Search

In offline reinforcement learning, a policy is learned using a static dataset in the absence of costly feedback from the environment. In contrast to the online setting, only using static datasets poses additional challenges, such as policies generating out-of-distribution samples. Model-based offline reinforcement learning methods try to overcome these by learning a model of the underlying dynamics of the environment and using it to guide policy search. It is beneficial but, with limited datasets, errors in the model and the issue of value overestimation among out-of-distribution states can worsen performance. Current model-based methods apply some notion of conservatism to the Bellman update, often implemented using uncertainty estimation derived from model ensembles. In this paper, we propose Constrained Latent Action Policies (C-LAP) which learns a generative model of the joint distribution of observations and actions. We cast policy learning as a constrained objective to always stay within the support of the latent action distribution, and use the generative capabilities of the model to impose an implicit constraint on the generated actions. Thereby eliminating the need to use additional uncertainty penalties on the Bellman update and significantly decreasing the number of gradient steps required to learn a policy. We empirically evaluate C-LAP on the D4RL and V-D4RL benchmark, and show that C-LAP is competitive to state-of-the-art methods, especially outperforming on datasets with visual observations., Comment: 38th Conference on Neural Information Processing Systems (NeurIPS 2024)

Regulatory frameworks for the use of AI are emerging. However, they trail behind the fast-evolving malicious AI technologies that can quickly cause lasting societal damage. In response, we introduce a pioneering Assistive AI framework designed to enhance human decision-making capabilities. This framework aims to establish a trust network across various fields, especially within legal contexts, serving as a proactive complement to ongoing regulatory efforts. Central to our framework are the principles of privacy, accountability, and credibility. In our methodology, the foundation of reliability of information and information sources is built upon the ability to uphold accountability, enhance security, and protect privacy. This approach supports, filters, and potentially guides communication, thereby empowering individuals and communities to make well-informed decisions based on cutting-edge advancements in AI. Our framework uses the concept of Boards as proxies to collectively ensure that AI-assisted decisions are reliable, accountable, and in alignment with societal values and legal standards. Through a detailed exploration of our framework, including its main components, operations, and sample use cases, the paper shows how AI can assist in the complex process of decision-making while maintaining human oversight. The proposed framework not only extends regulatory landscapes but also highlights the synergy between AI technology and human judgement, underscoring the potential of AI to serve as a vital instrument in discerning reality from fiction and thus enhancing the decision-making process. Furthermore, we provide domain-specific use cases to highlight the applicability of our framework., Comment: 37 pages, 6 figures; Changes: Revised references (k-anonimity)

The Finite Element Method (FEM) is a widely used technique for simulating crash scenarios with high accuracy and reliability. To reduce the significant computational costs associated with FEM, the Finite Element Method Integrated Networks (FEMIN) framework integrates neural networks (NNs) with FEM solvers. However, this integration can introduce errors and deviations from full-FEM simulations, highlighting the need for an additional metric to assess prediction confidence, especially when no ground truth data is available. In this study, we adapt the Deep Variational Bayes Filter (DVBF) to the FEMIN framework, incorporating a probabilistic approach to provide qualitative insights into prediction confidence during FEMIN simulations. The adaptation involves using the learned transition model for a predictive decoding step, generating a preliminary force prediction. This predictive force is used alongside the displacement and the velocity data from the FEM solver as input for the encoder model. The decoder reconstructs the likelihood distribution based on the posterior. The mean force of this distribution is applied to the FEM solver, while the predicted standard deviation can be used for uncertainty estimation. Our findings demonstrate that the DVBF outperforms deterministic NN architectures in terms of accuracy. Furthermore, the standard deviation derived from the decoder serves as a valuable qualitative metric for assessing the confidence in FEMIN simulations. This approach enhances the robustness of FEMIN by providing a measure of reliability alongside the simulation results.

Most recent successes in robot reinforcement learning involve learning a specialized single-task agent. However, robots capable of performing multiple tasks can be much more valuable in real-world applications. Multi-task reinforcement learning can be very challenging due to the increased sample complexity and the potentially conflicting task objectives. Previous work on this topic is dominated by model-free approaches. The latter can be very sample inefficient even when learning specialized single-task agents. In this work, we focus on model-based multi-task reinforcement learning. We propose a method for learning multi-task visual world models, leveraging pre-trained language models to extract semantically meaningful task representations. These representations are used by the world model and policy to reason about task similarity in dynamics and behavior. Our results highlight the benefits of using language-driven task representations for world models and a clear advantage of model-based multi-task learning over the more common model-free paradigm.

Robotic manipulation requires accurate motion and physical interaction control. However, current robot learning approaches focus on motion-centric action spaces that do not explicitly give the policy control over the interaction. In this paper, we discuss the repercussions of this choice and argue for more interaction-explicit action spaces in robot learning.

