Enzo Tartaglione

Olivier Laurent, Adrien Lafage, Enzo Tartaglione et al.

Deep Ensembles (DE) are a prominent approach for achieving excellent performance on key metrics such as accuracy, calibration, uncertainty estimation, and out-of-distribution detection. However, hardware limitations of real-world systems constrain to smaller ensembles and lower-capacity networks, significantly deteriorating their performance and properties. We introduce Packed-Ensembles (PE), a strategy to design and train lightweight structured ensembles by carefully modulating the dimension of their encoding space. We leverage grouped convolutions to parallelize the ensemble into a single shared backbone and forward pass to improve training and inference speeds. PE is designed to operate within the memory limits of a standard neural network. Our extensive research indicates that PE accurately preserves the properties of DE, such as diversity, and performs equally well in terms of accuracy, calibration, out-of-distribution detection, and robustness to distribution shift. We make our code available at https://github.com/ENSTA-U2IS/torch-uncertainty.

Daniele Perlo, Enzo Tartaglione, Umberto Gava et al.

The CT perfusion (CTP) is a medical exam for measuring the passage of a bolus of contrast solution through the brain on a pixel-by-pixel basis. The objective is to draw "perfusion maps" (namely cerebral blood volume, cerebral blood flow and time to peak) very rapidly for ischemic lesions, and to be able to distinguish between core and penumubra regions. A precise and quick diagnosis, in a context of ischemic stroke, can determine the fate of the brain tissues and guide the intervention and treatment in emergency conditions. In this work we present UniToBrain dataset, the very first open-source dataset for CTP. It comprises a cohort of more than a hundred of patients, and it is accompanied by patients metadata and ground truth maps obtained with state-of-the-art algorithms. We also propose a novel neural networks-based algorithm, using the European library ECVL and EDDL for the image processing and developing deep learning models respectively. The results obtained by the neural network models match the ground truth and open the road towards potential sub-sampling of the required number of CT maps, which impose heavy radiation doses to the patients.

Zhu Liao, Victor Quétu, Van-Tam Nguyen et al.

Pruning is a widely used technique for reducing the size of deep neural networks while maintaining their performance. However, such a technique, despite being able to massively compress deep models, is hardly able to remove entire layers from a model (even when structured): is this an addressable task? In this study, we introduce EGP, an innovative Entropy Guided Pruning algorithm aimed at reducing the size of deep neural networks while preserving their performance. The key focus of EGP is to prioritize pruning connections in layers with low entropy, ultimately leading to their complete removal. Through extensive experiments conducted on popular models like ResNet-18 and Swin-T, our findings demonstrate that EGP effectively compresses deep neural networks while maintaining competitive performance levels. Our results not only shed light on the underlying mechanism behind the advantages of unstructured pruning, but also pave the way for further investigations into the intricate relationship between entropy, pruning techniques, and deep learning performance. The EGP algorithm and its insights hold great promise for advancing the field of network compression and optimization. The source code for EGP is released open-source.

Imad Eddine Marouf, Subhankar Roy, Stéphane Lathuilière et al.

The integration of large pre-trained models (PTMs) into Class-Incremental Learning (CIL) has facilitated the development of computationally efficient strategies such as First-Session Adaptation (FSA), which fine-tunes the model solely on the first task while keeping it frozen for subsequent tasks. Although effective in homogeneous task sequences, these approaches struggle when faced with the heterogeneity of real-world task distributions. We introduce Plasticity-Enhanced Test-Time Adaptation in Class-Incremental Learning (PLASTIC), a method that reinstates plasticity in CIL while preserving model stability. PLASTIC leverages Test-Time Adaptation (TTA) by dynamically fine-tuning LayerNorm parameters on unlabeled test data, enabling adaptability to evolving tasks and improving robustness against data corruption. To prevent TTA-induced model divergence and maintain stable learning across tasks, we introduce a teacher-student distillation framework, ensuring that adaptation remains controlled and generalizable. Extensive experiments across multiple benchmarks demonstrate that PLASTIC consistently outperforms both conventional and state-of-the-art PTM-based CIL approaches, while also exhibiting inherent robustness to data corruptions. Code is available at: https://github.com/IemProg/PLASTIC.

Victor Quétu, Enzo Tartaglione

Neoteric works have shown that modern deep learning models can exhibit a sparse double descent phenomenon. Indeed, as the sparsity of the model increases, the test performance first worsens since the model is overfitting the training data; then, the overfitting reduces, leading to an improvement in performance, and finally, the model begins to forget critical information, resulting in underfitting. Such a behavior prevents using traditional early stop criteria. In this work, we have three key contributions. First, we propose a learning framework that avoids such a phenomenon and improves generalization. Second, we introduce an entropy measure providing more insights into the insurgence of this phenomenon and enabling the use of traditional stop criteria. Third, we provide a comprehensive quantitative analysis of contingent factors such as re-initialization methods, model width and depth, and dataset noise. The contributions are supported by empirical evidence in typical setups. Our code is available at https://github.com/VGCQ/DSD2.

Carl De Sousa Trias, Mihai Mitrea, Attilio Fiandrotti et al.

