Nina S. T. Hirata

Gabriel J. Perin, Runjin Chen, Xuxi Chen et al.

Large Language Models (LLMs) have become indispensable in real-world applications. However, their widespread adoption raises significant safety concerns, particularly in responding to socially harmful questions. Despite substantial efforts to improve model safety through alignment, aligned models can still have their safety protections undermined by subsequent fine-tuning - even when the additional training data appears benign. In this paper, we empirically demonstrate that this vulnerability stems from the sensitivity of safety-critical low-rank subspaces in LLM parameters to fine-tuning. Building on this insight, we propose a novel training-free method, termed Low-Rank Extrapolation (LoX), to enhance safety robustness by extrapolating the safety subspace of an aligned LLM. Our experimental results confirm the effectiveness of LoX, demonstrating significant improvements in robustness against both benign and malicious fine-tuning attacks while preserving the model's adaptability to new tasks. For instance, LoX leads to 11% to 54% absolute reductions in attack success rates (ASR) facing benign or malicious fine-tuning attacks. By investigating the ASR landscape of parameters, we attribute the success of LoX to that the extrapolation moves LLM parameters to a flatter zone, thereby less sensitive to perturbations. The code is available at github.com/VITA-Group/LoX.

Sergio Y. Hayashi, Nina S. T. Hirata

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, yet they struggle with spatial tasks that require mental simulation, such as mental rotation. This paper investigates whether equipping an LLM with an external ``Imagery Module'' -- a tool capable of rendering and rotating 3D models -- can bridge this gap, functioning as a ``cognitive prosthetic.'' We conducted experiments using a dual-module architecture in which a reasoning module (an MLLM) interacts with an imagery module on 3D model rotation tasks. Performance was lower than expected, with accuracy reaching at most 62.5%. Further investigation suggests that even when the burden of maintaining and manipulating a holistic 3D state is outsourced, the system still fails. This reveals that current frontier models lack the foundational visual-spatial primitives required to interface with imagery. Specifically, they lack: (1) the low-level sensitivity to extract spatial signals such as (a) depth, (b) motion, and (c) short-horizon dynamic prediction; and (2) the capacity to reason contemplatively over images, dynamically shifting visual focus and balancing imagery with symbolic and associative information.

Runjin Chen, Gabriel Jacob Perin, Xuxi Chen et al.

Alignment of large language models (LLMs) with human values and preferences, often achieved through fine-tuning based on human feedback, is essential for ensuring safe and responsible AI behaviors. However, the process typically requires substantial data and computation resources. Recent studies have revealed that alignment might be attainable at lower costs through simpler methods, such as in-context learning. This leads to the question: Is alignment predominantly superficial? In this paper, we delve into this question and provide a quantitative analysis. We formalize the concept of superficial knowledge, defining it as knowledge that can be acquired through easily token restyling, without affecting the model's ability to capture underlying causal relationships between tokens. We propose a method to extract and isolate superficial knowledge from aligned models, focusing on the shallow modifications to the final token selection process. By comparing models augmented only with superficial knowledge to fully aligned models, we quantify the superficial portion of alignment. Our findings reveal that while superficial knowledge constitutes a significant portion of alignment, particularly in safety and detoxification tasks, it is not the whole story. Tasks requiring reasoning and contextual understanding still rely on deeper knowledge. Additionally, we demonstrate two practical advantages of isolated superficial knowledge: (1) it can be transferred between models, enabling efficient offsite alignment of larger models using extracted superficial knowledge from smaller models, and (2) it is recoverable, allowing for the restoration of alignment in compromised models without sacrificing performance.

Victor Nascimento Ribeiro, Nina S. T. Hirata

Video-based vehicle detection and counting play a critical role in managing transport infrastructure. Traditional image-based counting methods usually involve two main steps: initial detection and subsequent tracking, which are applied to all video frames, leading to a significant increase in computational complexity. To address this issue, this work presents an alternative and more efficient method for vehicle detection and counting. The proposed approach eliminates the need for a tracking step and focuses solely on detecting vehicles in key video frames, thereby increasing its efficiency. To achieve this, we developed a system that combines YOLO, for vehicle detection, with Visual Rhythm, a way to create time-spatial images that allows us to focus on frames that contain useful information. Additionally, this method can be used for counting in any application involving unidirectional moving targets to be detected and identified. Experimental analysis using real videos shows that the proposed method achieves mean counting accuracy around 99.15% over a set of videos, with a processing speed three times faster than tracking based approaches.

