces and the resultant outputs. Notably, we identify a pervasive phenomenon we term \emph{reason-based deception}, where models either stop producing reasoning traces or produce seemingly ethical reasoning traces that belie the unethical nature of their final outputs. We further examine the efficacy of response strategies (polite refusal versus explicit rebuttal) in curbing the occurrence of undesired behavior in subsequent outputs of multi-turn interactions. Our findings reveal that explicit rebuttals significantly outperform polite refusals in preventing the continuation of undesired outputs and nearly eliminate reason-based deception, challenging current practices in model fine-tuning. Accordingly, the two key contributions of this paper are (1) defining and studying reason-based deception, a new type of hidden behavior, and (2) demonstrating that rebuttals provide a more robust response model to harmful requests than refusals, thereby highlighting the need to reconsider the response strategies in fine-tuning approaches.