Umberto Michieli

Donald Shenaj, Eros Fanì, Marco Toldo et al.

Federated Learning (FL) has recently emerged as a possible way to tackle the domain shift in real-world Semantic Segmentation (SS) without compromising the private nature of the collected data. However, most of the existing works on FL unrealistically assume labeled data in the remote clients. Here we propose a novel task (FFREEDA) in which the clients' data is unlabeled and the server accesses a source labeled dataset for pre-training only. To solve FFREEDA, we propose LADD, which leverages the knowledge of the pre-trained model by employing self-supervision with ad-hoc regularization techniques for local training and introducing a novel federated clustered aggregation scheme based on the clients' style. Our experiments show that our algorithm is able to efficiently tackle the new task outperforming existing approaches. The code is available at https://github.com/Erosinho13/LADD.

Paolo Testolina, Francesco Barbato, Umberto Michieli et al.

Accurate scene understanding from multiple sensors mounted on cars is a key requirement for autonomous driving systems. Nowadays, this task is mainly performed through data-hungry deep learning techniques that need very large amounts of data to be trained. Due to the high cost of performing segmentation labeling, many synthetic datasets have been proposed. However, most of them miss the multi-sensor nature of the data, and do not capture the significant changes introduced by the variation of daytime and weather conditions. To fill these gaps, we introduce SELMA, a novel synthetic dataset for semantic segmentation that contains more than 30K unique waypoints acquired from 24 different sensors including RGB, depth, semantic cameras and LiDARs, in 27 different atmospheric and daytime conditions, for a total of more than 20M samples. SELMA is based on CARLA, an open-source simulator for generating synthetic data in autonomous driving scenarios, that we modified to increase the variability and the diversity in the scenes and class sets, and to align it with other benchmark datasets. As shown by the experimental evaluation, SELMA allows the efficient training of standard and multi-modal deep learning architectures, and achieves remarkable results on real-world data. SELMA is free and publicly available, thus supporting open science and research.

Marco Toldo, Umberto Michieli, Pietro Zanuttigh

Deep learning models dealing with image understanding in real-world settings must be able to adapt to a wide variety of tasks across different domains. Domain adaptation and class incremental learning deal with domain and task variability separately, whereas their unified solution is still an open problem. We tackle both facets of the problem together, taking into account the semantic shift within both input and label spaces. We start by formally introducing continual learning under task and domain shift. Then, we address the proposed setup by using style transfer techniques to extend knowledge across domains when learning incremental tasks and a robust distillation framework to effectively recollect task knowledge under incremental domain shift. The devised framework (LwS, Learning with Style) is able to generalize incrementally acquired task knowledge across all the domains encountered, proving to be robust against catastrophic forgetting. Extensive experimental evaluation on multiple autonomous driving datasets shows how the proposed method outperforms existing approaches, which prove to be ill-equipped to deal with continual semantic segmentation under both task and domain shift.

Umberto Michieli, Pablo Peso Parada, Mete Ozay

Keyword Spotting (KWS) models on embedded devices should adapt fast to new user-defined words without forgetting previous ones. Embedded devices have limited storage and computational resources, thus, they cannot save samples or update large models. We consider the setup of embedded online continual learning (EOCL), where KWS models with frozen backbone are trained to incrementally recognize new words from a non-repeated stream of samples, seen one at a time. To this end, we propose Temporal Aware Pooling (TAP) which constructs an enriched feature space computing high-order moments of speech features extracted by a pre-trained backbone. Our method, TAP-SLDA, updates a Gaussian model for each class on the enriched feature space to effectively use audio representations. In experimental analyses, TAP-SLDA outperforms competitors on several setups, backbones, and baselines, bringing a relative average gain of 11.3% on the GSC dataset.

Chang Liu, Giulia Rizzoli, Francesco Barbato et al.

Catastrophic forgetting of previous knowledge is a critical issue in continual learning typically handled through various regularization strategies. However, existing methods struggle especially when several incremental steps are performed. In this paper, we extend our previous approach (RECALL) and tackle forgetting by exploiting unsupervised web-crawled data to retrieve examples of old classes from online databases. In contrast to the original methodology, which did not incorporate an assessment of web-based data, the present work proposes two advanced techniques: an adversarial approach and an adaptive threshold strategy. These methods are utilized to meticulously choose samples from web data that exhibit strong statistical congruence with the no longer available training data. Furthermore, we improved the pseudo-labeling scheme to achieve a more accurate labeling of web data that also considers classes being learned in the current step. Experimental results show that this enhanced approach achieves remarkable results, particularly when the incremental scenario spans multiple steps.

Elena Camuffo, Umberto Michieli, Simone Milani

Recent advances in autonomous robotic technologies have highlighted the growing need for precise environmental analysis. LiDAR semantic segmentation has gained attention to accomplish fine-grained scene understanding by acting directly on raw content provided by sensors. Recent solutions showed how different learning techniques can be used to improve the performance of the model, without any architectural or dataset change. Following this trend, we present a coarse-to-fine setup that LEArns from classification mistaKes (LEAK) derived from a standard model. First, classes are clustered into macro groups according to mutual prediction errors; then, the learning process is regularized by: (1) aligning class-conditional prototypical feature representation for both fine and coarse classes, (2) weighting instances with a per-class fairness index. Our LEAK approach is very general and can be seamlessly applied on top of any segmentation architecture; indeed, experimental results showed that it enables state-of-the-art performances on different architectures, datasets and tasks, while ensuring more balanced class-wise results and faster convergence.

