Radu Grosu

Mathias Lechner, Ramin Hasani, Zahra Babaiee et al.

Residual mappings have been shown to perform representation learning in the first layers and iterative feature refinement in higher layers. This interplay, combined with their stabilizing effect on the gradient norms, enables them to train very deep networks. In this paper, we take a step further and introduce entangled residual mappings to generalize the structure of the residual connections and evaluate their role in iterative learning representations. An entangled residual mapping replaces the identity skip connections with specialized entangled mappings such as orthogonal, sparse, and structural correlation matrices that share key attributes (eigenvalues, structure, and Jacobian norm) with identity mappings. We show that while entangled mappings can preserve the iterative refinement of features across various deep models, they influence the representation learning process in convolutional networks differently than attention-based models and recurrent neural networks. In general, we find that for CNNs and Vision Transformers entangled sparse mapping can help generalization while orthogonal mappings hurt performance. For recurrent networks, orthogonal residual mappings form an inductive bias for time-variant sequences, which degrades accuracy on time-invariant tasks.

Dung Phan, Junxing Yang, Matthew Clark et al.

We present Component-Based Simplex Architecture (CBSA), a new framework for assuring the runtime safety of component-based cyber-physical systems (CPSs). CBSA integrates Assume-Guarantee (A-G) reasoning with the core principles of the Simplex control architecture to allow component-based CPSs to run advanced, uncertified controllers while still providing runtime assurance that A-G contracts and global properties are satisfied. In CBSA, multiple Simplex instances, which can be composed in a nested, serial or parallel manner, coordinate to assure system-wide properties. Combining A-G reasoning and the Simplex architecture is a challenging problem that yields significant benefits. By utilizing A-G contracts, we are able to compositionally determine the switching logic for CBSAs, thereby alleviating the state explosion encountered by other approaches. Another benefit is that we can use A-G proof rules to decompose the proof of system-wide safety assurance into sub-proofs corresponding to the component-based structure of the system architecture. We also introduce the notion of coordinated switching between Simplex instances, a key component of our compositional approach to reasoning about CBSA switching logic. We illustrate our framework with a component-based control system for a ground rover. We formally prove that the CBSA for this system guarantees energy safety (the rover never runs out of power), and collision freedom (the rover never collides with a stationary obstacle). We also consider a CBSA for the rover that guarantees mission completion: all target destinations visited within a prescribed amount of time.

Usama Mehmood, Nicola Paoletti, Dung Phan et al.

The popularity of rule-based flocking models, such as Reynolds' classic flocking model, raises the question of whether more declarative flocking models are possible. This question is motivated by the observation that declarative models are generally simpler and easier to design, understand, and analyze than operational models. We introduce a very simple control law for flocking based on a cost function capturing cohesion (agents want to stay together) and separation (agents do not want to get too close). We refer to it as {\textit declarative flocking} (DF). We use model-predictive control (MPC) to define controllers for DF in centralized and distributed settings. A thorough performance comparison of our declarative flocking with Reynolds' model, and with more recent flocking models that use MPC with a cost function based on lattice structures, demonstrate that DF-MPC yields the best cohesion and least fragmentation, and maintains a surprisingly good level of geometric regularity while still producing natural flock shapes similar to those produced by Reynolds' model. We also show that DF-MPC has high resilience to sensor noise.

Scott A. Smolka, Ashish Tiwari, Lukas Esterle et al.

We introduce the concept of a V-formation game between a controller and an attacker, where controller's goal is to maneuver the plant (a simple model of flocking dynamics) into a V-formation, and the goal of the attacker is to prevent the controller from doing so. Controllers in V-formation games utilize a new formulation of model-predictive control we call Adaptive-Horizon MPC (AMPC), giving them extraordinary power: we prove that under certain controllability assumptions, an AMPC controller is able to attain V-formation with probability 1. We define several classes of attackers, including those that in one move can remove R birds from the flock, or introduce random displacement into flock dynamics. We consider both naive attackers, whose strategies are purely probabilistic, and AMPC-enabled attackers, putting them on par strategically with the controllers. While an AMPC-enabled controller is expected to win every game with probability 1, in practice, it is resource-constrained: its maximum prediction horizon and the maximum number of game execution steps are fixed. Under these conditions, an attacker has a much better chance of winning a V-formation game. Our extensive performance evaluation of V-formation games uses statistical model checking to estimate the probability an attacker can thwart the controller. Our results show that for the bird-removal game with R = 1, the controller almost always wins (restores the flock to a V-formation). For R = 2, the game outcome critically depends on which two birds are removed. For the displacement game, our results again demonstrate that an intelligent attacker, i.e. one that uses AMPC in this case, significantly outperforms its naive counterpart that randomly executes its attack.

Mónika Farsang, Mathias Lechner, David Lung et al.

Bio-inspired neural networks have the potential to advance our understanding of neural computation and improve the state-of-the-art of AI systems. Bio-electrical synapses directly transmit neural signals, by enabling fast current flow between neurons. In contrast, bio-chemical synapses transmit neural signals indirectly, through neurotransmitters. Prior work showed that interpretable dynamics for complex robotic control, can be achieved by using chemical synapses, within a sparse, bio-inspired architecture, called Neural Circuit Policies (NCPs). However, a comparison of these two synaptic models, within the same architecture, remains an unexplored area. In this work we aim to determine the impact of using chemical synapses compared to electrical synapses, in both sparse and all-to-all connected networks. We conduct experiments with autonomous lane-keeping through a photorealistic autonomous driving simulator to evaluate their performance under diverse conditions and in the presence of noise. The experiments highlight the substantial influence of the architectural and synaptic-model choices, respectively. Our results show that employing chemical synapses yields noticeable improvements compared to electrical synapses, and that NCPs lead to better results in both synaptic models.

Luigi Berducci, Shuo Yang, Rahul Mangharam et al.

