A Survey on Knowledge Editing of Neural Networks

Deep neural networks are becoming increasingly pervasive in academia and industry, matching and surpassing human performance on a wide variety of fields and related tasks. However, just as humans, even the largest artificial neural networks make mistakes, and once-correct predictions can become invalid as the world progresses in time. Augmenting datasets with samples that account for mistakes or up-to-date information has become a common workaround in practical applications. However, the well-known phenomenon of catastrophic forgetting poses a challenge in achieving precise changes in the implicitly memorized knowledge of neural network parameters, often requiring a full model re-training to achieve desired behaviors. That is expensive, unreliable, and incompatible with the current trend of large self-supervised pre-training, making it necessary to find more efficient and effective methods for adapting neural network models to changing data. To address this need, knowledge editing is emerging as a novel area of research that aims to enable reliable, data-efficient, and fast changes to a pre-trained target model, without affecting model behaviors on previously learned tasks. In this survey, we provide a brief review of this recent artificial intelligence field of research. We first introduce the problem of editing neural networks, formalize it in a common framework and differentiate it from more notorious branches of research such as continuous learning. Next, we provide a review of the most relevant knowledge editing approaches and datasets proposed so far, grouping works under four different families: regularization techniques, meta-learning, direct model editing, and architectural strategies. Finally, we outline some intersections with other fields of research and potential directions for future works.

Ultra-low-power Range Error Mitigation for Ultra-wideband Precise Localization

Precise and accurate localization in outdoor and indoor environments is a challenging problem that currently constitutes a significant limitation for several practical applications. Ultra-wideband (UWB) localization technology represents a valuable low-cost solution to the problem. However, non-line-of-sight (NLOS) conditions and complexity of the specific radio environment can easily introduce a positive bias in the ranging measurement, resulting in highly inaccurate and unsatisfactory position estimation. In the light of this, we leverage the latest advancement in deep neural network optimization techniques and their implementation on ultra-low-power microcontrollers to introduce an effective range error mitigation solution that provides corrections in either NLOS or LOS conditions with a few mW of power. Our extensive experimentation endorses the advantages and improvements of our low-cost and power-efficient methodology.

Back-to-Bones: Rediscovering the Role of Backbones in Domain Generalization

Domain Generalization (DG) studies the capability of a deep learning model to generalize to out-of-training distributions. In the last decade, literature has been massively filled with a collection of training methodologies that claim to obtain more abstract and robust data representations to tackle domain shifts. Recent research has provided a reproducible benchmark for DG, pointing out the effectiveness of naive empirical risk minimization (ERM) over existing algorithms. Nevertheless, researchers persist in using the same outdated feature extractors, and no attention has been given to the effects of different backbones yet. In this paper, we start back to backbones proposing a comprehensive analysis of their intrinsic generalization capabilities, so far ignored by the research community. We evaluate a wide variety of feature extractors, from standard residual solutions to transformer-based architectures, finding an evident linear correlation between large-scale single-domain classification accuracy and DG capability. Our extensive experimentation shows that by adopting competitive backbones in conjunction with effective data augmentation, plain ERM outperforms recent DG solutions and achieves state-of-the-art accuracy. Moreover, our additional qualitative studies reveal that novel backbones give more similar representations to same-class samples, separating different domains in the feature space. This boost in generalization capabilities leaves marginal room for DG algorithms and suggests a new paradigm for investigating the problem, placing backbones in the spotlight and encouraging the development of consistent algorithms on top of them.

A Deep Learning Driven Algorithmic Pipeline for Autonomous Navigation in Row-Based Crops

Expensive sensors and inefficient algorithmic pipelines significantly affect the overall cost of autonomous machines. However, affordable robotic solutions are essential to practical usage, and their financial impact constitutes a fundamental requirement to employ service robotics in most fields of application. Among all, researchers in the precision agriculture domain strive to devise robust and cost-effective autonomous platforms in order to provide genuinely large-scale competitive solutions. In this article, we present a complete algorithmic pipeline for row-based crops autonomous navigation, specifically designed to cope with low-range sensors and seasonal variations. Firstly, we build on a robust data-driven methodology to generate a viable path for the autonomous machine, covering the full extension of the crop with only the occupancy grid map information of the field. Moreover, our solution leverages on latest advancement of deep learning optimization techniques and synthetic generation of data to provide an affordable solution that efficiently tackles the well-known Global Navigation Satellite System unreliability and degradation due to vegetation growing inside rows. Extensive experimentation and simulations against computer-generated environments and real-world crops demonstrated the robustness and intrinsic generalizability of our methodology that opens the possibility of highly affordable and fully autonomous machines.

Action Transformer: A Self-Attention Model for Short-Time Human Action Recognition

Deep neural networks based purely on attention have been successful across several domains, relying on minimal architectural priors from the designer. In Human Action Recognition (HAR), attention mechanisms have been primarily adopted on top of standard convolutional or recurrent layers, improving the overall generalization capability. In this work, we introduce Action Transformer (AcT), a simple, fully self-attentional architecture that consistently outperforms more elaborated networks that mix convolutional, recurrent, and attentive layers. In order to limit computational and energy requests, building on previous human action recognition research, the proposed approach exploits 2D pose representations over small temporal windows, providing a low latency solution for accurate and effective real-time performance. Moreover, we open-source MPOSE2021, a new large-scale dataset, as an attempt to build a formal training and evaluation benchmark for real-time short-time human action recognition. Extensive experimentation on MPOSE2021 with our proposed methodology and several previous architectural solutions proves the effectiveness of the AcT model and poses the base for future work on HAR.