Umberto Michieli, Mete Ozay

Vision systems mounted on home robots need to interact with unseen classes in changing environments. Robots have limited computational resources, labelled data and storage capability. These requirements pose some unique challenges: models should adapt without forgetting past knowledge in a data- and parameter-efficient way. We characterize the problem as few-shot (FS) online continual learning (OCL), where robotic agents learn from a non-repeated stream of few-shot data updating only a few model parameters. Additionally, such models experience variable conditions at test time, where objects may appear in different poses (e.g., horizontal or vertical) and environments (e.g., day or night). To improve robustness of CL agents, we propose RobOCLe, which; 1) constructs an enriched feature space computing high order statistical moments from the embedded features of samples; and 2) computes similarity between high order statistics of the samples on the enriched feature space, and predicts their class labels. We evaluate robustness of CL models to train/test augmentations in various cases. We show that different moments allow RobOCLe to capture different properties of deformations, providing higher robustness with no decrease of inference speed.

Kirill Paramonov, Jia-Xing Zhong, Umberto Michieli et al.

In this paper, we address a recent trend in robotic home appliances to include vision systems on personal devices, capable of personalizing the appliances on the fly. In particular, we formulate and address an important technical task of personal object search, which involves localization and identification of personal items of interest on images captured by robotic appliances, with each item referenced only by a few annotated images. The task is crucial for robotic home appliances and mobile systems, which need to process personal visual scenes or to operate with particular personal objects (e.g., for grasping or navigation). In practice, personal object search presents two main technical challenges. First, a robot vision system needs to be able to distinguish between many fine-grained classes, in the presence of occlusions and clutter. Second, the strict resource requirements for the on-device system restrict the usage of most state-of-the-art methods for few-shot learning and often prevent on-device adaptation. In this work, we propose Swiss DINO: a simple yet effective framework for one-shot personal object search based on the recent DINOv2 transformer model, which was shown to have strong zero-shot generalization properties. Swiss DINO handles challenging on-device personalized scene understanding requirements and does not require any adaptation training. We show significant improvement (up to 55%) in segmentation and recognition accuracy compared to the common lightweight solutions, and significant footprint reduction of backbone inference time (up to 100x) and GPU consumption (up to 10x) compared to the heavy transformer-based solutions.

Edward Fish, Umberto Michieli, Mete Ozay

Recent advancement in Automatic Speech Recognition (ASR) has produced large AI models, which become impractical for deployment in mobile devices. Model quantization is effective to produce compressed general-purpose models, however such models may only be deployed to a restricted sub-domain of interest. We show that ASR models can be personalized during quantization while relying on just a small set of unlabelled samples from the target domain. To this end, we propose myQASR, a mixed-precision quantization method that generates tailored quantization schemes for diverse users under any memory requirement with no fine-tuning. myQASR automatically evaluates the quantization sensitivity of network layers by analysing the full-precision activation values. We are then able to generate a personalised mixed-precision quantization scheme for any pre-determined memory budget. Results for large-scale ASR models show how myQASR improves performance for specific genders, languages, and speakers.

Francesco Barbato, Umberto Michieli, Jijoong Moon et al.

Recent years have seen object detection robotic systems deployed in several personal devices (e.g., home robots and appliances). This has highlighted a challenge in their design, i.e., they cannot efficiently update their knowledge to distinguish between general classes and user-specific instances (e.g., a dog vs. user's dog). We refer to this challenging task as Instance-level Personalized Object Detection (IPOD). The personalization task requires many samples for model tuning and optimization in a centralized server, raising privacy concerns. An alternative is provided by approaches based on recent large-scale Foundation Models, but their compute costs preclude on-device applications. In our work we tackle both problems at the same time, designing a Few-Shot IPOD strategy called AuXFT. We introduce a conditional coarse-to-fine few-shot learner to refine the coarse predictions made by an efficient object detector, showing that using an off-the-shelf model leads to poor personalization due to neural collapse. Therefore, we introduce a Translator block that generates an auxiliary feature space where features generated by a self-supervised model (e.g., DINOv2) are distilled without impacting the performance of the detector. We validate AuXFT on three publicly available datasets and one in-house benchmark designed for the IPOD task, achieving remarkable gains in all considered scenarios with excellent time-complexity trade-off: AuXFT reaches a performance of 80% its upper bound at just 32% of the inference time, 13% of VRAM and 19% of the model size.

Massimo Bini, Ondrej Bohdal, Umberto Michieli et al.

Memory-augmented Large Language Models (LLMs) have demonstrated remarkable consistency during prolonged dialogues by storing relevant memories and incorporating them as context. Such memory-based personalization is also key in on-device settings that allow users to keep their conversations and data private. However, memory-augmented systems typically rely on LLMs that are too costly for local on-device deployment. Even though Small Language Models (SLMs) are more suitable for on-device inference than LLMs, they cannot achieve sufficient performance. Additionally, these LLM-based systems lack native visual capabilities, limiting their applicability in multimodal contexts. In this paper, we introduce (i) MemLoRA, a novel memory system that enables local deployment by equipping SLMs with specialized memory adapters, and (ii) its vision extension MemLoRA-V, which integrates small Vision-Language Models (SVLMs) to memory systems, enabling native visual understanding. Following knowledge distillation principles, each adapter is trained separately for specific memory operations$\unicode{x2013}$knowledge extraction, memory update, and memory-augmented generation. Equipped with memory adapters, small models enable accurate on-device memory operations without cloud dependency. On text-only operations, MemLoRA outperforms 10$\times$ larger baseline models (e.g., Gemma2-27B) and achieves performance comparable to 60$\times$ larger models (e.g., GPT-OSS-120B) on the LoCoMo benchmark. To evaluate visual understanding operations instead, we extend LoCoMo with challenging Visual Question Answering tasks that require direct visual reasoning. On this, our VLM-integrated MemLoRA-V shows massive improvements over caption-based approaches (81.3 vs. 23.7 accuracy) while keeping strong performance in text-based tasks, demonstrating the efficacy of our method in multimodal contexts.