Ensuring safety in dynamic multi-agent systems is challenging due to limited information about the other agents. Control Barrier Functions (CBFs) are showing promise for safety assurance but current methods make strong assumptions about other agents and often rely on manual tuning to balance safety, feasibility, and performance. In this work, we delve into the problem of adaptive safe learning for multi-agent systems with CBF. We show how emergent behavior can be profoundly influenced by the CBF configuration, highlighting the necessity for a responsive and dynamic approach to CBF design. We present ASRL, a novel adaptive safe RL framework, to fully automate the optimization of policy and CBF coefficients, to enhance safety and long-term performance through reinforcement learning. By directly interacting with the other agents, ASRL learns to cope with diverse agent behaviours and maintains the cost violations below a desired limit. We evaluate ASRL in a multi-robot system and a competitive multi-agent racing scenario, against learning-based and control-theoretic approaches. We empirically demonstrate the efficacy and flexibility of ASRL, and assess generalization and scalability to out-of-distribution scenarios. Code and supplementary material are public online.

Denise Ratasich, Michael Platzer, Radu Grosu et al.

The Internet of Things (IoT) connects millions of devices of different cyber-physical systems (CPSs) providing the CPSs additional (implicit) redundancy during runtime. However, the increasing level of dynamicity, heterogeneity, and complexity adds to the system's vulnerability, and challenges its ability to react to faults. Self-healing is an increasingly popular approach for ensuring resilience, that is, a proper monitoring and recovery, in CPSs. This work encodes and searches an adaptive knowledge base in Prolog/ProbLog that models relations among system variables given that certain implicit redundancy exists in the system. We exploit the redundancy represented in our knowledge base to generate adaptive runtime monitors which compares related signals by considering uncertainties in space and time. This enables the comparison of uncertain, asynchronous, multi-rate and delayed measurements. The monitor is used to trigger the recovery process of a self-healing mechanism. We demonstrate our approach by deploying it in a real-world CPS prototype of a rover whose sensors are susceptible to failure.

Julian Lemmel, Radu Grosu

We introduce a biologically plausible RL framework for solving tasks in partially observable Markov decision processes (POMDPs). The proposed algorithm combines three integral parts: (1) A Meta-RL architecture, resembling the mammalian basal ganglia; (2) A biologically plausible reinforcement learning algorithm, exploiting temporal difference learning and eligibility traces to train the policy and the value-function; (3) An online automatic differentiation algorithm for computing the gradients with respect to parameters of a shared recurrent network backbone. Our experimental results show that the method is capable of solving a diverse set of partially observable reinforcement learning tasks. The algorithm we call real-time recurrent reinforcement learning (RTRRL) serves as a model of learning in biological neural networks, mimicking reward pathways in the basal ganglia.

Shrajan Bhandary, Zahra Babaiee, Dejan Kostyszyn et al.

Despite the success of convolutional neural networks for 3D medical-image segmentation, the architectures currently used are still not robust enough to the protocols of different scanners, and the variety of image properties they produce. Moreover, access to large-scale datasets with annotated regions of interest is scarce, and obtaining good results is thus difficult. To overcome these challenges, we introduce IB-U-Nets, a novel architecture with inductive bias, inspired by the visual processing in vertebrates. With the 3D U-Net as the base, we add two 3D residual components to the second encoder blocks. They provide an inductive bias, helping U-Nets to segment anatomical structures from 3D images with increased robustness and accuracy. We compared IB-U-Nets with state-of-the-art 3D U-Nets on multiple modalities and organs, such as the prostate and spleen, using the same training and testing pipeline, including data processing, augmentation and cross-validation. Our results demonstrate the superior robustness and accuracy of IB-U-Nets, especially on small datasets, as is typically the case in medical-image analysis. IB-U-Nets source code and models are publicly available.

Julian Lemmel, Zahra Babaiee, Marvin Kleinlehner et al.

Modern tourism in the 21st century is facing numerous challenges. One of these challenges is the rapidly growing number of tourists in space limited regions such as historical city centers, museums or geographical bottlenecks like narrow valleys. In this context, a proper and accurate prediction of tourism volume and tourism flow within a certain area is important and critical for visitor management tasks such as visitor flow control and prevention of overcrowding. Static flow control methods like limiting access to hotspots or using conventional low level controllers could not solve the problem yet. In this paper, we empirically evaluate the performance of several state-of-the-art deep-learning methods in the field of visitor flow prediction with limited data by using available granular data supplied by a tourism region and comparing the results to ARIMA, a classical statistical method. Our results show that deep-learning models yield better predictions compared to the ARIMA method, while both featuring faster inference times and being able to incorporate additional input features.

Luigi Berducci, Radu Grosu

The automatic synthesis of a policy through reinforcement learning (RL) from a given set of formal requirements depends on the construction of a reward signal and consists of the iterative application of many policy-improvement steps. The synthesis algorithm has to balance target, safety, and comfort requirements in a single objective and to guarantee that the policy improvement does not increase the number of safety-requirements violations, especially for safety-critical applications. In this work, we present a solution to the synthesis problem by solving its two main challenges: reward-shaping from a set of formal requirements and safe policy update. For the former, we propose an automatic reward-shaping procedure, defining a scalar reward signal compliant with the task specification. For the latter, we introduce an algorithm ensuring that the policy is improved in a safe fashion with high-confidence guarantees. We also discuss the adoption of a model-based RL algorithm to efficiently use the collected data and train a model-free agent on the predicted trajectories, where the safety violation does not have the same impact as in the real world. Finally, we demonstrate in standard control benchmarks that the resulting learning procedure is effective and robust even under heavy perturbations of the hyperparameters.

Zahra Babaiee, Lucas Liebenwein, Ramin Hasani et al.

In this paper, we present a novel sensitivity-based filter pruning algorithm (SbF-Pruner) to learn the importance scores of filters of each layer end-to-end. Our method learns the scores from the filter weights, enabling it to account for the correlations between the filters of each layer. Moreover, by training the pruning scores of all layers simultaneously our method can account for layer interdependencies, which is essential to find a performant sparse sub-network. Our proposed method can train and generate a pruned network from scratch in a straightforward, one-stage training process without requiring a pretrained network. Ultimately, we do not need layer-specific hyperparameters and pre-defined layer budgets, since SbF-Pruner can implicitly determine the appropriate number of channels in each layer. Our experimental results on different network architectures suggest that SbF-Pruner outperforms advanced pruning methods. Notably, on CIFAR-10, without requiring a pretrained baseline network, we obtain 1.02% and 1.19% accuracy gain on ResNet56 and ResNet110, compared to the baseline reported for state-of-the-art pruning algorithms. This is while SbF-Pruner reduces parameter-count by 52.3% (for ResNet56) and 54% (for ResNet101), which is better than the state-of-the-art pruning algorithms with a high margin of 9.5% and 6.6%.