Nowadays, deep neural networks are used for solving complex tasks in several critical applications and protecting both their integrity and intellectual property rights (IPR) has become of utmost importance. To this end, we advance WaterMAS, a substitutive, white-box neural network watermarking method that improves the trade-off among robustness, imperceptibility, and computational complexity, while making provisions for increased data payload and security. WasterMAS insertion keeps unchanged the watermarked weights while sharpening their underlying gradient space. The robustness is thus ensured by limiting the attack's strength: even small alterations of the watermarked weights would impact the model's performance. The imperceptibility is ensured by inserting the watermark during the training process. The relationship among the WaterMAS data payload, imperceptibility, and robustness properties is discussed. The secret key is represented by the positions of the weights conveying the watermark, randomly chosen through multiple layers of the model. The security is evaluated by investigating the case in which an attacker would intercept the key. The experimental validations consider 5 models and 2 tasks (VGG16, ResNet18, MobileNetV3, SwinT for CIFAR10 image classification, and DeepLabV3 for Cityscapes image segmentation) as well as 4 types of attacks (Gaussian noise addition, pruning, fine-tuning, and quantization). The code will be released open-source upon acceptance of the article.

Maxime Girard, Victor Quétu, Samuel Tardieu et al.

Deploying Deep Neural Networks (DNNs) on different hardware platforms is challenging due to varying resource constraints. Besides handcrafted approaches aiming at making deep models hardware-friendly, Neural Architectures Search is rising as a toolbox to craft more efficient DNNs without sacrificing performance. Among these, the Once-For-All (OFA) approach offers a solution by allowing the sampling of well-performing sub-networks from a single supernet -- this leads to evident advantages in terms of computation. However, OFA does not fully utilize the potential memory capacity of the target device, focusing instead on limiting maximum memory usage per layer. This leaves room for an unexploited potential in terms of model generalizability. In this paper, we introduce a Memory-Optimized OFA (MOOFA) supernet, designed to enhance DNN deployment on resource-limited devices by maximizing memory usage (and for instance, features diversity) across different configurations. Tested on ImageNet, our MOOFA supernet demonstrates improvements in memory exploitation and model accuracy compared to the original OFA supernet. Our code is available at https://github.com/MaximeGirard/memory-optimized-once-for-all.

Carlo Alberto Barbano, Benoit Dufumier, Enzo Tartaglione et al.

Many datasets are biased, namely they contain easy-to-learn features that are highly correlated with the target class only in the dataset but not in the true underlying distribution of the data. For this reason, learning unbiased models from biased data has become a very relevant research topic in the last years. In this work, we tackle the problem of learning representations that are robust to biases. We first present a margin-based theoretical framework that allows us to clarify why recent contrastive losses (InfoNCE, SupCon, etc.) can fail when dealing with biased data. Based on that, we derive a novel formulation of the supervised contrastive loss (epsilon-SupInfoNCE), providing more accurate control of the minimal distance between positive and negative samples. Furthermore, thanks to our theoretical framework, we also propose FairKL, a new debiasing regularization loss, that works well even with extremely biased data. We validate the proposed losses on standard vision datasets including CIFAR10, CIFAR100, and ImageNet, and we assess the debiasing capability of FairKL with epsilon-SupInfoNCE, reaching state-of-the-art performance on a number of biased datasets, including real instances of biases in the wild.

Ivan Luiz De Moura Matos, Abdel Djalil Sad Saoud, Ekaterina Iakovleva et al.

The issue of algorithmic biases in deep learning has led to the development of various debiasing techniques, many of which perform complex training procedures or dataset manipulation. However, an intriguing question arises: is it possible to extract fair and bias-agnostic subnetworks from standard vanilla-trained models without relying on additional data, such as unbiased training set? In this work, we introduce Bias-Invariant Subnetwork Extraction (BISE), a learning strategy that identifies and isolates "bias-free" subnetworks that already exist within conventionally trained models, without retraining or finetuning the original parameters. Our approach demonstrates that such subnetworks can be extracted via pruning and can operate without modification, effectively relying less on biased features and maintaining robust performance. Our findings contribute towards efficient bias mitigation through structural adaptation of pre-trained neural networks via parameter removal, as opposed to costly strategies that are either data-centric or involve (re)training all model parameters. Extensive experiments on common benchmarks show the advantages of our approach in terms of the performance and computational efficiency of the resulting debiased model.

Nathan Roos, Ekaterina Iakovleva, Ani Gjergji et al.

Diffusion-based generative models demonstrate state-of-the-art performance across various image synthesis tasks, yet their tendency to replicate and amplify dataset biases remains poorly understood. Although previous research has viewed bias amplification as an inherent characteristic of diffusion models, this work provides the first analysis of how sampling algorithms and their hyperparameters influence bias amplification. We empirically demonstrate that samplers for diffusion models -- commonly optimized for sample quality and speed -- have a significant and measurable effect on bias amplification. Through controlled studies with models trained on Biased MNIST, Multi-Color MNIST and BFFHQ, and with Stable Diffusion, we show that sampling hyperparameters can induce both bias reduction and amplification, even when the trained model is fixed. Source code is available at https://github.com/How-I-met-your-bias/how_i_met_your_bias.

Andrea Bragagnolo, Enzo Tartaglione, Marco Grangetto

Recent advances in deep learning optimization showed that, with some a-posteriori information on fully-trained models, it is possible to match the same performance by simply training a subset of their parameters. Such a discovery has a broad impact from theory to applications, driving the research towards methods to identify the minimum subset of parameters to train without look-ahead information exploitation. However, the methods proposed do not match the state-of-the-art performance, and rely on unstructured sparsely connected models. In this work we shift our focus from the single parameters to the behavior of the whole neuron, exploiting the concept of neuronal equilibrium (NEq). When a neuron is in a configuration at equilibrium (meaning that it has learned a specific input-output relationship), we can halt its update; on the contrary, when a neuron is at non-equilibrium, we let its state evolve towards an equilibrium state, updating its parameters. The proposed approach has been tested on different state-of-the-art learning strategies and tasks, validating NEq and observing that the neuronal equilibrium depends on the specific learning setup.