Kirill Paramonov, Mete Ozay, Aristeidis Mystakidis et al.

The proliferation of AI models in everyday devices has highlighted a critical challenge: prediction errors that degrade user experience. While existing solutions focus on error detection, they rarely provide efficient correction mechanisms, especially for resource-constrained devices. We present a novel system enabling users to correct AI misclassifications through few-shot learning, requiring minimal computational resources and storage. Our approach combines server-side foundation model training with on-device prototype-based classification, enabling efficient error correction through prototype updates rather than model retraining. The system consists of two key components: (1) a server-side pipeline that leverages knowledge distillation to transfer robust feature representations from foundation models to device-compatible architectures, and (2) a device-side mechanism that enables ultra-efficient error correction through prototype adaptation. We demonstrate our system's effectiveness on both image classification and object detection tasks, achieving over 50% error correction in one-shot scenarios on Food-101 and Flowers-102 datasets while maintaining minimal forgetting (less than 0.02%) and negligible computational overhead. Our implementation, validated through an Android demonstration app, proves the system's practicality in real-world scenarios.

Peyman Hosseini, Ondrej Bohdal, Taha Ceritli et al.

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

Elena Camuffo, Umberto Michieli, Simone Milani et al.

Developing a reliable vision system is a fundamental challenge for robotic technologies (e.g., indoor service robots and outdoor autonomous robots) which can ensure reliable navigation even in challenging environments such as adverse weather conditions (e.g., fog, rain), poor lighting conditions (e.g., over/under exposure), or sensor degradation (e.g., blurring, noise), and can guarantee high performance in safety-critical functions. Current solutions proposed to improve model robustness usually rely on generic data augmentation techniques or employ costly test-time adaptation methods. In addition, most approaches focus on addressing a single vision task (typically, image recognition) utilising synthetic data. In this paper, we introduce Per-corruption Adaptation of Normalization statistics (PAN) to enhance the model robustness of vision systems. Our approach entails three key components: (i) a corruption type identification module, (ii) dynamic adjustment of normalization layer statistics based on identified corruption type, and (iii) real-time update of these statistics according to input data. PAN can integrate seamlessly with any convolutional model for enhanced accuracy in several robot vision tasks. In our experiments, PAN obtains robust performance improvement on challenging real-world corrupted image datasets (e.g., OpenLoris, ExDark, ACDC), where most of the current solutions tend to fail. Moreover, PAN outperforms the baseline models by 20-30% on synthetic benchmarks in object recognition tasks.

Donald Shenaj, Ondrej Bohdal, Mete Ozay et al.

Recent advancements in image generation models have enabled personalized image creation with both user-defined subjects (content) and styles. Prior works achieved personalization by merging corresponding low-rank adapters (LoRAs) through optimization-based methods, which are computationally demanding and unsuitable for real-time use on resource-constrained devices like smartphones. To address this, we introduce LoRA$.$rar, a method that not only improves image quality but also achieves a remarkable speedup of over $4000\times$ in the merging process. We collect a dataset of style and subject LoRAs and pre-train a hypernetwork on a diverse set of content-style LoRA pairs, learning an efficient merging strategy that generalizes to new, unseen content-style pairs, enabling fast, high-quality personalization. Moreover, we identify limitations in existing evaluation metrics for content-style quality and propose a new protocol using multimodal large language models (MLLMs) for more accurate assessment. Our method significantly outperforms the current state of the art in both content and style fidelity, as validated by MLLM assessments and human evaluations.

Emanuele Aiello, Umberto Michieli, Diego Valsesia et al.

Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models.

Umberto Michieli, Mete Ozay

Continual Learning (CL) aims to learn a sequence of problems (i.e., tasks and domains) by transferring knowledge acquired on previous problems, whilst avoiding forgetting of past ones. Different from previous approaches which focused on CL for one NLP task or domain in a specific use-case, in this paper, we address a more general CL setting to learn from a sequence of problems in a unique framework. Our method, HOP, permits to hop across tasks and domains by addressing the CL problem along three directions: (i) we employ a set of adapters to generalize a large pre-trained model to unseen problems, (ii) we compute high-order moments over the distribution of embedded representations to distinguish independent and correlated statistics across different tasks and domains, (iii) we process this enriched information with auxiliary heads specialized for each end problem. Extensive experimental campaign on 4 NLP applications, 5 benchmarks and 2 CL setups demonstrates the effectiveness of our HOP.

Taha Ceritli, Ondrej Bohdal, Mete Ozay et al.

Large language models (LLMs) often leverage adapters, such as low-rank-based adapters, to achieve strong performance on downstream tasks. However, storing a separate adapter for each task significantly increases memory requirements, posing a challenge for resource-constrained environments such as mobile devices. Although model merging techniques can reduce storage costs, they typically result in substantial performance degradation. In this work, we introduce HydraOpt, a new model merging technique that capitalizes on the inherent similarities between the matrices of low-rank adapters. Unlike existing methods that produce a fixed trade-off between storage size and performance, HydraOpt allows us to navigate this spectrum of efficiency and performance. Our experiments show that HydraOpt significantly reduces storage size (48% reduction) compared to storing all adapters, while achieving competitive performance (0.2-1.8% drop). Furthermore, it outperforms existing merging techniques in terms of performance at the same or slightly worse storage efficiency.