Shrajan Bhandary, Dejan Kuhn, Zahra Babaiee et al.

Accurate segmentation of prostate tumours from PET images presents a formidable challenge in medical image analysis. Despite considerable work and improvement in delineating organs from CT and MR modalities, the existing standards do not transfer well and produce quality results in PET related tasks. Particularly, contemporary methods fail to accurately consider the intensity-based scaling applied by the physicians during manual annotation of tumour contours. In this paper, we observe that the prostate-localised uptake threshold ranges are beneficial for suppressing outliers. Therefore, we utilize the intensity threshold values, to implement a new custom-feature-clipping normalisation technique. We evaluate multiple, established U-Net variants under different normalisation schemes, using the nnU-Net framework. All models were trained and tested on multiple datasets, obtained with two radioactive tracers: [68-Ga]Ga-PSMA-11 and [18-F]PSMA-1007. Our results show that the U-Net models achieve much better performance when the PET scans are preprocessed with our novel clipping technique.

Julian Lemmel, Radu Grosu

The approximation capability of ANNs and their RNN instantiations, is strongly correlated with the number of parameters packed into these networks. However, the complexity barrier for human understanding, is arguably related to the number of neurons and synapses in the networks, and to the associated nonlinear transformations. In this paper we show that the use of biophysical synapses, as found in LTCs, have two main benefits. First, they allow to pack more parameters for a given number of neurons and synapses. Second, they allow to formulate the nonlinear-network transformation, as a linear system with state-dependent coefficients. Both increase interpretability, as for a given task, they allow to learn a system linear in its input features, that is smaller in size compared to the state of the art. We substantiate the above claims on various time-series prediction tasks, but we believe that our results are applicable to any feedforward or recurrent ANN.

Zahra Babaiee, Lucas Liebenwein, Ramin Hasani et al.

Filter pruning of a CNN is typically achieved by applying discrete masks on the CNN's filter weights or activation maps, post-training. Here, we present a new filter-importance-scoring concept named pruning by active attention manipulation (PAAM), that sparsifies the CNN's set of filters through a particular attention mechanism, during-training. PAAM learns analog filter scores from the filter weights by optimizing a cost function regularized by an additive term in the scores. As the filters are not independent, we use attention to dynamically learn their correlations. Moreover, by training the pruning scores of all layers simultaneously, PAAM can account for layer inter-dependencies, which is essential to finding a performant sparse sub-network. PAAM can also train and generate a pruned network from scratch in a straightforward, one-stage training process without requiring a pre-trained network. Finally, PAAM does not need layer-specific hyperparameters and pre-defined layer budgets, since it can implicitly determine the appropriate number of filters in each layer. Our experimental results on different network architectures suggest that PAAM outperforms state-of-the-art structured-pruning methods (SOTA). On CIFAR-10 dataset, without requiring a pre-trained baseline network, we obtain 1.02% and 1.19% accuracy gain and 52.3% and 54% parameters reduction, on ResNet56 and ResNet110, respectively. Similarly, on the ImageNet dataset, PAAM achieves 1.06% accuracy gain while pruning 51.1% of the parameters on ResNet50. For Cifar-10, this is better than the SOTA with a margin of 9.5% and 6.6%, respectively, and on ImageNet with a margin of 11%.

Mihaela-Larisa Clement, Mónika Farsang, Agnes Poks et al.

The practical deployment of nonlinear model predictive control (NMPC) is often limited by online computation: solving a nonlinear program at high control rates can be expensive on embedded hardware, especially when models are complex or horizons are long. Learning-based NMPC approximations shift this computation offline but typically demand large expert datasets and costly training. We propose Sequential-AMPC, a sequential neural policy that generates MPC candidate control sequences by sharing parameters across the prediction horizon. For deployment, we wrap the policy in a safety-augmented online evaluation and fallback mechanism, yielding Safe Sequential-AMPC. Compared to a naive feedforward policy baseline across several benchmarks, Sequential-AMPC requires substantially fewer expert MPC rollouts and yields candidate sequences with higher feasibility rates and improved closed-loop safety. On high-dimensional systems, it also exhibits better learning dynamics and performance in fewer epochs while maintaining stable validation improvement where the feedforward baseline can stagnate.

Fatemeh Valeh, Monika Farsang, Radu Grosu et al.

State space models (SSMs) have recently achieved strong performance on long sequence modeling tasks while offering improved memory and computational efficiency compared to transformer based architectures. However, their evaluation has been largely limited to synthetic benchmarks and application domains such as language and audio, leaving their behavior on sparse and stochastic temporal processes in biological imaging unexplored. In this work, we introduce the Single Molecule Localization Microscopy Challenge (SMLM-C), a benchmark dataset consisting of ten SMLM simulations spanning dSTORM and DNA-PAINT modalities with varying hyperparameter designed to evaluate state space models on biologically realistic spatiotemporal point process data with known ground truth. Using a controlled subset of these simulations, we evaluate state space models and find that performance degrades substantially as temporal discontinuity increases, revealing fundamental challenges in modeling heavy-tailed blinking dynamics. These results highlight the need for sequence models better suited to sparse, irregular temporal processes encountered in real world scientific imaging data.

Mónika Farsang, Radu Grosu

In this paper, we present a unified framework for various bio-inspired models to better understand their structural and functional differences. We show that liquid-capacitance-extended models lead to interpretable behavior even in dense, all-to-all recurrent neural network (RNN) policies. We further demonstrate that incorporating chemical synapses improves interpretability and that combining chemical synapses with synaptic activation yields the most accurate and interpretable RNN models. To assess the accuracy and interpretability of these RNN policies, we consider the challenging lane-keeping control task and evaluate performance across multiple metrics, including turn-weighted validation loss, neural activity during driving, absolute correlation between neural activity and road trajectory, saliency maps of the networks' attention, and the robustness of their saliency maps measured by the structural similarity index.