Chenxi Lola Deng, Enzo Tartaglione

NeRFs have revolutionized the world of per-scene radiance field reconstruction because of their intrinsic compactness. One of the main limitations of NeRFs is their slow rendering speed, both at training and inference time. Recent research focuses on the optimization of an explicit voxel grid (EVG) that represents the scene, which can be paired with neural networks to learn radiance fields. This approach significantly enhances the speed both at train and inference time, but at the cost of large memory occupation. In this work we propose Re:NeRF, an approach that specifically targets EVG-NeRFs compressibility, aiming to reduce memory storage of NeRF models while maintaining comparable performance. We benchmark our approach with three different EVG-NeRF architectures on four popular benchmarks, showing Re:NeRF's broad usability and effectiveness.

Imad Eddine Marouf, Enzo Tartaglione, Stéphane Lathuilière

Vision Transformers (ViTs) have become one of the dominant architectures in computer vision, and pre-trained ViT models are commonly adapted to new tasks via fine-tuning. Recent works proposed several parameter-efficient transfer learning methods, such as adapters, to avoid the prohibitive training and storage cost of finetuning. In this work, we observe that adapters perform poorly when the dimension of adapters is small, and we propose MiMi, a training framework that addresses this issue. We start with large adapters which can reach high performance, and iteratively reduce their size. To enable automatic estimation of the hidden dimension of every adapter, we also introduce a new scoring function, specifically designed for adapters, that compares the neuron importance across layers. Our method outperforms existing methods in finding the best trade-off between accuracy and trained parameters across the three dataset benchmarks DomainNet, VTAB, and Multi-task, for a total of 29 datasets.

Enzo Tartaglione, Francesca Gennari, Marco Grangetto

Deep learning models are nowadays broadly deployed to solve an incredibly large variety of tasks. However, little attention has been devoted to connected legal aspects. In 2016, the European Union approved the General Data Protection Regulation which entered into force in 2018. Its main rationale was to protect the privacy and data protection of its citizens by the way of operating of the so-called "Data Economy". As data is the fuel of modern Artificial Intelligence, it is argued that the GDPR can be partly applicable to a series of algorithmic decision making tasks before a more structured AI Regulation enters into force. In the meantime, AI should not allow undesired information leakage deviating from the purpose for which is created. In this work we propose DisP, an approach for deep learning models disentangling the information related to some classes we desire to keep private, from the data processed by AI. In particular, DisP is a regularization strategy de-correlating the features belonging to the same private class at training time, hiding the information of private classes membership. Our experiments on state-of-the-art deep learning models show the effectiveness of DisP, minimizing the risk of extraction for the classes we desire to keep private.

Riccardo Renzulli, Enzo Tartaglione, Marco Grangetto

Capsule networks are biologically inspired neural networks that group neurons into vectors called capsules, each explicitly representing an object or one of its parts. The routing mechanism connects capsules in consecutive layers, forming a hierarchical structure between parts and objects, also known as a parse tree. Capsule networks often attempt to model all elements in an image, requiring large network sizes to handle complexities such as intricate backgrounds or irrelevant objects. However, this comprehensive modeling leads to increased parameter counts and computational inefficiencies. Our goal is to enable capsule networks to focus only on the object of interest, reducing the number of parse trees. We accomplish this with REM (Routing Entropy Minimization), a technique that minimizes the entropy of the parse tree-like structure. REM drives the model parameters distribution towards low entropy configurations through a pruning mechanism, significantly reducing the generation of intra-class parse trees. This empowers capsules to learn more stable and succinct representations with fewer parameters and negligible performance loss.

Yinghao Wang, Rémi Nahon, Enzo Tartaglione et al.

In this paper, we present a new approach to mental state classification from EEG signals by combining signal processing techniques and machine learning (ML) algorithms. We evaluate the performance of the proposed method on a dataset of EEG recordings collected during a cognitive load task and compared it to other state-of-the-art methods. The results show that the proposed method achieves high accuracy in classifying mental states and outperforms state-of-the-art methods in terms of classification accuracy and computational efficiency.

Carlo Alberto Barbano, Enzo Tartaglione, Marco Grangetto

Deep neural networks often struggle to learn robust representations in the presence of dataset biases, leading to suboptimal generalization on unbiased datasets. This limitation arises because the models heavily depend on peripheral and confounding factors, inadvertently acquired during training. Existing approaches to address this problem typically involve explicit supervision of bias attributes or reliance on prior knowledge about the biases. In this study, we address the challenging scenario where no explicit annotations of bias are available, and there's no prior knowledge about its nature. We present a fully unsupervised debiasing framework with three key steps: firstly, leveraging the inherent tendency to learn malignant biases to acquire a bias-capturing model; next, employing a pseudo-labeling process to obtain bias labels; and finally, applying cutting-edge supervised debiasing techniques to achieve an unbiased model. Additionally, we introduce a theoretical framework for evaluating model biasedness and conduct a detailed analysis of how biases impact neural network training. Experimental results on both synthetic and real-world datasets demonstrate the effectiveness of our method, showcasing state-of-the-art performance in various settings, occasionally surpassing fully supervised debiasing approaches.

Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano et al.