Yingtao Zhang, Diego Cerretti, Jialin Zhao et al.

Dynamic sparse training (DST) can reduce the computational demands in ANNs, but faces difficulties in keeping peak performance at high sparsity levels. The Cannistraci-Hebb training (CHT) is a brain-inspired method for growing connectivity in DST. CHT leverages a gradient-free, topology-driven link regrowth, which has shown ultra-sparse (less than 1% connectivity) advantage across various tasks compared to fully connected networks. Yet, CHT suffers two main drawbacks: (i) its time complexity is $O(Nd^3)$ - N node network size, d node degree - restricting it to ultra-sparse regimes. (ii) it selects top link prediction scores, which is inappropriate for the early training epochs, when the network presents unreliable connections. Here, we design the first brain-inspired network model - termed bipartite receptive field (BRF) - to initialize the connectivity of sparse artificial neural networks. We further introduce a GPU-friendly matrix-based approximation of CH link prediction, reducing complexity to $O(N^3)$. We introduce the Cannistraci-Hebb training soft rule (CHTs), which adopts a flexible strategy for sampling connections in both link removal and regrowth, balancing the exploration and exploitation of network topology. Additionally, we integrate CHTs with a sigmoid gradual density decay (CHTss). Empirical results show that BRF offers performance advantages over previous network science models. Using 1% of connections, CHTs outperforms fully connected networks in MLP architectures on image classification tasks, compressing some networks to less than 30% of the nodes. Using 5% of the connections, CHTss outperforms fully connected networks in two Transformer-based machine translation tasks. Finally, at 30% connectivity, both CHTs and CHTss outperform other DST methods in language modeling and even exceed fully connected baselines in zero-shot tasks.

Umberto Michieli, Jijoong Moon, Daehyun Kim et al.

Nowadays, users demand for increased personalization of vision systems to localize and identify personal instances of objects (e.g., my dog rather than dog) from a few-shot dataset only. Despite outstanding results of deep networks on classical label-abundant benchmarks (e.g., those of the latest YOLOv8 model for standard object detection), they struggle to maintain within-class variability to represent different instances rather than object categories only. We construct an Object-conditioned Bag of Instances (OBoI) based on multi-order statistics of extracted features, where generic object detection models are extended to search and identify personal instances from the OBoI's metric space, without need for backpropagation. By relying on multi-order statistics, OBoI achieves consistent superior accuracy in distinguishing different instances. In the results, we achieve 77.1% personal object recognition accuracy in case of 18 personal instances, showing about 12% relative gain over the state of the art.

Elena Camuffo, Umberto Michieli, Jijoong Moon et al.

Improving model robustness in case of corrupted images is among the key challenges to enable robust vision systems on smart devices, such as robotic agents. Particularly, robust test-time performance is imperative for most of the applications. This paper presents a novel approach to improve robustness of any classification model, especially on severely corrupted images. Our method (FROST) employs high-frequency features to detect input image corruption type, and select layer-wise feature normalization statistics. FROST provides the state-of-the-art results for different models and datasets, outperforming competitors on ImageNet-C by up to 37.1% relative gain, improving baseline of 40.9% mCE on severe corruptions.

Hafiz Tiomoko Ali, Umberto Michieli, Ji Joong Moon et al.

The recently discovered Neural collapse (NC) phenomenon states that the last-layer weights of Deep Neural Networks (DNN), converge to the so-called Equiangular Tight Frame (ETF) simplex, at the terminal phase of their training. This ETF geometry is equivalent to vanishing within-class variability of the last layer activations. Inspired by NC properties, we explore in this paper the transferability of DNN models trained with their last layer weight fixed according to ETF. This enforces class separation by eliminating class covariance information, effectively providing implicit regularization. We show that DNN models trained with such a fixed classifier significantly improve transfer performance, particularly on out-of-domain datasets. On a broad range of fine-grained image classification datasets, our approach outperforms i) baseline methods that do not perform any covariance regularization (up to 22%), as well as ii) methods that explicitly whiten covariance of activations throughout training (up to 19%). Our findings suggest that DNNs trained with fixed ETF classifiers offer a powerful mechanism for improving transfer learning across domains.

Francesco Barbato, Umberto Michieli, Mehmet Kerim Yucel et al.

In multimedia understanding tasks, corrupted samples pose a critical challenge, because when fed to machine learning models they lead to performance degradation. In the past, three groups of approaches have been proposed to handle noisy data: i) enhancer and denoiser modules to improve the quality of the noisy data, ii) data augmentation approaches, and iii) domain adaptation strategies. All the aforementioned approaches come with drawbacks that limit their applicability; the first has high computational costs and requires pairs of clean-corrupted data for training, while the others only allow deployment of the same task/network they were trained on (\ie, when upstream and downstream task/network are the same). In this paper, we propose SyMPIE to solve these shortcomings. To this end, we design a small, modular, and efficient (just 2GFLOPs to process a Full HD image) system to enhance input data for robust downstream multimedia understanding with minimal computational cost. Our SyMPIE is pre-trained on an upstream task/network that should not match the downstream ones and does not need paired clean-corrupted samples. Our key insight is that most input corruptions found in real-world tasks can be modeled through global operations on color channels of images or spatial filters with small kernels. We validate our approach on multiple datasets and tasks, such as image classification (on ImageNetC, ImageNetC-Bar, VizWiz, and a newly proposed mixed corruption benchmark named ImageNetC-mixed) and semantic segmentation (on Cityscapes, ACDC, and DarkZurich) with consistent improvements of about 5\% relative accuracy gain across the board. The code of our approach and the new ImageNetC-mixed benchmark will be made available upon publication.