Anna Lukina, Ashish Tiwari, Scott A. Smolka et al.

We present DAMPC, a distributed, adaptive-horizon and adaptive-neighborhood algorithm for solving the stochastic reachability problem in multi-agent systems, in particular flocking modeled as a Markov decision process. At each time step, every agent calls a centralized, adaptive-horizon model-predictive control (AMPC) algorithm to obtain an optimal solution for its local neighborhood. Second, the agents derive the flock-wide optimal solution through a sequence of consensus rounds. Third, the neighborhood is adaptively resized using a flock-wide, cost-based Lyapunov function V. This way DAMPC improves efficiency without compromising convergence. We evaluate DAMPC's performance using statistical model checking. Our results demonstrate that, compared to AMPC, DAMPC achieves considerable speed-up (two-fold in some cases) with only a slightly lower rate of convergence. The smaller average neighborhood size and lookahead horizon demonstrate the benefits of the DAMPC approach for stochastic reachability problems involving any controllable multi-agent system that possesses a cost function.

Dung Phan, Scott A. Smolka, Radu Grosu et al.

We present a novel approach to the problem of model checking cyber-physical systems. We transform the model checking problem to an optimization one by designing an objective function that measures how close a state is to a violation of a property. We use particle swarm optimization (PSO) to effectively search for a state that minimizes the objective function. Such states, if found, are counter-examples describing safe states from which the system can reach an unsafe state in one time step. We illustrate our approach with a controller for the Quickbot ground rover. Our PSO model checker quickly found a bug in the controller that could cause the rover to collide with an obstacle.

Julian Lemmel, Felix Resch, Mónika Farsang et al.

Deploying pretrained policies in real-world applications presents substantial challenges that fundamentally limit the practical applicability of learning-based control systems. When autonomous systems encounter environmental changes in system dynamics, sensor drift, or task objectives, fixed policies rapidly degrade in performance. We show that employing Real-Time Recurrent Reinforcement Learning (RTRRL), a biologically plausible algorithm for online adaptation, can effectively fine-tune a pretrained policy to improve autonomous agents' performance on driving tasks. We further show that RTRRL synergizes with a recent biologically inspired recurrent network model, the Liquid-Resistance Liquid-Capacitance RNN. We demonstrate the effectiveness of this closed-loop approach in a simulated CarRacing environment and in a real-world line-following task with a RoboRacer car equipped with an event camera.

Daniel Scheuchenstuhl, Stefan Ulmer, Felix Resch et al.

Robust and efficient learning remains a challenging problem in robotics, in particular with complex visual inputs. Inspired by human attention mechanism, with which we quickly process complex visual scenes and react to changes in the environment, we think that embedding auxiliary information about focus point into robot learning would enhance efficiency and robustness of the learning process. In this paper, we propose a novel approach to model and emulate the human attention with an approximate prediction model. We then leverage this output and feed it as a structured auxiliary feature map into downstream learning tasks. We validate this idea by learning a prediction model from human-gaze recordings of manual driving in the real world. We test our approach on two learning tasks - object detection and imitation learning. Our experiments demonstrate that the inclusion of predicted human attention leads to improved robustness of the trained models to out-of-distribution samples and faster learning in low-data regime settings. Our work highlights the potential of incorporating structured auxiliary information in representation learning for robotics and opens up new avenues for research in this direction. All code and data are available online.

Julian Lemmel, Zahra Babaiee, Marvin Kleinlehner et al.

Modern tourism in the 21st century is facing numerous challenges. Among these the rapidly growing number of tourists visiting space-limited regions like historical cities, museums and bottlenecks such as bridges is one of the biggest. In this context, a proper and accurate prediction of tourism volume and tourism flow within a certain area is important and critical for visitor management tasks such as sustainable treatment of the environment and prevention of overcrowding. Static flow control methods like conventional low-level controllers or limiting access to overcrowded venues could not solve the problem yet. In this paper, we empirically evaluate the performance of state-of-the-art deep-learning methods such as RNNs, GNNs, and Transformers as well as the classic statistical ARIMA method. Granular limited data supplied by a tourism region is extended by exogenous data such as geolocation trajectories of individual tourists, weather and holidays. In the field of visitor flow prediction with sparse data, we are thereby capable of increasing the accuracy of our predictions, incorporating modern input feature handling as well as mapping geolocation data on top of discrete POI data.

Sophie A. Neubauer, Radu Grosu

This paper presents, in a unified fashion, deterministic as well as statistical Lagrangian-verification techniques. They formally quantify the behavioral robustness of any time-continuous process, formulated as a continuous-depth model. To this end, we review LRT-NG, SLR, and GoTube, algorithms for constructing a tight reachtube, that is, an over-approximation of the set of states reachable within a given time-horizon, and provide guarantees for the reachtube bounds. We compare the usage of the variational equations, associated to the system equations, the mean value theorem, and the Lipschitz constants, in achieving deterministic and statistical guarantees. In LRT-NG, the Lipschitz constant is used as a bloating factor of the initial perturbation, to compute the radius of an ellipsoid in an optimal metric, which over-approximates the set of reachable states. In SLR and GoTube, we get statistical guarantees, by using the Lipschitz constants to compute local balls around samples. These are needed to calculate the probability of having found an upper bound, of the true maximum perturbation at every timestep. Our experiments demonstrate the superior performance of Lagrangian techniques, when compared to LRT, Flow*, and CAPD, and illustrate their use in the robustness analysis of various continuous-depth models.

Axel Brunnbauer, Luigi Berducci, Peter Priller et al.