In the last few years, due to the broad applicability of deep learning to downstream tasks and end-to-end training capabilities, increasingly more concerns about potential biases to specific, non-representative patterns have been raised. Many works focusing on unsupervised debiasing usually leverage the tendency of deep models to learn ``easier'' samples, for example by clustering the latent space to obtain bias pseudo-labels. However, the interpretation of such pseudo-labels is not trivial, especially for a non-expert end user, as it does not provide semantic information about the bias features. To address this issue, we introduce ``Say My Name'' (SaMyNa), the first tool to identify biases within deep models semantically. Unlike existing methods, our approach focuses on biases learned by the model. Our text-based pipeline enhances explainability and supports debiasing efforts: applicable during either training or post-hoc validation, our method can disentangle task-related information and proposes itself as a tool to analyze biases. Evaluation on traditional benchmarks demonstrates its effectiveness in detecting biases and even disclaiming them, showcasing its broad applicability for model diagnosis.

Dorian Gailhard, Enzo Tartaglione, Lirida Naviner et al.

Hypergraphs are powerful mathematical structures that can model complex, high-order relationships in various domains, including social networks, bioinformatics, and recommender systems. However, generating realistic and diverse hypergraphs remains challenging due to their inherent complexity and lack of effective generative models. In this paper, we introduce a diffusion-based Hypergraph Generation (HYGENE) method that addresses these challenges through a progressive local expansion approach. HYGENE works on the bipartite representation of hypergraphs, starting with a single pair of connected nodes and iteratively expanding it to form the target hypergraph. At each step, nodes and hyperedges are added in a localized manner using a denoising diffusion process, which allows for the construction of the global structure before refining local details. Our experiments demonstrated the effectiveness of HYGENE, proving its ability to closely mimic a variety of properties in hypergraphs. To the best of our knowledge, this is the first attempt to employ deep learning models for hypergraph generation, and our work aims to lay the groundwork for future research in this area.

Imad Eddine Marouf, Subhankar Roy, Enzo Tartaglione et al.

In this work, we study the problem of continual learning (CL) where the goal is to learn a model on a sequence of tasks, such that the data from the previous tasks becomes unavailable while learning on the current task data. CL is essentially a balancing act between being able to learn on the new task (i.e., plasticity) and maintaining the performance on the previously learned concepts (i.e., stability). Intending to address the stability-plasticity trade-off, we propose to perform weight-ensembling of the model parameters of the previous and current tasks. This weighted-ensembled model, which we call Continual Model Averaging (or CoMA), attains high accuracy on the current task by leveraging plasticity, while not deviating too far from the previous weight configuration, ensuring stability. We also propose an improved variant of CoMA, named Continual Fisher-weighted Model Averaging (or CoFiMA), that selectively weighs each parameter in the weights ensemble by leveraging the Fisher information of the weights of the model. Both variants are conceptually simple, easy to implement, and effective in attaining state-of-the-art performance on several standard CL benchmarks. Code is available at: https://github.com/IemProg/CoFiMA.

Victor Quétu, Marta Milovanovic, Enzo Tartaglione

Vision transformers (ViT) have been of broad interest in recent theoretical and empirical works. They are state-of-the-art thanks to their attention-based approach, which boosts the identification of key features and patterns within images thanks to the capability of avoiding inductive bias, resulting in highly accurate image analysis. Meanwhile, neoteric studies have reported a ``sparse double descent'' phenomenon that can occur in modern deep-learning models, where extremely over-parametrized models can generalize well. This raises practical questions about the optimal size of the model and the quest over finding the best trade-off between sparsity and performance is launched: are Vision Transformers also prone to sparse double descent? Can we find a way to avoid such a phenomenon? Our work tackles the occurrence of sparse double descent on ViTs. Despite some works that have shown that traditional architectures, like Resnet, are condemned to the sparse double descent phenomenon, for ViTs we observe that an optimally-tuned $\ell_2$ regularization relieves such a phenomenon. However, everything comes at a cost: optimal lambda will sacrifice the potential compression of the ViT.

Victor Quétu, Enzo Tartaglione

Finding the optimal size of deep learning models is very actual and of broad impact, especially in energy-saving schemes. Very recently, an unexpected phenomenon, the ``double descent'', has caught the attention of the deep learning community. As the model's size grows, the performance gets first worse, and then goes back to improving. It raises serious questions about the optimal model's size to maintain high generalization: the model needs to be sufficiently over-parametrized, but adding too many parameters wastes training resources. Is it possible to find, in an efficient way, the best trade-off? Our work shows that the double descent phenomenon is potentially avoidable with proper conditioning of the learning problem, but a final answer is yet to be found. We empirically observe that there is hope to dodge the double descent in complex scenarios with proper regularization, as a simple $\ell_2$ regularization is already positively contributing to such a perspective.

Enzo Tartaglione

Deep learning models are nowadays broadly deployed to solve an incredibly large variety of tasks. Commonly, leveraging over the availability of "big data", deep neural networks are trained as black-boxes, minimizing an objective function at its output. This however does not allow control over the propagation of some specific features through the model, like gender or race, for solving some an uncorrelated task. This raises issues either in the privacy domain (considering the propagation of unwanted information) and of bias (considering that these features are potentially used to solve the given task). In this work we propose IRENE, a method to achieve information removal at the bottleneck of deep neural networks, which explicitly minimizes the estimated mutual information between the features to be kept ``private'' and the target. Experiments on a synthetic dataset and on CelebA validate the effectiveness of the proposed approach, and open the road towards the development of approaches guaranteeing information removal in deep neural networks.