Kirill Paramonov, Mete Ozay, Eunju Yang et al.

Class-incremental learning in the context of limited personal labeled samples (few-shot) is critical for numerous real-world applications, such as smart home devices. A key challenge in these scenarios is balancing the trade-off between adapting to new, personalized classes and maintaining the performance of the model on the original, base classes. Fine-tuning the model on novel classes often leads to the phenomenon of catastrophic forgetting, where the accuracy of base classes declines unpredictably and significantly. In this paper, we propose a simple yet effective mechanism to address this challenge by controlling the trade-off between novel and base class accuracy. We specifically target the ultra-low-shot scenario, where only a single example is available per novel class. Our approach introduces a Novel Class Detection (NCD) rule, which adjusts the degree of forgetting a priori while simultaneously enhancing performance on novel classes. We demonstrate the versatility of our solution by applying it to state-of-the-art Few-Shot Class-Incremental Learning (FSCIL) methods, showing consistent improvements across different settings. To better quantify the trade-off between novel and base class performance, we introduce new metrics: NCR@2FOR and NCR@5FOR. Our approach achieves up to a 30% improvement in novel class accuracy on the CIFAR100 dataset (1-shot, 1 novel class) while maintaining a controlled base class forgetting rate of 2%.

Jack Foster, Kirill Paramonov, Mete Ozay et al.

Few-shot class-incremental learning (FSCIL) is a paradigm where a model, initially trained on a dataset of base classes, must adapt to an expanding problem space by recognizing novel classes with limited data. We focus on the challenging FSCIL setup where a model receives only a single sample (1-shot) for each novel class and no further training or model alterations are allowed after the base training phase. This makes generalization to novel classes particularly difficult. We propose a novel approach predicated on the hypothesis that base and novel class embeddings have structural similarity. We map the original embedding space into a residual space by subtracting the class prototype (i.e., the average class embedding) of input samples. Then, we leverage generative modeling with VAE or diffusion models to learn the multi-modal distribution of residuals over the base classes, and we use this as a valuable structural prior to improve recognition of novel classes. Our approach, Gen1S, consistently improves novel class recognition over the state of the art across multiple benchmarks and backbone architectures.

Donald Shenaj, Ondrej Bohdal, Taha Ceritli et al.

On-device deployment of Large Language Models (LLMs) frequently leverages Low-Rank Adapters (LoRAs) to support diverse downstream tasks under tight resource constraints. To address the limited storage capacity of mobile devices, recent works have explored model merging techniques to fuse multiple LoRAs into a single one. In practice, however, LoRAs are often delivered incrementally, as users request support for new tasks (e.g., novel problem types or languages). This scenario introduces a new challenge: on-device online continual merging, where the objective is to incorporate new LoRAs while preserving the performance on previously supported tasks. In this paper, we propose a data-free and computationally efficient strategy for selecting and merging LoRAs when a new one becomes available, assuming the device can store only a limited number of adapters. Extensive experiments across real-world tasks demonstrate the superiority of our approach compared to alternative strategies while adhering to the storage budget and compute limitations of on-device settings.

Ondrej Bohdal, Konstantinos Theodosiadis, Asterios Mpatziakas et al.

Large language models (LLMs) are commonly adapted for diverse downstream tasks via parameter-efficient fine-tuning techniques such as Low-Rank Adapters (LoRA). While adapters can be combined to handle multiple tasks separately, standard approaches struggle when targeting the simultaneous execution of complex tasks, such as generating a translated summary from a long conversation. To address this challenge, we propose a novel approach tailored specifically for compositional multi-tasking scenarios involving summarization and translation. Our technique involves adding a learnable projection layer on top of the combined summarization and translation adapters. This design enables effective integration while maintaining efficiency through reduced computational overhead compared to alternative strategies requiring extensive retraining or sequential processing. We demonstrate the practical viability of our method within an on-device environment by developing an Android app capable of executing compositional tasks seamlessly. Experimental results indicate our solution performs well and is fast in both cloud-based and on-device implementations, highlighting the potential benefits of adopting our framework in real-world applications demanding high-speed operation alongside resource constraints.

Elena Camuffo, Francesco Barbato, Mete Ozay et al.

We introduce MOCHA (Multi-modal Objects-aware Cross-arcHitecture Alignment), a knowledge distillation approach that transfers region-level multimodal semantics from a large vision-language teacher (e.g., LLaVa) into a lightweight vision-only object detector student (e.g., YOLO). A translation module maps student features into a joint space, where the training of the student and translator is guided by a dual-objective loss that enforces both local alignment and global relational consistency. Unlike prior approaches focused on dense or global alignment, MOCHA operates at the object level, enabling efficient transfer of semantics without modifying the teacher or requiring textual input at inference. We validate our method across four personalized detection benchmarks under few-shot regimes. Results show consistent gains over baselines, with a +10.1 average score improvement. Despite its compact architecture, MOCHA reaches performance on par with larger multimodal models, proving its suitability for real-world deployment.

Matteo Caligiuri, Francesco Barbato, Donald Shenaj et al.

