Recent blind super-resolution (SR) methods typically consist of two branches, one for degradation prediction and the other for conditional restoration. However, our experiments show that a one-branch network can achieve comparable performance to the two-branch scheme. Then we wonder: how can one-branch networks automatically learn to distinguish degradations? To find the answer, we propose a new diagnostic tool -- Filter Attribution method based on Integral Gradient (FAIG). Unlike previous integral gradient methods, our FAIG aims at finding the most discriminative filters instead of input pixels/features for degradation removal in blind SR networks. With the discovered filters, we further develop a simple yet effective method to predict the degradation of an input image. Based on FAIG, we show that, in one-branch blind SR networks, 1) we are able to find a very small number of (1%) discriminative filters for each specific degradation; 2) The weights, locations and connections of the discovered filters are all important to determine the specific network function. 3) The task of degradation prediction can be implicitly realized by these discriminative filters without explicit supervised learning. Our findings can not only help us better understand network behaviors inside one-branch blind SR networks, but also provide guidance on designing more efficient architectures and diagnosing networks for blind SR.

Learning-based methods are growing prominence for planning purposes. However, there are very few approaches for learning-assisted constrained path-planning on graphs, while there are multiple downstream practical applications. This is the case for constrained path-planning for Autonomous Unmanned Ground Vehicles (AUGV), typically deployed in disaster relief or search and rescue applications. In off-road environments, the AUGV must dynamically optimize a source-destination path under various operational constraints, out of which several are difficult to predict in advance and need to be addressed on-line. We propose a hybrid solving planner that combines machine learning models and an optimal solver. More specifically, a graph convolutional network (GCN) is used to assist a branch and bound (B&B) algorithm in handling the constraints. We conduct experiments on realistic scenarios and show that GCN support enables substantial speedup and smoother scaling to harder problems.

Humans effortlessly solve pushing tasks in everyday life but unlocking these capabilities remains a challenge in robotics because physics models of these tasks are often inaccurate or unattainable. State-of-the-art data-driven approaches learn to compensate for these inaccuracies or replace the approximated physics models altogether. Nevertheless, approaches like Deep Q-Networks (DQNs) suffer from local optima in large state-action spaces. Furthermore, they rely on well-chosen deep learning architectures and learning paradigms. In this paper, we propose to frame the learning of pushing policies (where to push and how) by DQNs as an image-to-image translation problem and exploit an Hourglass-based architecture. We present an architecture combining a predictor of which pushes lead to changes in the environment with a state-action value predictor dedicated to the pushing task. Moreover, we investigate positional information encoding to learn position-dependent policy behaviors. We demonstrate in simulation experiments with a UR5 robot arm that our overall architecture helps the DQN learn faster and achieve higher performance in a pushing task involving objects with unknown dynamics.

This thesis explores the benefits machine learning algorithms can bring to online planning and scheduling for autonomous vehicles in off-road situations. Mainly, we focus on typical problems of interest which include computing itineraries that meet certain objectives, as well as computing scheduling strategies to execute synchronized maneuvers with other vehicles. We present a range of learning-based heuristics to assist different itinerary planners. We show that these heuristics allow a significant increase in performance for optimal planners. Furthermore, in the case of approximate planning, we show that not only does the running time decrease, the quality of the itinerary found also becomes almost always better. Finally, in order to synthesize strategies to execute synchronized maneuvers, we propose a novel type of scheduling controllability and a learning-assisted algorithm. The proposed framework achieves significant improvement on known benchmarks in this controllability type over the performance of state-of-the-art works in a related controllability type. Moreover, it is able to find strategies on complex scheduling problems for which previous works fail to do so.

Planning for Autonomous Unmanned Ground Vehicles (AUGV) is still a challenge, especially in difficult, off-road, critical situations. Automatic planning can be used to reach mission objectives, to perform navigation or maneuvers. Most of the time, the problem consists in finding a path from a source to a destination, while satisfying some operational constraints. In a graph without negative cycles, the computation of the single-pair shortest path from a start node to an end node is solved in polynomial time. Additional constraints on the solution path can however make the problem harder to solve. This becomes the case when we need the path to pass through a few mandatory nodes without requiring a specific order of visit. The complexity grows exponentially with the number of mandatory nodes to visit. In this paper, we focus on shortest path search with mandatory nodes on a given connected graph. We propose a hybrid model that combines a constraint-based solver and a graph convolutional neural network to improve search performance. Promising results are obtained on realistic scenarios.