Cem Eteke, Enzo Tartaglione

3D Gaussian Splatting (3DGS) revolutionized novel view rendering. Instead of inferring from dense spatial points, as implicit representations do, 3DGS uses sparse Gaussians. This enables real-time performance but increases space requirements, hindering applications such as immersive communication. 3DGS compression emerged as a field aimed at alleviating this issue. While impressive progress has been made, at low rates, compression introduces artifacts that degrade visual quality significantly. We introduce NiFi, a method for extreme 3DGS compression through restoration via artifact-aware, diffusion-based one-step distillation. We show that our method achieves state-of-the-art perceptual quality at extremely low rates, down to 0.1 MB, and towards 1000x rate improvement over 3DGS at comparable perceptual performance. The code will be open-sourced upon acceptance.

Francesco Di Sario, Riccardo Renzulli, Enzo Tartaglione et al.

Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance.

Victor Quétu, Zhu Liao, Enzo Tartaglione

While deep neural networks are highly effective at solving complex tasks, large pre-trained models are commonly employed even to solve consistently simpler downstream tasks, which do not necessarily require a large model's complexity. Motivated by the awareness of the ever-growing AI environmental impact, we propose an efficiency strategy that leverages prior knowledge transferred by large models. Simple but effective, we propose a method relying on an Entropy-bASed Importance mEtRic (EASIER) to reduce the depth of over-parametrized deep neural networks, which alleviates their computational burden. We assess the effectiveness of our method on traditional image classification setups. Our code is available at https://github.com/VGCQ/EASIER.

Gabriele Spadaro, Alberto Presta, Jhony H. Giraldo et al.

Efficient compression of low-bit-rate point clouds is critical for bandwidth-constrained applications. However, existing techniques mainly focus on high-fidelity reconstruction, requiring many bits for compression. This paper proposes a "Denoising Diffusion Probabilistic Model" (DDPM) architecture for point cloud compression (DDPM-PCC) at low bit-rates. A PointNet encoder produces the condition vector for the generation, which is then quantized via a learnable vector quantizer. This configuration allows to achieve a low bitrates while preserving quality. Experiments on ShapeNet and ModelNet40 show improved rate-distortion at low rates compared to standardized and state-of-the-art approaches. We publicly released the code at https://github.com/EIDOSLAB/DDPM-PCC.

Victor Quétu, Zhu Liao, Nour Hezbri et al.

Although deep neural networks are well-known for their outstanding performance in tackling complex tasks, their hunger for computational resources remains a significant hurdle, posing energy-consumption issues and restricting their deployment on resource-constrained devices, preventing their widespread adoption. In this paper, we present an optimal transport-based method to reduce the depth of over-parametrized deep neural networks, alleviating their computational burden. More specifically, we propose a new regularization strategy based on the Max-Sliced Wasserstein distance to minimize the distance between the intermediate feature distributions in the neural network. We show that minimizing this distance enables the complete removal of intermediate layers in the network, achieving better performance/depth trade-off compared to existing techniques. We assess the effectiveness of our method on traditional image classification setups and extend it to generative image models. Our code is available at https://github.com/VGCQ/LaCoOT.

Massimiliano Ciranni, Vito Paolo Pastore, Roberto Di Via et al.

Deep learning model effectiveness in classification tasks is often challenged by the quality and quantity of training data whenever they are affected by strong spurious correlations between specific attributes and target labels. This results in a form of bias affecting training data, which typically leads to unrecoverable weak generalization in prediction. This paper aims at facing this problem by leveraging bias amplification with generated synthetic data: we introduce Diffusing DeBias (DDB), a novel approach acting as a plug-in for common methods of unsupervised model debiasing exploiting the inherent bias-learning tendency of diffusion models in data generation. Specifically, our approach adopts conditional diffusion models to generate synthetic bias-aligned images, which replace the original training set for learning an effective bias amplifier model that we subsequently incorporate into an end-to-end and a two-step unsupervised debiasing approach. By tackling the fundamental issue of bias-conflicting training samples memorization in learning auxiliary models, typical of this type of techniques, our proposed method beats current state-of-the-art in multiple benchmark datasets, demonstrating its potential as a versatile and effective tool for tackling bias in deep learning models. Code is available at https://github.com/Malga-Vision/DiffusingDeBias

Imad Eddine Marouf, Enzo Tartaglione, Stephane Lathuiliere et al.

Continual Learning in Visual Question Answering (VQACL) requires models to acquire new visual-linguistic skills (plasticity) while preserving previously learned knowledge (stability). The inherent multimodality of VQACL exacerbates this challenge, as models must balance stability across visual and textual domains while adapting to novel objects and reasoning tasks. Existing methods, primarily designed for unimodal settings, often fall short in addressing this dual requirement. In this work, we present QUestion-only replay with Attention Distillation (QUAD), a novel approach for VQACL that leverages only past task questions for regularization. By eliminating the need to store visual data, QUAD not only reduces memory overhead, but also alleviates privacy concerns. Our method introduces a Question-only Replay mechanism that selectively reuses prior task questions to counteract overfitting to the answer space of the current task, addressing the problem out of answer set. Complementing this, we propose Attention Consistency Distillation to enforce both intra-modal and inter-modal attention consistency across tasks, preserving essential visual-linguistic associations. Extensive experiments on VQAv2 and NExT-QA demonstrate that QUAD significantly outperforms state-of-the-art methods, achieving robust performance in continual VQA. Code is available at: https://github.com/IemProg/QUAD.

Muhammad Salman Ali, Sung-Ho Bae, Enzo Tartaglione

3D models have recently been popularized by the potentiality of end-to-end training offered first by Neural Radiance Fields and most recently by 3D Gaussian Splatting models. The latter has the big advantage of naturally providing fast training convergence and high editability. However, as the research around these is still in its infancy, there is still a gap in the literature regarding the model's scalability. In this work, we propose an approach enabling both memory and computation scalability of such models. More specifically, we propose an iterative pruning strategy that removes redundant information encoded in the model. We also enhance compressibility for the model by including in the optimization strategy a differentiable quantization and entropy coding estimator. Our results on popular benchmarks showcase the effectiveness of the proposed approach and open the road to the broad deployability of such a solution even on resource-constrained devices.