Federated Learning (FL) is an established paradigm for training deep learning models on decentralized data. However, as the size of the models grows, conventional FL approaches often require significant computational resources on client devices, which may not be feasible. We introduce FedPromo, a novel framework that enables efficient adaptation of large-scale foundation models stored on a central server to new domains encountered only by remote clients. Instead of directly training the large model on client devices, FedPromo optimizes lightweight proxy models via FL, significantly reducing computational overhead while maintaining privacy. Our method follows a two-stage process: first, server-side knowledge distillation aligns the representations of a large-scale foundation model (e.g., a transformer) with those of a compact counterpart (e.g., a CNN). Then, the compact model encoder is deployed to client devices, where trainable classifiers are learned locally. These classifiers are subsequently aggregated and seamlessly transferred back to the foundation model, facilitating personalized adaptation without requiring direct access to user data. Through novel regularization strategies, our framework enables decentralized multi-domain learning, balancing performance, privacy, and resource efficiency. Extensive experiments on five image classification benchmarks demonstrate that FedPromo outperforms existing methods while assuming limited-resource clients.

Ondrej Bohdal, Mete Ozay, Jijoong Moon et al.

Adapter parameters provide a mechanism to modify the behavior of machine learning models and have gained significant popularity in the context of large language models (LLMs) and generative AI. These parameters can be merged to support multiple tasks via a process known as task merging. However, prior work on merging in LLMs, particularly in natural language processing, has been limited to scenarios where each test example addresses only a single task. In this paper, we focus on on-device settings and study the problem of text-based compositional multi-tasking, where each test example involves the simultaneous execution of multiple tasks. For instance, generating a translated summary of a long text requires solving both translation and summarization tasks concurrently. To facilitate research in this setting, we propose a benchmark comprising four practically relevant compositional tasks. We also present an efficient method (Learnable Calibration) tailored for on-device applications, where computational resources are limited, emphasizing the need for solutions that are both resource-efficient and high-performing. Our contributions lay the groundwork for advancing the capabilities of LLMs in real-world multi-tasking scenarios, expanding their applicability to complex, resource-constrained use cases.

Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio Pizzati et al.

Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model, retaining the expertise of the original ones. However, current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. This work investigates the effects of model merging on alignment. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment. We propose a simple two-step approach to address this problem: (i) generating synthetic safety and domain-specific data, and (ii) incorporating these generated data into the optimization process of existing data-aware model merging techniques. This allows us to treat alignment as a skill that can be maximized in the resulting merged LLM. Our experiments illustrate the effectiveness of integrating alignment-related data during merging, resulting in models that excel in both domain expertise and alignment.

Donald Shenaj, Francesco Barbato, Umberto Michieli et al.

Deep neural networks are typically trained in a single shot for a specific task and data distribution, but in real world settings both the task and the domain of application can change. The problem becomes even more challenging in dense predictive tasks, such as semantic segmentation, and furthermore most approaches tackle the two problems separately. In this paper we introduce the novel task of coarse-to-fine learning of semantic segmentation architectures in presence of domain shift. We consider subsequent learning stages progressively refining the task at the semantic level; i.e., the finer set of semantic labels at each learning step is hierarchically derived from the coarser set of the previous step. We propose a new approach (CCDA) to tackle this scenario. First, we employ the maximum squares loss to align source and target domains and, at the same time, to balance the gradients between well-classified and harder samples. Second, we introduce a novel coarse-to-fine knowledge distillation constraint to transfer network capabilities acquired on a coarser set of labels to a set of finer labels. Finally, we design a coarse-to-fine weight initialization rule to spread the importance from each coarse class to the respective finer classes. To evaluate our approach, we design two benchmarks where source knowledge is extracted from the GTA5 dataset and it is transferred to either the Cityscapes or the IDD datasets, and we show how it outperforms the main competitors.

Andrea Maracani, Umberto Michieli, Marco Toldo et al.

Deep networks allow to obtain outstanding results in semantic segmentation, however they need to be trained in a single shot with a large amount of data. Continual learning settings where new classes are learned in incremental steps and previous training data is no longer available are challenging due to the catastrophic forgetting phenomenon. Existing approaches typically fail when several incremental steps are performed or in presence of a distribution shift of the background class. We tackle these issues by recreating no longer available data for the old classes and outlining a content inpainting scheme on the background class. We propose two sources for replay data. The first resorts to a generative adversarial network to sample from the class space of past learning steps. The second relies on web-crawled data to retrieve images containing examples of old classes from online databases. In both scenarios no samples of past steps are stored, thus avoiding privacy concerns. Replay data are then blended with new samples during the incremental steps. Our approach, RECALL, outperforms state-of-the-art methods.

Francesco Barbato, Umberto Michieli, Marco Toldo et al.

Deep learning models obtain impressive accuracy in road scenes understanding, however they need a large quantity of labeled samples for their training. Additionally, such models do not generalise well to environments where the statistical properties of data do not perfectly match those of training scenes, and this can be a significant problem for intelligent vehicles. Hence, domain adaptation approaches have been introduced to transfer knowledge acquired on a label-abundant source domain to a related label-scarce target domain. In this work, we design and carefully analyse multiple latent space-shaping regularisation strategies that work together to reduce the domain shift. More in detail, we devise a feature clustering strategy to increase domain alignment, a feature perpendicularity constraint to space apart features belonging to different semantic classes, including those not present in the current batch, and a feature norm alignment strategy to separate active and inactive channels. In addition, we propose a novel evaluation metric to capture the relative performance of an adapted model with respect to supervised training. We validate our framework in driving scenarios, considering both synthetic-to-real and real-to-real adaptation, outperforming previous feature-level state-of-the-art methods on multiple road scenes benchmarks.