Along with predictive performance and runtime speed, reliability is a key requirement for real-world semantic segmentation. Reliability encompasses robustness, predictive uncertainty and reduced bias. To improve reliability, we introduce Superpixel-mix, a new superpixel-based data augmentation method with teacher-student consistency training. Unlike other mixing-based augmentation techniques, mixing superpixels between images is aware of object boundaries, while yielding consistent gains in segmentation accuracy. Our proposed technique achieves state-of-the-art results in semi-supervised semantic segmentation on the Cityscapes dataset. Moreover, Superpixel-mix improves the reliability of semantic segmentation by reducing network uncertainty and bias, as confirmed by competitive results under strong distributions shift (adverse weather, image corruptions) and when facing out-of-distribution data.

The Travelling Salesman Problem (TSP) is one of the most investigated problems in the Combinatorial Optimization (CO) field. This is partly due to the fact that it belongs to the set of NP-Hard problems, which makes it particularly challenging. Moreover, the many practical problems that can be reduced to this - such as in Ratnesh et al. [10] where models of the TSP are presented to be used in the manufacture of microchips - make it even more attractive. At the same time, the full potentials of Machine Learning (ML) and Deep Learning (DL) techniques are becoming increasingly recognized in the CO field [2]. Mele et al. [17] recently introduced ML-Constructive, a promising constructive approach that computes fast solutions in two separate phases.

Federated learning (FL), proposed by Google at the very beginning [1], is recently a burgeoning research area of machine learning, which aims to protect individual data privacy in distributed machine learning process, especially in finance [2], smart healthcare [3, 4] and edge computing [5, 6]. Different from traditional data-centered distributed machine learning, participants in FL setting utilize localized data to train local model, then leverages specific strategies with other participants to acquire the final model collaboratively, avoiding direct data sharing behavior. Though it might differ in specific methodologies, current FL schemes can be summarized as repetition of training rounds, with each integrated by several basic steps: i) local update on client's model using their own localized data; ii) clients upload their local trained model parameters to server; iii) server performs aggregation strategy on collected clients' model parameters to obtain global model; iv) server selects a subset of clients and distributes the latest global model to them. Many FL researches try to improve algorithm effectiveness or efficiency on only one or more steps in this workflow with different scenarios: [7] suggests to add regularization term in step i) to achieve more robust convergence in heterogeneous settings; [8] applies gradient compression method in step ii) to reduce communication bandwidth; [9] tries to modify in step i), ii) and iii) for privacy-preserving purpose; [10] proposes better sample strategy in step iv) to address suboptimal result problem in Federated Multi-Task Learning. These indicate that the implementation of many FL algorithms only requires modification on several components of common workflow, without the necessity of repetitive implementation on basic FL workflow. The paradigm of FL and related research points are as depicted in figure 1.

This paper seeks to tackle the bin packing problem (BPP) through a learning perspective. Building on self-attention-based encoding and deep reinforcement learning algorithms, we propose a new end-to-end learning model for this task of interest. By decomposing the combinatorial action space, as well as utilizing a new training technique denoted as prioritized oversampling, which is a general scheme to speed up on-policy learning, we achieve state-of-the-art performance in a range of experimental settings. Moreover, although the proposed approach attend2pack targets offline-BPP, we strip our method down to the strict online-BPP setting where it is also able to achieve state-of-the-art performance. With a set of ablation studies as well as comparisons against a range of previous works, we hope to offer as a valid baseline approach to this field of study.

We propose three improvements to vision transformers (ViT) to reduce the number of trainable parameters without compromising classification accuracy. We address two shortcomings of the early ViT architectures -- quadratic bottleneck of the attention mechanism and the lack of an inductive bias in their architectures that rely on unrolling the two-dimensional image structure. Linear attention mechanisms overcome the bottleneck of quadratic complexity, which restricts application of transformer models in vision tasks. We modify the ViT architecture to work on longer sequence data by replacing the quadratic attention with efficient transformers, such as Performer, Linformer and Nystr\"omformer of linear complexity creating Vision X-formers (ViX). We show that all three versions of ViX may be more accurate than ViT for image classification while using far fewer parameters and computational resources. We also compare their performance with FNet and multi-layer perceptron (MLP) mixer. We further show that replacing the initial linear embedding layer by convolutional layers in ViX further increases their performance. Furthermore, our tests on recent vision transformer models, such as LeViT, Convolutional vision Transformer (CvT), Compact Convolutional Transformer (CCT) and Pooling-based Vision Transformer (PiT) show that replacing the attention with Nystr\"omformer or Performer saves GPU usage and memory without deteriorating the classification accuracy. We also show that replacing the standard learnable 1D position embeddings in ViT with Rotary Position Embedding (RoPE) give further improvements in accuracy. Incorporating these changes can democratize transformers by making them accessible to those with limited data and computing resources.