Haicheng Wang, Zhemeng Yu, Gabriele Spadaro et al.

Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their abilities to generate and understand cross-modal data. However, processing long sequences of visual tokens extracted from visual backbones poses a challenge for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating both computational and memory demands during training and inference. Through a comprehensive analysis of the token reduction process, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We showcase the effectiveness of FOLDER by integrating it into the visual backbone of several MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. In both contexts, FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to 70% of visual tokens.

Aël Quélennec, Enzo Tartaglione, Pavlo Mozharovskyi et al.

In the realm of efficient on-device learning under extreme memory and computation constraints, a significant gap in successful approaches persists. Although considerable effort has been devoted to efficient inference, the main obstacle to efficient learning is the prohibitive cost of backpropagation. The resources required to compute gradients and update network parameters often exceed the limits of tightly constrained memory budgets. This paper challenges conventional wisdom and proposes a series of experiments that reveal the existence of superior sub-networks. Furthermore, we hint at the potential for substantial gains through a dynamic neuron selection strategy when fine-tuning a target task. Our efforts extend to the adaptation of a recent dynamic neuron selection strategy pioneered by Bragagnolo et al. (NEq), revealing its effectiveness in the most stringent scenarios. Our experiments demonstrate, in the average case, the superiority of a NEq-inspired approach over a random selection. This observation prompts a compelling avenue for further exploration in the area, highlighting the opportunity to design a new class of algorithms designed to facilitate parameter update selection. Our findings usher in a new era of possibilities in the field of on-device learning under extreme constraints and encourage the pursuit of innovative strategies for efficient, resource-friendly model fine-tuning.

Francesco Di Sario, Riccardo Renzulli, Marco Grangetto et al.

3D Gaussian Splatting enhances real-time performance in novel view synthesis by representing scenes with mixtures of Gaussians and utilizing differentiable rasterization. However, it typically requires large storage capacity and high VRAM, demanding the design of effective pruning and compression techniques. Existing methods, while effective in some scenarios, struggle with scalability and fail to adapt models based on critical factors such as computing capabilities or bandwidth, requiring to re-train the model under different configurations. In this work, we propose a novel, model-agnostic technique that organizes Gaussians into several hierarchical layers, enabling progressive Level of Detail (LoD) strategy. This method, combined with recent approach of compression of 3DGS, allows a single model to instantly scale across several compression ratios, with minimal to none impact to quality compared to a single non-scalable model and without requiring re-training. We validate our approach on typical datasets and benchmarks, showcasing low distortion and substantial gains in terms of scalability and adaptability.

Le-Trung Nguyen, Aël Quélennec, Enzo Tartaglione et al.

Internet of Things and Deep Learning are synergetically and exponentially growing industrial fields with a massive call for their unification into a common framework called Edge AI. While on-device inference is a well-explored topic in recent research, backpropagation remains an open challenge due to its prohibitive computational and memory costs compared to the extreme resource constraints of embedded devices. Drawing on tensor decomposition research, we tackle the main bottleneck of backpropagation, namely the memory footprint of activation map storage. We investigate and compare the effects of activation compression using Singular Value Decomposition and its tensor variant, High-Order Singular Value Decomposition. The application of low-order decomposition results in considerable memory savings while preserving the features essential for learning, and also offers theoretical guarantees to convergence. Experimental results obtained on main-stream architectures and tasks demonstrate Pareto-superiority over other state-of-the-art solutions, in terms of the trade-off between generalization and memory footprint.

Alberto Presta, Gabriele Spadaro, Enzo Tartaglione et al.

In Learned Image Compression (LIC), a model is trained at encoding and decoding images sampled from a source domain, often outperforming traditional codecs on natural images; yet its performance may be far from optimal on images sampled from different domains. In this work, we tackle the problem of adapting a pre-trained model to multiple target domains by plugging into the decoder an adapter module for each of them, including the source one. Each adapter improves the decoder performance on a specific domain, without the model forgetting about the images seen at training time. A gate network computes the weights to optimally blend the contributions from the adapters when the bitstream is decoded. We experimentally validate our method over two state-of-the-art pre-trained models, observing improved rate-distortion efficiency on the target domains without penalties on the source domain. Furthermore, the gate's ability to find similarities with the learned target domains enables better encoding efficiency also for images outside them.

Le-Trung Nguyen, Ael Quelennec, Van-Tam Nguyen et al.

On-device learning has emerged as a promising direction for AI development, particularly because of its potential to reduce latency issues and mitigate privacy risks associated with device-server communication, while improving energy efficiency. Despite these advantages, significant memory and computational constraints still represent major challenges for its deployment. Drawing on previous studies on low-rank decomposition methods that address activation memory bottlenecks in backpropagation, we propose a novel shortcut approach as an alternative. Our analysis and experiments demonstrate that our method can reduce activation memory usage, even up to $120.09\times$ compared to vanilla training, while also reducing overall training FLOPs up to $1.86\times$ when evaluated on traditional benchmarks.