Umberto Michieli, Mete Ozay

Federated Learning (FL) is a framework which enables distributed model training using a large corpus of decentralized training data. Existing methods aggregate models disregarding their internal representations, which are crucial for training models in vision tasks. System and statistical heterogeneity (e.g., highly imbalanced and non-i.i.d. data) further harm model training. To this end, we introduce a method, called FedProto, which computes client deviations using margins of prototypical representations learned on distributed data, and applies them to drive federated optimization via an attention mechanism. In addition, we propose three methods to analyse statistical properties of feature representations learned in FL, in order to elucidate the relationship between accuracy, margins and feature discrepancy of FL models. In experimental analyses, FedProto demonstrates state-of-the-art accuracy and convergence rate across image classification and semantic segmentation benchmarks by enabling maximum margin training of FL models. Moreover, FedProto reduces uncertainty of predictions of FL models compared to the baseline. To our knowledge, this is the first work evaluating FL models in dense prediction tasks, such as semantic segmentation.

Francesco Barbato, Marco Toldo, Umberto Michieli et al.

Deep convolutional neural networks for semantic segmentation achieve outstanding accuracy, however they also have a couple of major drawbacks: first, they do not generalize well to distributions slightly different from the one of the training data; second, they require a huge amount of labeled data for their optimization. In this paper, we introduce feature-level space-shaping regularization strategies to reduce the domain discrepancy in semantic segmentation. In particular, for this purpose we jointly enforce a clustering objective, a perpendicularity constraint and a norm alignment goal on the feature vectors corresponding to source and target samples. Additionally, we propose a novel measure able to capture the relative efficacy of an adaptation strategy compared to supervised training. We verify the effectiveness of such methods in the autonomous driving setting achieving state-of-the-art results in multiple synthetic-to-real road scenes benchmarks.

Umberto Michieli, Pietro Zanuttigh

Deep neural networks suffer from the major limitation of catastrophic forgetting old tasks when learning new ones. In this paper we focus on class incremental continual learning in semantic segmentation, where new categories are made available over time while previous training data is not retained. The proposed continual learning scheme shapes the latent space to reduce forgetting whilst improving the recognition of novel classes. Our framework is driven by three novel components which we also combine on top of existing techniques effortlessly. First, prototypes matching enforces latent space consistency on old classes, constraining the encoder to produce similar latent representation for previously seen classes in the subsequent steps. Second, features sparsification allows to make room in the latent space to accommodate novel classes. Finally, contrastive learning is employed to cluster features according to their semantics while tearing apart those of different classes. Extensive evaluation on the Pascal VOC2012 and ADE20K datasets demonstrates the effectiveness of our approach, significantly outperforming state-of-the-art methods.

Marco Toldo, Umberto Michieli, Pietro Zanuttigh

Deep learning frameworks allowed for a remarkable advancement in semantic segmentation, but the data hungry nature of convolutional networks has rapidly raised the demand for adaptation techniques able to transfer learned knowledge from label-abundant domains to unlabeled ones. In this paper we propose an effective Unsupervised Domain Adaptation (UDA) strategy, based on a feature clustering method that captures the different semantic modes of the feature distribution and groups features of the same class into tight and well-separated clusters. Furthermore, we introduce two novel learning objectives to enhance the discriminative clustering performance: an orthogonality loss forces spaced out individual representations to be orthogonal, while a sparsity loss reduces class-wise the number of active feature channels. The joint effect of these modules is to regularize the structure of the feature space. Extensive evaluations in the synthetic-to-real scenario show that we achieve state-of-the-art performance.

Umberto Michieli, Edoardo Borsato, Luca Rossi et al.

The semantic segmentation of parts of objects in the wild is a challenging task in which multiple instances of objects and multiple parts within those objects must be detected in the scene. This problem remains nowadays very marginally explored, despite its fundamental importance towards detailed object understanding. In this work, we propose a novel framework combining higher object-level context conditioning and part-level spatial relationships to address the task. To tackle object-level ambiguity, a class-conditioning module is introduced to retain class-level semantics when learning parts-level semantics. In this way, mid-level features carry also this information prior to the decoding stage. To tackle part-level ambiguity and localization we propose a novel adjacency graph-based module that aims at matching the relative spatial relationships between ground truth and predicted parts. The experimental evaluation on the Pascal-Part dataset shows that we achieve state-of-the-art results on this task.

Marco Toldo, Andrea Maracani, Umberto Michieli et al.

The aim of this paper is to give an overview of the recent advancements in the Unsupervised Domain Adaptation (UDA) of deep networks for semantic segmentation. This task is attracting a wide interest, since semantic segmentation models require a huge amount of labeled data and the lack of data fitting specific requirements is the main limitation in the deployment of these techniques. This problem has been recently explored and has rapidly grown with a large number of ad-hoc approaches. This motivates us to build a comprehensive overview of the proposed methodologies and to provide a clear categorization. In this paper, we start by introducing the problem, its formulation and the various scenarios that can be considered. Then, we introduce the different levels at which adaptation strategies may be applied: namely, at the input (image) level, at the internal features representation and at the output level. Furthermore, we present a detailed overview of the literature in the field, dividing previous methods based on the following (non mutually exclusive) categories: adversarial learning, generative-based, analysis of the classifier discrepancies, self-teaching, entropy minimization, curriculum learning and multi-task learning. Novel research directions are also briefly introduced to give a hint of interesting open problems in the field. Finally, a comparison of the performance of the various methods in the widely used autonomous driving scenario is presented.