Victor Nascimento Ribeiro, Nina S. T. Hirata

Video-based Automatic License Plate Recognition (ALPR) involves extracting vehicle license plate text information from video captures. Traditional systems typically rely heavily on high-end computing resources and utilize multiple frames to recognize license plates, leading to increased computational overhead. In this paper, we propose two methods capable of efficiently extracting exactly one frame per vehicle and recognizing its license plate characters from this single image, thus significantly reducing computational demands. The first method uses Visual Rhythm (VR) to generate time-spatial images from videos, while the second employs Accumulative Line Analysis (ALA), a novel algorithm based on single-line video processing for real-time operation. Both methods leverage YOLO for license plate detection within the frame and a Convolutional Neural Network (CNN) for Optical Character Recognition (OCR) to extract textual information. Experiments on real videos demonstrate that the proposed methods achieve results comparable to traditional frame-by-frame approaches, with processing speeds three times faster.

Giulio Cesare Mastrocinque Santo, Patrícia Izar, Irene Delval et al.

Video recordings of nonhuman primates in their natural habitat are a common source for studying their behavior in the wild. We fine-tune pre-trained video-text foundational models for the specific domain of capuchin monkeys, with the goal of developing useful computational models to help researchers to retrieve useful clips from videos. We focus on the challenging problem of training a model based solely on raw, unlabeled video footage, using weak audio descriptions sometimes provided by field collaborators. We leverage recent advances in Multimodal Large Language Models (MLLMs) and Vision-Language Models (VLMs) to address the extremely noisy nature of both video and audio content. Specifically, we propose a two-folded approach: an agentic data treatment pipeline and a fine-tuning process. The data processing pipeline automatically extracts clean and semantically aligned video-text pairs from the raw videos, which are subsequently used to fine-tune a pre-trained Microsoft's X-CLIP model through Low-Rank Adaptation (LoRA). We obtained an uplift in $Hits@5$ of $167\%$ for the 16 frames model and an uplift of $114\%$ for the 8 frame model on our domain data. Moreover, based on $NDCG@K$ results, our model is able to rank well most of the considered behaviors, while the tested raw pre-trained models are not able to rank them at all. The code will be made available upon acceptance.

Luis Vitor Zerkowski, Nina S. T. Hirata

Indigenous communities face ongoing challenges in preserving their cultural heritage, particularly in the face of systemic marginalization and urban development. In Brazil, the Museu Nacional dos Povos Indigenas through the Tainacan platform hosts the country's largest online collection of Indigenous objects and iconographies, providing a critical resource for cultural engagement. Using publicly available data from this repository, we present a data-driven initiative that applies artificial intelligence to enhance accessibility, interpretation, and exploration. We develop two semantic pipelines: a visual pipeline that models image-based similarity and a textual pipeline that captures semantic relationships from item descriptions. These embedding spaces are projected into two dimensions and integrated into an interactive visualization tool we also developed. In addition to similarity-based navigation, users can explore the collection through temporal and geographic lenses, enabling both semantic and contextualized perspectives. The system supports curatorial tasks, aids public engagement, and reveals latent connections within the collection. This work demonstrates how AI can ethically contribute to cultural preservation practices.

Victor Nascimento Ribeiro, Nina S. T. Hirata

Automatic License Plate Recognition (ALPR) involves extracting vehicle license plate information from image or a video capture. These systems have gained popularity due to the wide availability of low-cost surveillance cameras and advances in Deep Learning. Typically, video-based ALPR systems rely on multiple frames to detect the vehicle and recognize the license plates. Therefore, we propose a system capable of extracting exactly one frame per vehicle and recognizing its license plate characters from this singular image using an Optical Character Recognition (OCR) model. Early experiments show that this methodology is viable.

Sergio Y. Hayashi, Nina S. T. Hirata

Deep neural networks are largely used for complex prediction tasks. There is plenty of empirical evidence of their successful end-to-end training for a diversity of tasks. Success is often measured based solely on the final performance of the trained network, and explanations on when, why and how they work are less emphasized. In this paper we study encoder-decoder recurrent neural networks with attention mechanisms for the task of reading handwritten chess scoresheets. Rather than prediction performance, our concern is to better understand how learning occurs in these type of networks. We characterize the task in terms of three subtasks, namely input-output alignment, sequential pattern recognition, and handwriting recognition, and experimentally investigate which factors affect their learning. We identify competition, collaboration and dependence relations between the subtasks, and argue that such knowledge might help one to better balance factors to properly train a network.