The automated generation of diverse and complex training scenarios has been an important ingredient in many complex learning tasks. Especially in real-world application domains, such as autonomous driving, auto-curriculum generation is considered vital for obtaining robust and general policies. However, crafting traffic scenarios with multiple, heterogeneous agents is typically considered as a tedious and time-consuming task, especially in more complex simulation environments. In our work, we introduce MATS-Gym, a Multi-Agent Traffic Scenario framework to train agents in CARLA, a high-fidelity driving simulator. MATS-Gym is a multi-agent training framework for autonomous driving that uses partial scenario specifications to generate traffic scenarios with variable numbers of agents. This paper unifies various existing approaches to traffic scenario description into a single training framework and demonstrates how it can be integrated with techniques from unsupervised environment design to automate the generation of adaptive auto-curricula. The code is available at https://github.com/AutonomousDrivingExaminer/mats-gym.

Mónika Farsang, Ramin Hasani, Daniela Rus et al.

State Space Models (SSMs) are inherently recurrent along the sequence dimension, yet depth-recurrence - reusing the same block repeatedly across layers, as recently applied in looped transformers - has not been explored in this model family. We show that a looped SSM with $k$ parameters iterated $L$ times consistently closely matches or outperforms a standard SSM with $k \cdot L$ independent parameters across four architectures (LRU, S5, LinOSS, LrcSSM) and six time series classification benchmarks, despite operating within a strictly smaller hypothesis space, as we formally establish. Since the larger model contains the looped model as a special case, this dominance cannot be explained by expressivity and instead points to parameter sharing across depth as a beneficial inductive bias that simplifies optimization. These results demonstrate that depth-recurrence is orthogonal to sequence-recurrence and independently beneficial. We further show that input reshaping is an equally neglected design axis: concatenating timesteps for low-dimensional inputs, or flattening and rechunking the joint feature-time dimension for high-dimensional ones, yields accuracy gains of 1-6% across all models, confirmed over 5 random seeds. Both techniques provide standalone improvements that compound when combined, suggesting that depth and input reshaping are two independent and underexplored design axes for SSMs on time series.

Ramin Hasani, Mathias Lechner, Alexander Amini et al.

We introduce a new class of time-continuous recurrent neural network models. Instead of declaring a learning system's dynamics by implicit nonlinearities, we construct networks of linear first-order dynamical systems modulated via nonlinear interlinked gates. The resulting models represent dynamical systems with varying (i.e., liquid) time-constants coupled to their hidden state, with outputs being computed by numerical differential equation solvers. These neural networks exhibit stable and bounded behavior, yield superior expressivity within the family of neural ordinary differential equations, and give rise to improved performance on time-series prediction tasks. To demonstrate these properties, we first take a theoretical approach to find bounds over their dynamics and compute their expressive power by the trajectory length measure in latent trajectory space. We then conduct a series of time-series prediction experiments to manifest the approximation capability of Liquid Time-Constant Networks (LTCs) compared to classical and modern RNNs. Code and data are available at https://github.com/raminmh/liquid_time_constant_networks

Mónika Farsang, Ramin Hasani, Daniela Rus et al.

We present LrcSSM, a $\textit{non-linear}$ recurrent model that processes long sequences as fast as today's linear state-space layers. By forcing its Jacobian matrix to be diagonal, the full sequence can be solved in parallel, giving $\mathcal{O}(TD)$ time and memory and only $\mathcal{O}(\log T)$ sequential depth, for input-sequence length $T$ and a state dimension $D$. Moreover, LrcSSM offers a formal gradient-stability guarantee that other input-varying systems such as Liquid-S4 and Mamba do not provide. Importantly, the diagonal Jacobian structure of our model results in no performance loss compared to the original model with dense Jacobian, and the approach can be generalized to other non-linear recurrent models, demonstrating broader applicability. On a suite of long-range forecasting tasks, we demonstrate that LrcSSM outperforms Transformers, LRU, S5, and Mamba.

Zahra Babaiee, Peyman M. Kiasari, Daniela Rus et al.

This paper challenges the prevailing view that convolutional neural network (CNN) filters become increasingly specialized in deeper layers. Motivated by recent observations of clusterable repeating patterns in depthwise separable CNNs (DS-CNNs) trained on ImageNet, we extend this investigation across various domains and datasets. Our analysis of DS-CNNs reveals that deep filters maintain generality, contradicting the expected transition to class-specific filters. We demonstrate the generalizability of these filters through transfer learning experiments, showing that frozen filters from models trained on different datasets perform well and can be further improved when sourced from larger datasets. Our findings indicate that spatial features learned by depthwise separable convolutions remain generic across all layers, domains, and architectures. This research provides new insights into the nature of generalization in neural networks, particularly in DS-CNNs, and has significant implications for transfer learning and model design.

Zahra Babaiee, Peyman M. Kiassari, Daniela Rus et al.

A recent study has proposed the "Master Key Filters Hypothesis" for convolutional neural network filters. This paper extends this hypothesis by radically constraining its scope to a single set of just 8 universal filters that depthwise separable convolutional networks inherently converge to. While conventional DS-CNNs employ thousands of distinct trained filters, our analysis reveals these filters are predominantly linear shifts (ax+b) of our discovered universal set. Through systematic unsupervised search, we extracted these fundamental patterns across different architectures and datasets. Remarkably, networks initialized with these 8 unique frozen filters achieve over 80% ImageNet accuracy, and even outperform models with thousands of trainable parameters when applied to smaller datasets. The identified master key filters closely match Difference of Gaussians (DoGs), Gaussians, and their derivatives, structures that are not only fundamental to classical image processing but also strikingly similar to receptive fields in mammalian visual systems. Our findings provide compelling evidence that depthwise convolutional layers naturally gravitate toward this fundamental set of spatial operators regardless of task or architecture. This work offers new insights for understanding generalization and transfer learning through the universal language of these master key filters.

Dominik Freinberger, Julian Lemmel, Radu Grosu et al.

Recent advances in reinforcement learning have demonstrated the potential of quantum learning models based on parametrized quantum circuits as an alternative to deep learning models. On the one hand, these findings have shown the ultimate exponential speed-ups in learning that full-blown quantum models can offer in certain -- artificially constructed -- environments. On the other hand, they have demonstrated the ability of experimentally accessible PQCs to solve OpenAI Gym benchmarking tasks. However, it remains an open question whether these near-term QRL techniques can be successfully applied to more complex problems exhibiting high-dimensional observation spaces. In this work, we bridge this gap and present a hybrid model combining a PQC with classical feature encoding and post-processing layers that is capable of tackling Atari games. A classical model, subjected to architectural restrictions similar to those present in the hybrid model is constructed to serve as a reference. Our numerical investigation demonstrates that the proposed hybrid model is capable of solving the Pong environment and achieving scores comparable to the classical reference in Breakout. Furthermore, our findings shed light on important hyperparameter settings and design choices that impact the interplay of the quantum and classical components. This work contributes to the understanding of near-term quantum learning models and makes an important step towards their deployment in real-world RL scenarios.