Ziyu Lin, Enzo Tartaglione, Van-Tam Nguyen

On-device training is an emerging approach in machine learning where models are trained on edge devices, aiming to enhance privacy protection and real-time performance. However, edge devices typically possess restricted computational power and resources, making it challenging to perform computationally intensive model training tasks. Consequently, reducing resource consumption during training has become a pressing concern in this field. To this end, we propose SCoTTi (Save Computation at Training Time), an adaptive framework that addresses the aforementioned challenge. It leverages an optimizable threshold parameter to effectively reduce the number of neuron updates during training which corresponds to a decrease in memory and computation footprint. Our proposed approach demonstrates superior performance compared to the state-of-the-art methods regarding computational resource savings on various commonly employed benchmarks and popular architectures, including ResNets, MobileNet, and Swin-T.

Carl De Sousa Trias, Mihai Petru Mitrea, Attilio Fiandrotti et al.

Deep neural networks are characterized by multiple symmetrical, equi-loss solutions that are redundant. Thus, the order of neurons in a layer and feature maps can be given arbitrary permutations, without affecting (or minimally affecting) their output. If we shuffle these neurons, or if we apply to them some perturbations (like fine-tuning) can we put them back in the original order i.e. re-synchronize? Is there a possible corruption threat? Answering these questions is important for applications like neural network white-box watermarking for ownership tracking and integrity verification. We advance a method to re-synchronize the order of permuted neurons. Our method is also effective if neurons are further altered by parameter pruning, quantization, and fine-tuning, showing robustness to integrity attacks. Additionally, we provide theoretical and practical evidence for the usual means to corrupt the integrity of the model, resulting in a solution to counter it. We test our approach on popular computer vision datasets and models, and we illustrate the threat and our countermeasure on a popular white-box watermarking method.

Zhu Liao, Victor Quétu, Van-Tam Nguyen et al.

While deep neural networks are highly effective at solving complex tasks, their computational demands can hinder their usefulness in real-time applications and with limited-resources systems. Besides, for many tasks it is known that these models are over-parametrized: neoteric works have broadly focused on reducing the width of these networks, rather than their depth. In this paper, we aim to reduce the depth of over-parametrized deep neural networks: we propose an eNtropy-basEd Pruning as a nEural Network depTH's rEducer (NEPENTHE) to alleviate deep neural networks' computational burden. Based on our theoretical finding, NEPENTHE focuses on un-structurally pruning connections in layers with low entropy to remove them entirely. We validate our approach on popular architectures such as MobileNet and Swin-T, showing that when encountering an over-parametrization regime, it can effectively linearize some layers (hence reducing the model's depth) with little to no performance loss. The code will be publicly available upon acceptance of the article.

Aël Quélennec, Nour Hezbri, Pavlo Mozharovskyi et al.

Memory-efficient training of deep neural networks has become increasingly important as models grow larger while deployment environments impose strict resource constraints. We propose TraDy, a novel transfer learning scheme leveraging two key insights: layer importance for updates is architecture-dependent and determinable a priori, while dynamic stochastic channel selection provides superior gradient approximation compared to static approaches. We introduce a dynamic channel selection approach that stochastically resamples channels between epochs within preselected layers. Extensive experiments demonstrate TraDy achieves state-of-the-art performance across various downstream tasks and architectures while maintaining strict memory constraints, achieving up to 99% activation sparsity, 95% weight derivative sparsity, and 97% reduction in FLOPs for weight derivative computation.

Alaa Mazouz, Sumanta Chaudhuri, Marco Cagnanzzo et al.

Learnable Image Compression (LIC) has shown the potential to outperform standardized video codecs in RD efficiency, prompting the research for hardware-friendly implementations. Most existing LIC hardware implementations prioritize latency to RD-efficiency and through an extensive exploration of the hardware design space. We present a novel design paradigm where the burden of tuning the design for a specific hardware platform is shifted towards model dimensioning and without compromising on RD-efficiency. First, we design a framework for distilling a leaner student LIC model from a reference teacher: by tuning a single model hyperparameters, we can meet the constraints of different hardware platforms without a complex hardware design exploration. Second, we propose a hardware-friendly implementation of the Generalized Divisive Normalization - GDN activation that preserves RD efficiency even post parameter quantization. Third, we design a pipelined FPGA configuration which takes full advantage of available FPGA resources by leveraging parallel processing and optimizing resource allocation. Our experiments with a state of the art LIC model show that we outperform all existing FPGA implementations while performing very close to the original model.

Alberto Presta, Enzo Tartaglione, Attilio Fiandrotti et al.

Learned progressive image compression is gaining momentum as it allows improved image reconstruction as more bits are decoded at the receiver. We propose a progressive image compression method in which an image is first represented as a pair of base-quality and top-quality latent representations. Next, a residual latent representation is encoded as the element-wise difference between the top and base representations. Our scheme enables progressive image compression with element-wise granularity by introducing a masking system that ranks each element of the residual latent representation from most to least important, dividing it into complementary components, which can be transmitted separately to the decoder in order to obtain different reconstruction quality. The masking system does not add further parameters nor complexity. At the receiver, any elements of the top latent representation excluded from the transmitted components can be independently replaced with the mean predicted by the hyperprior architecture, ensuring reliable reconstructions at any intermediate quality level. We also introduced Rate Enhancement Modules (REMs), which refine the estimation of entropy parameters using already decoded components. We obtain results competitive with state-of-the-art competitors, while significantly reducing computational complexity, decoding time, and number of parameters.

Rémi Nahon, Ivan Luiz De Moura Matos, Van-Tam Nguyen et al.