Deep Semantic Segmentation at the Edge for Autonomous Navigation in Vineyard Rows

Precision agriculture is a fast-growing field that aims at introducing affordable and effective automation into agricultural processes. Nowadays, algorithmic solutions for navigation in vineyards require expensive sensors and high computational workloads that preclude large-scale applicability of autonomous robotic platforms in real business case scenarios. From this perspective, our novel proposed control leverages the latest advancement in machine perception and edge AI techniques to achieve highly affordable and reliable navigation inside vineyard rows with low computational and power consumption. Indeed, using a custom-trained segmentation network and a low-range RGB-D camera, we are able to take advantage of the semantic information of the environment to produce smooth trajectories and stable control in different vineyards scenarios. Moreover, the segmentation maps generated by the control algorithm itself could be directly exploited as filters for a vegetative assessment of the crop status. Extensive experimentations and evaluations against real-world data and simulated environments demonstrated the effectiveness and intrinsic robustness of our methodology.

Domain-Adversarial Training of Self-Attention Based Networks for Land Cover Classification using Multi-temporal Sentinel-2 Satellite Imagery

The increasing availability of large-scale remote sensing labeled data has prompted researchers to develop increasingly precise and accurate data-driven models for land cover and crop classification (LC&CC). Moreover, with the introduction of self-attention and introspection mechanisms, deep learning approaches have shown promising results in processing long temporal sequences in the multi-spectral domain with a contained computational request. Nevertheless, most practical applications cannot rely on labeled data, and in the field, surveys are a time consuming solution that poses strict limitations to the number of collected samples. Moreover, atmospheric conditions and specific geographical region characteristics constitute a relevant domain gap that does not allow direct applicability of a trained model on the available dataset to the area of interest. In this paper, we investigate adversarial training of deep neural networks to bridge the domain discrepancy between distinct geographical zones. In particular, we perform a thorough analysis of domain adaptation applied to challenging multi-spectral, multi-temporal data, accurately highlighting the advantages of adapting state-of-the-art self-attention based models for LC&CC to different target zones where labeled data are not available. Extensive experimentation demonstrated significant performance and generalization gain in applying domain-adversarial training to source and target regions with marked dissimilarities between the distribution of extracted features.

Efficient-CapsNet: Capsule Network with Self-Attention Routing

Deep convolutional neural networks, assisted by architectural design strategies, make extensive use of data augmentation techniques and layers with a high number of feature maps to embed object transformations. That is highly inefficient and for large datasets implies a massive redundancy of features detectors. Even though capsules networks are still in their infancy, they constitute a promising solution to extend current convolutional networks and endow artificial visual perception with a process to encode more efficiently all feature affine transformations. Indeed, a properly working capsule network should theoretically achieve higher results with a considerably lower number of parameters count due to intrinsic capability to generalize to novel viewpoints. Nevertheless, little attention has been given to this relevant aspect. In this paper, we investigate the efficiency of capsule networks and, pushing their capacity to the limits with an extreme architecture with barely 160K parameters, we prove that the proposed architecture is still able to achieve state-of-the-art results on three different datasets with only 2% of the original CapsNet parameters. Moreover, we replace dynamic routing with a novel non-iterative, highly parallelizable routing algorithm that can easily cope with a reduced number of capsules. Extensive experimentation with other capsule implementations has proved the effectiveness of our methodology and the capability of capsule networks to efficiently embed visual representations more prone to generalization.

Robust Ultra-wideband Range Error Mitigation with Deep Learning at the Edge

Ultra-wideband (UWB) is the state-of-the-art and most popular technology for wireless localization. Nevertheless, precise ranging and localization in non-line-of-sight (NLoS) conditions is still an open research topic. Indeed, multipath effects, reflections, refractions and complexity of the indoor radio environment can easily introduce a positive bias in the ranging measurement, resulting in highly inaccurate and unsatisfactory position estimation. This article proposes an efficient representation learning methodology that exploits the latest advancement in deep learning and graph optimization techniques to achieve effective ranging error mitigation at the edge. Channel Impulse Response (CIR) signals are directly exploited to extract high semantic features to estimate corrections in either NLoS or LoS conditions. Extensive experimentation with different settings and configurations have proved the effectiveness of our methodology and demonstrated the feasibility of a robust and low computational power UWB range error mitigation.

Indoor Point-to-Point Navigation with Deep Reinforcement Learning and Ultra-wideband

Indoor autonomous navigation requires a precise and accurate localization system able to guide robots through cluttered, unstructured and dynamic environments. Ultra-wideband (UWB) technology, as an indoor positioning system, offers precise localization and tracking, but moving obstacles and non-line-of-sight occurrences can generate noisy and unreliable signals. That, combined with sensors noise, unmodeled dynamics and environment changes can result in a failure of the guidance algorithm of the robot. We demonstrate how a power-efficient and low computational cost point-to-point local planner, learnt with deep reinforcement learning (RL), combined with UWB localization technology can constitute a robust and resilient to noise short-range guidance system complete solution. We trained the RL agent on a simulated environment that encapsulates the robot dynamics and task constraints and then, we tested the learnt point-to-point navigation policies in a real setting with more than two-hundred experimental evaluations using UWB localization. Our results show that the computational efficient end-to-end policy learnt in plain simulation, that directly maps low-range sensors signals to robot controls, deployed in combination with ultra-wideband noisy localization in a real environment, can provide a robust, scalable and at-the-edge low-cost navigation system solution.