Mugdim Bublin, Heimo Hirner, Antoine-Martin Lanners et al.

The exponential growth of IoT networks necessitates a paradigm shift towards architectures that offer high flexibility and learning capabilities while maintaining low energy consumption, minimal communication overhead, and low latency. Traditional IoT systems, particularly when integrated with machine learning approaches, often suffer from high communication overhead and significant energy consumption. This work addresses these challenges by proposing a neuromorphic architecture inspired by biological systems. To illustrate the practical application of our proposed architecture, we present a case study focusing on water management in the Carinthian community of Neuhaus. Preliminary results regarding water consumption prediction and anomaly detection in this community are presented. We also introduce a novel neuromorphic IoT architecture that integrates biological principles into the design of IoT systems. This architecture is specifically tailored for edge computing scenarios, where low power and high efficiency are crucial. Our approach leverages the inherent advantages of neuromorphic computing, such as asynchronous processing and event-driven communication, to create an IoT framework that is both energy-efficient and responsive. This case study demonstrates how the neuromorphic IoT architecture can be deployed in a real-world scenario, highlighting its benefits in terms of energy savings, reduced communication overhead, and improved system responsiveness.

Julian Lemmel, Manuel Kranzl, Adam Lamine et al.

Interactive Imitation Learning deals with training a novice policy from expert demonstrations in an online fashion. The established DAgger algorithm trains a robust novice policy by alternating between interacting with the environment and retraining of the network. Many variants thereof exist, that differ in the method of discerning whether to allow the novice to act or return control to the expert. We propose the use of stochastic reachtubes - common in verification of dynamical systems - as a novel method for estimating the necessity of expert intervention. Our approach does not require fine-tuning of decision thresholds per environment and effectively reduces the number of expert interventions, especially when compared with related approaches that make use of a doubt classification model.

Julian Lemmel, Manuel Kranzl, Adam Lamine et al.

This paper introduces a new approach for fine-tuning the predictions of structured state space models (SSMs) at inference time using real-time recurrent learning. While SSMs are known for their efficiency and long-range modeling capabilities, they are typically trained offline and remain static during deployment. Our method enables online adaptation by continuously updating model parameters in response to incoming data. We evaluate our approach for linear-recurrent-unit SSMs using a small carbon emission dataset collected from embedded automotive hardware. Experimental results show that our method consistently reduces prediction error online during inference, demonstrating its potential for dynamic, resource-constrained environments.

Jie He, Vincent Theo Willem Kenbeek, Zhantao Yang et al.

We introduce TD-Interpreter, a specialized ML tool that assists engineers in understanding complex timing diagrams (TDs), originating from a third party, during their design and verification process. TD-Interpreter is a visual question-answer environment which allows engineers to input a set of TDs and ask design and verification queries regarding these TDs. We implemented TD-Interpreter with multimodal learning by fine-tuning LLaVA, a lightweight 7B Multimodal Large Language Model (MLLM). To address limited training data availability, we developed a synthetic data generation workflow that aligns visual information with its textual interpretation. Our experimental evaluation demonstrates the usefulness of TD-Interpreter which outperformed untuned GPT-4o by a large margin on the evaluated benchmarks.

Claas Voelcker, Axel Brunnbauer, Marcel Hussing et al.

Score-function based methods for policy learning, such as REINFORCE and PPO, have delivered strong results in game-playing and robotics, yet their high variance often undermines training stability. Using pathwise policy gradients, i.e. computing a derivative by differentiating the objective function, alleviates the variance issues. However, they require an accurate action-conditioned value function, which is notoriously hard to learn without relying on replay buffers for reusing past off-policy data. We present an on-policy algorithm that trains Q-value models purely from on-policy trajectories, unlocking the possibility of using pathwise policy updates in the context of on-policy learning. We show how to combine stochastic policies for exploration with constrained updates for stable training, and evaluate important architectural components that stabilize value function learning. The result, Relative Entropy Pathwise Policy Optimization (REPPO), is an efficient on-policy algorithm that combines the stability of pathwise policy gradients with the simplicity and minimal memory footprint of standard on-policy learning. Compared to state-of-the-art on two standard GPU-parallelized benchmarks, REPPO provides strong empirical performance at superior sample efficiency, wall-clock time, memory footprint, and hyperparameter robustness.

Elpiniki Maria Lygizou, Mónika Farsang, Radu Grosu

Transformers excel across a large variety of tasks but remain susceptible to corrupted inputs, since standard self-attention treats all query-key interactions uniformly. Inspired by lateral inhibition in biological neural circuits and building on the recent use by the Differential Transformer's use of two parallel softmax subtraction for noise cancellation, we propose Multihead Differential Gated Self-Attention (M-DGSA) that learns per-head input-dependent gating to dynamically suppress attention noise. Each head splits into excitatory and inhibitory branches whose dual softmax maps are fused by a sigmoid gate predicted from the token embedding, yielding a context-aware contrast enhancement. M-DGSA integrates seamlessly into existing Transformer stacks with minimal computational overhead. We evaluate on both vision and language benchmarks, demonstrating consistent robustness gains over vanilla Transformer, Vision Transformer, and Differential Transformer baselines. Our contributions are (i) a novel input-dependent gating mechanism for self-attention grounded in lateral inhibition, (ii) a principled synthesis of biological contrast-enhancement and self-attention theory, and (iii) comprehensive experiments demonstrating noise resilience and cross-domain applicability.

Mihaela-Larisa Clement, Mónika Farsang, Felix Resch et al.