The growing adoption and deployment of Machine Learning (ML) systems came with its share of ethical incidents and societal concerns. It also unveiled the necessity to properly audit these systems in light of ethical principles. For such a novel type of algorithmic auditing to become standard practice, two main prerequisites need to be available: A lifecycle model that is tailored towards transparency and accountability, and a principled risk assessment procedure that allows the proper scoping of the audit. Aiming to make a pragmatic step towards a wider adoption of ML auditing, we present a respective procedure that extends the AI-HLEG guidelines published by the European Commission. Our audit procedure is based on an ML lifecycle model that explicitly focuses on documentation, accountability, and quality assurance; and serves as a common ground for alignment between the auditors and the audited organisation. We describe two pilots conducted on real-world use cases from two different organisations and discuss the shortcomings of ML algorithmic auditing as well as future directions thereof.

Incorporating the successful paradigm of pretraining and finetuning from Computer Vision and Natural Language Processing into decision-making has become increasingly popular in recent years. In this paper, we study Imitation Learning from Observation with pretrained models and find existing approaches such as BCO and AIME face knowledge barriers, specifically the Embodiment Knowledge Barrier (EKB) and the Demonstration Knowledge Barrier (DKB), greatly limiting their performance. The EKB arises when pretrained models lack knowledge about unseen observations, leading to errors in action inference. The DKB results from policies trained on limited demonstrations, hindering adaptability to diverse scenarios. We thoroughly analyse the underlying mechanism of these barriers and propose AIME-v2 upon AIME as a solution. AIME-v2 uses online interactions with data-driven regulariser to alleviate the EKB and mitigates the DKB by introducing a surrogate reward function to enhance policy training. Experimental results on tasks from the DeepMind Control Suite and Meta-World benchmarks demonstrate the effectiveness of these modifications in improving both sample-efficiency and converged performance. The study contributes valuable insights into resolving knowledge barriers for enhanced decision-making in pretraining-based approaches. Code will be available at https://github.com/argmax-ai/aime-v2., Comment: 19 pages, 7 figures

Multi-modality magnetic resonance imaging data with various sequences facilitate the early diagnosis, tumor segmentation, and disease staging in the management of nasopharyngeal carcinoma (NPC). The lack of publicly available, comprehensive datasets limits advancements in diagnosis, treatment planning, and the development of machine learning algorithms for NPC. Addressing this critical need, we introduce the first comprehensive NPC MRI dataset, encompassing MR axial imaging of 277 primary NPC patients. This dataset includes T1-weighted, T2-weighted, and contrast-enhanced T1-weighted sequences, totaling 831 scans. In addition to the corresponding clinical data, manually annotated and labeled segmentations by experienced radiologists offer high-quality data resources from untreated primary NPC.

We present a novel motion generation approach for robot arms, with high degrees of freedom, in complex settings that can adapt online to obstacles or new via points. Learning from Demonstration facilitates rapid adaptation to new tasks and optimizes the utilization of accumulated expertise by allowing robots to learn and generalize from demonstrated trajectories. We train a transformer architecture, based on conditional variational autoencoder, on a large dataset of simulated trajectories used as demonstrations. Our architecture learns essential motion generation skills from these demonstrations and is able to adapt them to meet auxiliary tasks. Additionally, our approach implements auto-regressive motion generation to enable real-time adaptations, as, for example, introducing or changing via-points, and velocity and acceleration constraints. Using beam search, we present a method for further adaption of our motion generator to avoid obstacles. We show that our model successfully generates motion from different initial and target points and that is capable of generating trajectories that navigate complex tasks across different robotic platforms., Comment: IEEE-RAS International Conference on Humanoid Robots, 2024

Objective: Subcutaneous Immunotherapy (SCIT) is the long-lasting causal treatment of allergic rhinitis (AR). How to enhance the adherence of patients to maximize the benefit of allergen immunotherapy (AIT) plays a crucial role in the management of AIT. This study aims to leverage novel machine learning models to precisely predict the risk of non-adherence of AR patients and related local symptom scores in three years SCIT. Methods: The research develops and analyzes two models, sequential latent-variable model (SLVM) of Stochastic Latent Actor-Critic (SLAC) and Long Short-Term Memory (LSTM) evaluating them based on scoring and adherence prediction capabilities. Results: Excluding the biased samples at the first time step, the predictive adherence accuracy of the SLAC models is from 60\% to 72\%, and for LSTM models, it is 66\% to 84\%, varying according to the time steps. The range of Root Mean Square Error (RMSE) for SLAC models is between 0.93 and 2.22, while for LSTM models it is between 1.09 and 1.77. Notably, these RMSEs are significantly lower than the random prediction error of 4.55. Conclusion: We creatively apply sequential models in the long-term management of SCIT with promising accuracy in the prediction of SCIT nonadherence in AR patients. While LSTM outperforms SLAC in adherence prediction, SLAC excels in score prediction for patients undergoing SCIT for AR. The state-action-based SLAC adds flexibility, presenting a novel and effective approach for managing long-term AIT., Comment: Frontiers in Pharmacology, research topic: Methods and Metrics to Measure Medication Adherence

We study the choice of action space in robot manipulation learning and sim-to-real transfer. We define metrics that assess the performance, and examine the emerging properties in the different action spaces. We train over 250 reinforcement learning~(RL) agents in simulated reaching and pushing tasks, using 13 different control spaces. The choice of spaces spans combinations of common action space design characteristics. We evaluate the training performance in simulation and the transfer to a real-world environment. We identify good and bad characteristics of robotic action spaces and make recommendations for future designs. Our findings have important implications for the design of RL algorithms for robot manipulation tasks, and highlight the need for careful consideration of action spaces when training and transferring RL agents for real-world robotics.