Nowadays an ever-growing concerning phenomenon, the emergence of algorithmic biases that can lead to unfair models, emerges. Several debiasing approaches have been proposed in the realm of deep learning, employing more or less sophisticated approaches to discourage these models from massively employing these biases. However, a question emerges: is this extra complexity really necessary? Is a vanilla-trained model already embodying some ``unbiased sub-networks'' that can be used in isolation and propose a solution without relying on the algorithmic biases? In this work, we show that such a sub-network typically exists, and can be extracted from a vanilla-trained model without requiring additional training. We further validate that such specific architecture is incapable of learning a specific bias, suggesting that there are possible architectural countermeasures to the problem of biases in deep neural networks.

Aël Quélennec, Pavlo Mozharovskyi, Van-Tam Nguyen et al.

On-device neural network training faces critical memory constraints that limit the adaptation of pre-trained models to downstream tasks. We present MeDyate, a theoretically-grounded framework for memory-constrained dynamic subnetwork adaptation. Our approach introduces two key innovations: LaRa (Layer Ranking), an improved layer importance metric that enables principled layer pre-selection, and a dynamic channel sampling strategy that exploits the temporal stability of channel importance distributions during fine-tuning. MeDyate dynamically resamples channels between epochs according to importance-weighted probabilities, ensuring comprehensive parameter space exploration while respecting strict memory budgets. Extensive evaluation across a large panel of tasks and architectures demonstrates that MeDyate achieves state-of-the-art performance under extreme memory constraints, consistently outperforming existing static and dynamic approaches while maintaining high computational efficiency. Our method represents a significant step towards enabling efficient on-device learning by demonstrating effective fine-tuning with memory budgets as low as a few hundred kB of RAM.

Le-Trung Nguyen, Enzo Tartaglione, Van-Tam Nguyen

As AI increasingly shapes daily life, energy consumption and data privacy have become pressing concerns. On-device learning trains models directly on edge devices, cutting energy consumption and safeguarding data privacy. However, the expanding scale of modern neural networks creates a major obstacle for on-device training. Although prior work has concentrated on compact convolutional architectures, we instead apply subspace-based training to transformer models. Motivated by the idea that a model's essential information lies in a fixed subspace, we introduce Weight-Activation Subspace Iteration (WASI), a method that mitigates the memory bottleneck of backpropagation and boosts inference efficiency in transformer models by restricting training to this subspace. Our results demonstrate that WASI maintains accuracy comparable to vanilla training while reducing memory usage by up to $62\times$ and computational cost (FLOPs) by up to $2\times$. On a Raspberry Pi 5, WASI achieves roughly $1.5\times$ faster training and inference than vanilla training.

Andrea Diecidue, Carlo Alberto Barbano, Piero Fraternali et al.

Transformer-based models have become the state of the art across multiple domains, from natural language processing to machine listening, thanks to attention mechanisms. However, the attention layers require a large number of parameters and high-end hardware for both training and inference. We propose a novel pruning technique targeted explicitly at the attention mechanism, where we decouple the pruning of the four layers in the attention block, namely: query, keys, values and outputs' projection matrices. We also investigate pruning strategies to prune along the head and channel dimensions, and compare the performance of the Audio Spectrogram Transformer (AST) model under different pruning scenarios. Our results show that even by pruning 50\% of the attention parameters we incur in performance degradation of less than 1\%

Gianluca Dalmasso, Andrea Bragagnolo, Enzo Tartaglione et al.

Hyperparameter tuning, such as learning rate decay and defining a stopping criterion, often relies on monitoring the validation loss. This paper presents NeVe, a dynamic training approach that adjusts the learning rate and defines the stop criterion based on the novel notion of "neural velocity". The neural velocity measures the rate of change of each neuron's transfer function and is an indicator of model convergence: sampling neural velocity can be performed even by forwarding noise in the network, reducing the need for a held-out dataset. Our findings show the potential of neural velocity as a key metric for optimizing neural network training efficiently

Dorian Gailhard, Enzo Tartaglione, Lirida Naviner et al.

Graph generative models have shown strong results in molecular design but struggle to scale to large, complex structures. While hierarchical methods improve scalability, they usually ignore node and edge features, which are critical in real-world applications. This issue is amplified in hypergraphs, where hyperedges capture higher-order relationships among multiple nodes. Despite their importance in domains such as 3D geometry, molecular systems, and circuit design, existing generative models rarely support both hypergraphs and feature generation at scale. In this paper, we introduce FAHNES (feature-aware hypergraph generation via next-scale prediction), a hierarchical framework that jointly generates hypergraph topology and features. FAHNES builds multi-scale representations through node coarsening and refines them via localized expansion, guided by a novel node budget mechanism that controls granularity and ensures consistency across scales. Experiments on synthetic, 3D mesh and graph point cloud datasets show that FAHNES achieves state-of-the-art performance in jointly generating features and structure, advancing scalable hypergraph and graph generation.

Ali Haider, Muhammad Salman Ali, Maryam Qamar et al.

Implicit Neural Representations (INRs) have revolutionized signal processing and computer vision by modeling signals as continuous, differentiable functions parameterized by neural networks. However, their inherent formulation as a regression problem makes them prone to regression to the mean, limiting their ability to capture fine details, retain high-frequency information, and handle noise effectively. To address these challenges, we propose Iterative Implicit Neural Representations (I-INRs) a novel plug-and-play framework that enhances signal reconstruction through an iterative refinement process. I-INRs effectively recover high-frequency details, improve robustness to noise, and achieve superior reconstruction quality. Our framework seamlessly integrates with existing INR architectures, delivering substantial performance gains across various tasks. Extensive experiments show that I-INRs outperform baseline methods, including WIRE, SIREN, and Gauss, in diverse computer vision applications such as image restoration, image denoising, and object occupancy prediction.