Autonomous agents that rely purely on perception to make real-time control decisions require efficient and robust architectures. In this work, we demonstrate that augmenting RGB input with depth information significantly enhances our agents' ability to predict steering commands compared to using RGB alone. We benchmark lightweight recurrent controllers that leverage the fused RGB-D features for sequential decision-making. To train our models, we collect high-quality data using a small-scale autonomous car controlled by an expert driver via a physical steering wheel, capturing varying levels of steering difficulty. Our models were successfully deployed on real hardware and inherently avoided dynamic and static obstacles, under out-of-distribution conditions. Specifically, our findings reveal that the early fusion of depth data results in a highly robust controller, which remains effective even with frame drops and increased noise levels, without compromising the network's focus on the task.

Axel Brunnbauer, Julian Lemmel, Zahra Babaiee et al.

Reinforcement learning algorithms for mean-field games offer a scalable framework for optimizing policies in large populations of interacting agents. Existing methods often depend on online interactions or access to system dynamics, limiting their practicality in real-world scenarios where such interactions are infeasible or difficult to model. In this paper, we present Offline Munchausen Mirror Descent (Off-MMD), a novel mean-field RL algorithm that approximates equilibrium policies in mean-field games using purely offline data. By leveraging iterative mirror descent and importance sampling techniques, Off-MMD estimates the mean-field distribution from static datasets without relying on simulation or environment dynamics. Additionally, we incorporate techniques from offline reinforcement learning to address common issues like Q-value overestimation, ensuring robust policy learning even with limited data coverage. Our algorithm scales to complex environments and demonstrates strong performance on benchmark tasks like crowd exploration or navigation, highlighting its applicability to real-world multi-agent systems where online experimentation is infeasible. We empirically demonstrate the robustness of Off-MMD to low-quality datasets and conduct experiments to investigate its sensitivity to hyperparameter choices.

Elpiniki Maria Lygizou, Michael Reiter, Margarita Maurer-Granofszky et al.

Acute Leukemia is the most common hematologic malignancy in children and adolescents. A key methodology in the diagnostic evaluation of this malignancy is immunophenotyping based on Multiparameter Flow Cytometry (FCM). However, this approach is manual, and thus time-consuming and subjective. To alleviate this situation, we propose in this paper the FCM-Former, a machine learning, self-attention based FCM-diagnostic tool, automating the immunophenotyping assessment in Childhood Acute Leukemia. The FCM-Former is trained in a supervised manner, by directly using flow cytometric data. Our FCM-Former achieves an accuracy of 96.5% assigning lineage to each sample among 960 cases of either acute B-cell, T-cell lymphoblastic, and acute myeloid leukemia (B-ALL, T-ALL, AML). To the best of our knowledge, the FCM-Former is the first work that automates the immunophenotyping assessment with FCM data in diagnosing pediatric Acute Leukemia.

Zahra Babaiee, Peyman M. Kiasari, Daniela Rus et al.

Recent advances in depthwise-separable convolutional neural networks (DS-CNNs) have led to novel architectures, that surpass the performance of classical CNNs, by a considerable scalability and accuracy margin. This paper reveals another striking property of DS-CNN architectures: discernible and explainable patterns emerge in their trained depthwise convolutional kernels in all layers. Through an extensive analysis of millions of trained filters, with different sizes and from various models, we employed unsupervised clustering with autoencoders, to categorize these filters. Astonishingly, the patterns converged into a few main clusters, each resembling the difference of Gaussian (DoG) functions, and their first and second-order derivatives. Notably, we were able to classify over 95\% and 90\% of the filters from state-of-the-art ConvNextV2 and ConvNeXt models, respectively. This finding is not merely a technological curiosity; it echoes the foundational models neuroscientists have long proposed for the vision systems of mammals. Our results thus deepen our understanding of the emergent properties of trained DS-CNNs and provide a bridge between artificial and biological visual processing systems. More broadly, they pave the way for more interpretable and biologically-inspired neural network designs in the future.

Zahra Babaiee, Peyman M. Kiasari, Daniela Rus et al.

In this study, we present evidence suggesting that depthwise convolutional kernels are effectively replicating the structural intricacies of the biological receptive fields observed in the mammalian retina. We provide analytics of trained kernels from various state-of-the-art models substantiating this evidence. Inspired by this intriguing discovery, we propose an initialization scheme that draws inspiration from the biological receptive fields. Experimental analysis of the ImageNet dataset with multiple CNN architectures featuring depthwise convolutions reveals a marked enhancement in the accuracy of the learned model when initialized with biologically derived weights. This underlies the potential for biologically inspired computational models to further our understanding of vision processing systems and to improve the efficacy of convolutional networks.

Shrajan Bhandary, Zahra Babaiee, Dejan Kostyszyn et al.

Despite the great success of convolutional neural networks (CNN) in 3D medical image segmentation tasks, the methods currently in use are still not robust enough to the different protocols utilized by different scanners, and to the variety of image properties or artefacts they produce. To this end, we introduce OOCS-enhanced networks, a novel architecture inspired by the innate nature of visual processing in the vertebrates. With different 3D U-Net variants as the base, we add two 3D residual components to the second encoder blocks: on and off center-surround (OOCS). They generalise the ganglion pathways in the retina to a 3D setting. The use of 2D-OOCS in any standard CNN network complements the feedforward framework with sharp edge-detection inductive biases. The use of 3D-OOCS also helps 3D U-Nets to scrutinise and delineate anatomical structures present in 3D images with increased accuracy.We compared the state-of-the-art 3D U-Nets with their 3D-OOCS extensions and showed the superior accuracy and robustness of the latter in automatic prostate segmentation from 3D Magnetic Resonance Images (MRIs). For a fair comparison, we trained and tested all the investigated 3D U-Nets with the same pipeline, including automatic hyperparameter optimisation and data augmentation.

Luigi Berducci, Edgar A. Aguilar, Dejan Ničković et al.