Teo Spadotto, Marco Toldo, Umberto Michieli et al.

Unsupervised Domain Adaptation (UDA) aims at improving the generalization capability of a model trained on a source domain to perform well on a target domain for which no labeled data is available. In this paper, we consider the semantic segmentation of urban scenes and we propose an approach to adapt a deep neural network trained on synthetic data to real scenes addressing the domain shift between the two different data distributions. We introduce a novel UDA framework where a standard supervised loss on labeled synthetic data is supported by an adversarial module and a self-training strategy aiming at aligning the two domain distributions. The adversarial module is driven by a couple of fully convolutional discriminators dealing with different domains: the first discriminates between ground truth and generated maps, while the second between segmentation maps coming from synthetic or real world data. The self-training module exploits the confidence estimated by the discriminators on unlabeled data to select the regions used to reinforce the learning process. Furthermore, the confidence is thresholded with an adaptive mechanism based on the per-class overall confidence. Experimental results prove the effectiveness of the proposed strategy in adapting a segmentation network trained on synthetic datasets like GTA5 and SYNTHIA, to real world datasets like Cityscapes and Mapillary.

Marco Toldo, Umberto Michieli, Gianluca Agresti et al.

The supervised training of deep networks for semantic segmentation requires a huge amount of labeled real world data. To solve this issue, a commonly exploited workaround is to use synthetic data for training, but deep networks show a critical performance drop when analyzing data with slightly different statistical properties with respect to the training set. In this work, we propose a novel Unsupervised Domain Adaptation (UDA) strategy to address the domain shift issue between real world and synthetic representations. An adversarial model, based on the cycle consistency framework, performs the mapping between the synthetic and real domain. The data is then fed to a MobileNet-v2 architecture that performs the semantic segmentation task. An additional couple of discriminators, working at the feature level of the MobileNet-v2, allows to better align the features of the two domain distributions and to further improve the performance. Finally, the consistency of the semantic maps is exploited. After an initial supervised training on synthetic data, the whole UDA architecture is trained end-to-end considering all its components at once. Experimental results show how the proposed strategy is able to obtain impressive performance in adapting a segmentation network trained on synthetic data to real world scenarios. The usage of the lightweight MobileNet-v2 architecture allows its deployment on devices with limited computational resources as the ones employed in autonomous vehicles.

Umberto Michieli, Pietro Zanuttigh

Deep learning architectures have shown remarkable results in scene understanding problems, however they exhibit a critical drop of performances when they are required to learn incrementally new tasks without forgetting old ones. This catastrophic forgetting phenomenon impacts on the deployment of artificial intelligence in real world scenarios where systems need to learn new and different representations over time. Current approaches for incremental learning deal only with image classification and object detection tasks, while in this work we formally introduce incremental learning for semantic segmentation. We tackle the problem applying various knowledge distillation techniques on the previous model. In this way, we retain the information about learned classes, whilst updating the current model to learn the new ones. We developed four main methodologies of knowledge distillation working on both output layers and internal feature representations. We do not store any image belonging to previous training stages and only the last model is used to preserve high accuracy on previously learned classes. Extensive experimental results on the Pascal VOC2012 and MSRC-v2 datasets show the effectiveness of the proposed approaches in several incremental learning scenarios.

Umberto Michieli, Matteo Biasetton, Gianluca Agresti et al.

Deep learning techniques have been widely used in autonomous driving systems for the semantic understanding of urban scenes. However, they need a huge amount of labeled data for training, which is difficult and expensive to acquire. A recently proposed workaround is to train deep networks using synthetic data, but the domain shift between real world and synthetic representations limits the performance. In this work, a novel Unsupervised Domain Adaptation (UDA) strategy is introduced to solve this issue. The proposed learning strategy is driven by three components: a standard supervised learning loss on labeled synthetic data; an adversarial learning module that exploits both labeled synthetic data and unlabeled real data; finally, a self-teaching strategy applied to unlabeled data. The last component exploits a region growing framework guided by the segmentation confidence. Furthermore, we weighted this component on the basis of the class frequencies to enhance the performance on less common classes. Experimental results prove the effectiveness of the proposed strategy in adapting a segmentation network trained on synthetic datasets, like GTA5 and SYNTHIA, to real world datasets like Cityscapes and Mapillary.

Umberto Michieli, Pietro Zanuttigh

Deep learning architectures exhibit a critical drop of performance due to catastrophic forgetting when they are required to incrementally learn new tasks. Contemporary incremental learning frameworks focus on image classification and object detection while in this work we formally introduce the incremental learning problem for semantic segmentation in which a pixel-wise labeling is considered. To tackle this task we propose to distill the knowledge of the previous model to retain the information about previously learned classes, whilst updating the current model to learn the new ones. We propose various approaches working both on the output logits and on intermediate features. In opposition to some recent frameworks, we do not store any image from previously learned classes and only the last model is needed to preserve high accuracy on these classes. The experimental evaluation on the Pascal VOC2012 dataset shows the effectiveness of the proposed approaches.

Umberto Michieli

Who are the most significant players in the history of men tennis? Is the official ATP ranking system fair in evaluating players scores? Which players deserved the most contemplation looking at their match records? Which players have never faced yet and are likely to play against in the future? Those are just some of the questions developed in this paper supported by data updated at April 2018. In order to give an answer to the aforementioned questions, complex network science techniques have been applied to some representations of the network of men singles tennis matches. Additionally, a new predictive algorithm is proposed in order to forecast the winner of a match.