Ana Martinazzo, Mateus Espadoto, Nina S. T. Hirata

In Astronomy, a huge amount of image data is generated daily by photometric surveys, which scan the sky to collect data from stars, galaxies and other celestial objects. In this paper, we propose a technique to leverage unlabeled astronomical images to pre-train deep convolutional neural networks, in order to learn a domain-specific feature extractor which improves the results of machine learning techniques in setups with small amounts of labeled data available. We show that our technique produces results which are in many cases better than using ImageNet pre-training.

Artur A. M. Oliveira, Nina S. T. Hirata, Roberto Hirata

Vegetation is a relevant feature in the urban scenery and its awareness can be measured in an image by the Green View Index (GVI). Previous approaches to estimate the GVI were based upon heuristics image processing approaches and recently by deep learning networks (DLN). By leveraging some recent DLN architectures tuned to the image segmentation problem and exploiting a weighting strategy in the loss function (LF) we improved previously reported results in similar datasets.

Mateus Espadoto, Nina S. T. Hirata, Alexandru C. Telea

Dimensionality reduction methods, also known as projections, are frequently used for exploring multidimensional data in machine learning, data science, and information visualization. Among these, t-SNE and its variants have become very popular for their ability to visually separate distinct data clusters. However, such methods are computationally expensive for large datasets, suffer from stability problems, and cannot directly handle out-of-sample data. We propose a learning approach to construct such projections. We train a deep neural network based on a collection of samples from a given data universe, and their corresponding projections, and next use the network to infer projections of data from the same, or similar, universes. Our approach generates projections with similar characteristics as the learned ones, is computationally two to three orders of magnitude faster than SNE-class methods, has no complex-to-set user parameters, handles out-of-sample data in a stable manner, and can be used to learn any projection technique. We demonstrate our proposal on several real-world high dimensional datasets from machine learning.

Frank D. Julca-Aguilar, Nina S. T. Hirata

Symbol detection techniques in online handwritten graphics (e.g. diagrams and mathematical expressions) consist of methods specifically designed for a single graphic type. In this work, we evaluate the Faster R-CNN object detection algorithm as a general method for detection of symbols in handwritten graphics. We evaluate different configurations of the Faster R-CNN method, and point out issues relative to the handwritten nature of the data. Considering the online recognition context, we evaluate efficiency and accuracy trade-offs of using Deep Neural Networks of different complexities as feature extractors. We evaluate the method on publicly available flowchart and mathematical expression (CROHME-2016) datasets. Results show that Faster R-CNN can be effectively used on both datasets, enabling the possibility of developing general methods for symbol detection, and furthermore, general graphic understanding methods that could be built on top of the algorithm.

Frank D. Julca-Aguilar, Nina S. T. Hirata

Many image transformations can be modeled by image operators that are characterized by pixel-wise local functions defined on a finite support window. In image operator learning, these functions are estimated from training data using machine learning techniques. Input size is usually a critical issue when using learning algorithms, and it limits the size of practicable windows. We propose the use of convolutional neural networks (CNNs) to overcome this limitation. The problem of removing staff-lines in music score images is chosen to evaluate the effects of window and convolutional mask sizes on the learned image operator performance. Results show that the CNN based solution outperforms previous ones obtained using conventional learning algorithms or heuristic algorithms, indicating the potential of CNNs as base classifiers in image operator learning. The implementations will be made available on the TRIOSlib project site.

Frank Julca-Aguilar, Harold Mouchère, Christian Viard-Gaudin et al.

We propose a new framework for the recognition of online handwritten graphics. Three main features of the framework are its ability to treat symbol and structural level information in an integrated way, its flexibility with respect to different families of graphics, and means to control the tradeoff between recognition effectiveness and computational cost. We model a graphic as a labeled graph generated from a graph grammar. Non-terminal vertices represent subcomponents, terminal vertices represent symbols, and edges represent relations between subcomponents or symbols. We then model the recognition problem as a graph parsing problem: given an input stroke set, we search for a parse tree that represents the best interpretation of the input. Our graph parsing algorithm generates multiple interpretations (consistent with the grammar) and then we extract an optimal interpretation according to a cost function that takes into consideration the likelihood scores of symbols and structures. The parsing algorithm consists in recursively partitioning the stroke set according to structures defined in the grammar and it does not impose constraints present in some previous works (e.g. stroke ordering). By avoiding such constraints and thanks to the powerful representativeness of graphs, our approach can be adapted to the recognition of different graphic notations. We show applications to the recognition of mathematical expressions and flowcharts. Experimentation shows that our method obtains state-of-the-art accuracy in both applications.