The automatic synthesis of policies for robotic-control tasks through reinforcement learning relies on a reward signal that simultaneously captures many possibly conflicting requirements. In this paper, we in\-tro\-duce a novel, hierarchical, potential-based reward-shaping approach (HPRS) for defining effective, multivariate rewards for a large family of such control tasks. We formalize a task as a partially-ordered set of safety, target, and comfort requirements, and define an automated methodology to enforce a natural order among requirements and shape the associated reward. Building upon potential-based reward shaping, we show that HPRS preserves policy optimality. Our experimental evaluation demonstrates HPRS's superior ability in capturing the intended behavior, resulting in task-satisfying policies with improved comfort, and converging to optimal behavior faster than other state-of-the-art approaches. We demonstrate the practical usability of HPRS on several robotics applications and the smooth sim2real transition on two autonomous-driving scenarios for F1TENTH race cars.

Jie He, Ezio Bartocci, Dejan Ničković et al.

Formal methods provide very powerful tools and techniques for the design and analysis of complex systems. Their practical application remains however limited, due to the widely accepted belief that formal methods require extensive expertise and a steep learning curve. Writing correct formal specifications in form of logical formulas is still considered to be a difficult and error prone task. In this paper we propose DeepSTL, a tool and technique for the translation of informal requirements, given as free English sentences, into Signal Temporal Logic (STL), a formal specification language for cyber-physical systems, used both by academia and advanced research labs in industry. A major challenge to devise such a translator is the lack of publicly available informal requirements and formal specifications. We propose a two-step workflow to address this challenge. We first design a grammar-based generation technique of synthetic data, where each output is a random STL formula and its associated set of possible English translations. In the second step, we use a state-of-the-art transformer-based neural translation technique, to train an accurate attentional translator of English to STL. The experimental results show high translation quality for patterns of English requirements that have been well trained, making this workflow promising to be extended for processing more complex translation tasks.

Sophie Gruenbacher, Mathias Lechner, Ramin Hasani et al.

We introduce a new stochastic verification algorithm that formally quantifies the behavioral robustness of any time-continuous process formulated as a continuous-depth model. Our algorithm solves a set of global optimization (Go) problems over a given time horizon to construct a tight enclosure (Tube) of the set of all process executions starting from a ball of initial states. We call our algorithm GoTube. Through its construction, GoTube ensures that the bounding tube is conservative up to a desired probability and up to a desired tightness. GoTube is implemented in JAX and optimized to scale to complex continuous-depth neural network models. Compared to advanced reachability analysis tools for time-continuous neural networks, GoTube does not accumulate overapproximation errors between time steps and avoids the infamous wrapping effect inherent in symbolic techniques. We show that GoTube substantially outperforms state-of-the-art verification tools in terms of the size of the initial ball, speed, time-horizon, task completion, and scalability on a large set of experiments. GoTube is stable and sets the state-of-the-art in terms of its ability to scale to time horizons well beyond what has been previously possible.

Zahra Babaiee, Ramin Hasani, Mathias Lechner et al.

Robustness to variations in lighting conditions is a key objective for any deep vision system. To this end, our paper extends the receptive field of convolutional neural networks with two residual components, ubiquitous in the visual processing system of vertebrates: On-center and off-center pathways, with excitatory center and inhibitory surround; OOCS for short. The on-center pathway is excited by the presence of a light stimulus in its center but not in its surround, whereas the off-center one is excited by the absence of a light stimulus in its center but not in its surround. We design OOCS pathways via a difference of Gaussians, with their variance computed analytically from the size of the receptive fields. OOCS pathways complement each other in their response to light stimuli, ensuring this way a strong edge-detection capability, and as a result, an accurate and robust inference under challenging lighting conditions. We provide extensive empirical evidence showing that networks supplied with the OOCS edge representation gain accuracy and illumination-robustness compared to standard deep models.

Mathias Lechner, Ramin Hasani, Radu Grosu et al.

Adversarial training is an effective method to train deep learning models that are resilient to norm-bounded perturbations, with the cost of nominal performance drop. While adversarial training appears to enhance the robustness and safety of a deep model deployed in open-world decision-critical applications, counterintuitively, it induces undesired behaviors in robot learning settings. In this paper, we show theoretically and experimentally that neural controllers obtained via adversarial training are subjected to three types of defects, namely transient, systematic, and conditional errors. We first generalize adversarial training to a safety-domain optimization scheme allowing for more generic specifications. We then prove that such a learning process tends to cause certain error profiles. We support our theoretical results by a thorough experimental safety analysis in a robot-learning task. Our results suggest that adversarial training is not yet ready for robot learning.

Axel Brunnbauer, Luigi Berducci, Andreas Brandstätter et al.

World models learn behaviors in a latent imagination space to enhance the sample-efficiency of deep reinforcement learning (RL) algorithms. While learning world models for high-dimensional observations (e.g., pixel inputs) has become practicable on standard RL benchmarks and some games, their effectiveness in real-world robotics applications has not been explored. In this paper, we investigate how such agents generalize to real-world autonomous vehicle control tasks, where advanced model-free deep RL algorithms fail. In particular, we set up a series of time-lap tasks for an F1TENTH racing robot, equipped with a high-dimensional LiDAR sensor, on a set of test tracks with a gradual increase in their complexity. In this continuous-control setting, we show that model-based agents capable of learning in imagination substantially outperform model-free agents with respect to performance, sample efficiency, successful task completion, and generalization. Moreover, we show that the generalization ability of model-based agents strongly depends on the choice of their observation model. We provide extensive empirical evidence for the effectiveness of world models provided with long enough memory horizons in sim2real tasks.

Sophie Gruenbacher, Ramin Hasani, Mathias Lechner et al.

We show that Neural ODEs, an emerging class of time-continuous neural networks, can be verified by solving a set of global-optimization problems. For this purpose, we introduce Stochastic Lagrangian Reachability (SLR), an abstraction-based technique for constructing a tight Reachtube (an over-approximation of the set of reachable states over a given time-horizon), and provide stochastic guarantees in the form of confidence intervals for the Reachtube bounds. SLR inherently avoids the infamous wrapping effect (accumulation of over-approximation errors) by performing local optimization steps to expand safe regions instead of repeatedly forward-propagating them as is done by deterministic reachability methods. To enable fast local optimizations, we introduce a novel forward-mode adjoint sensitivity method to compute gradients without the need for backpropagation. Finally, we establish asymptotic and non-asymptotic convergence rates for SLR.