Unlike most reinforcement learning agents which require an unrealistic amount of environment interactions to learn a new behaviour, humans excel at learning quickly by merely observing and imitating others. This ability highly depends on the fact that humans have a model of their own embodiment that allows them to infer the most likely actions that led to the observed behaviour. In this paper, we propose Action Inference by Maximising Evidence (AIME) to replicate this behaviour using world models. AIME consists of two distinct phases. In the first phase, the agent learns a world model from its past experience to understand its own body by maximising the ELBO. While in the second phase, the agent is given some observation-only demonstrations of an expert performing a novel task and tries to imitate the expert's behaviour. AIME achieves this by defining a policy as an inference model and maximising the evidence of the demonstration under the policy and world model. Our method is "zero-shot" in the sense that it does not require further training for the world model or online interactions with the environment after given the demonstration. We empirically validate the zero-shot imitation performance of our method on the Walker and Cheetah embodiment of the DeepMind Control Suite and find it outperforms the state-of-the-art baselines. Code is available at: https://github.com/argmax-ai/aime., Comment: NeurIPS 2023

Type Ia supernovae remain poorly understood despite decades of investigation. Massive computationally intensive hydrodynamic simulations have been developed and run to model an ever-growing number of proposed progenitor channels. Further complicating the matter, a large number of sub-types of Type Ia supernovae have been identified in recent decades. Due to the massive computational load required, inference of the internal structure of Type Ia supernovae ejecta directly from observations using simulations has previously been computationally intractable. However, deep-learning emulators for radiation transport simulations have alleviated such barriers. We perform abundance tomography on 40 Type Ia supernovae from optical spectra using the radiative transfer code TARDIS accelerated by the probabilistic DALEK deep-learning emulator. We apply a parametric model of potential ejecta structures to comparatively investigate abundance distributions and internal ionization fractions of intermediate-mass elements between normal and 1991T-like Type Ia supernovae. Our inference shows that 1991T-like Type Ia supernovae are under-abundant in the typical intermediate mass elements that heavily contribute to the spectral line formation seen in normal Type Ia supernovae at early times. Additionally, we find that the intermediate-mass elements present in 1991T-like Type Ia supernovae are highly ionized compared to those in the normal Type Ia population. Finally, we conclude that the transition between normal and 1991T-like Type Ia supernovae appears to be continuous observationally and that the observed differences come out of a combination of both abundance and ionization fractions in these supernovae populations., Comment: 15 Pages, 6 Figures, Article, Submitted to ApJ

Partially-observable problems pose a trade-off between reducing costs and gathering information. They can be solved optimally by planning in belief space, but that is often prohibitively expensive. Model-predictive control (MPC) takes the alternative approach of using a state estimator to form a belief over the state, and then plan in state space. This ignores potential future observations during planning and, as a result, cannot actively increase or preserve the certainty of its own state estimate. We find a middle-ground between planning in belief space and completely ignoring its dynamics by only reasoning about its future accuracy. Our approach, filter-aware MPC, penalises the loss of information by what we call "trackability", the expected error of the state estimator. We show that model-based simulation allows condensing trackability into a neural network, which allows fast planning. In experiments involving visual navigation, realistic every-day environments and a two-link robot arm, we show that filter-aware MPC vastly improves regular MPC.

We introduce PRISM, a method for real-time filtering in a probabilistic generative model of agent motion and visual perception. Previous approaches either lack uncertainty estimates for the map and agent state, do not run in real-time, do not have a dense scene representation or do not model agent dynamics. Our solution reconciles all of these aspects. We start from a predefined state-space model which combines differentiable rendering and 6-DoF dynamics. Probabilistic inference in this model amounts to simultaneous localisation and mapping (SLAM) and is intractable. We use a series of approximations to Bayesian inference to arrive at probabilistic map and state estimates. We take advantage of well-established methods and closed-form updates, preserving accuracy and enabling real-time capability. The proposed solution runs at 10Hz real-time and is similarly accurate to state-of-the-art SLAM in small to medium-sized indoor environments, with high-speed UAV and handheld camera agents (Blackbird, EuRoC and TUM-RGBD)., Comment: Will appear in PMLR, CoRL 2022

Multi-robot manipulation tasks involve various control entities that can be separated into dynamically independent parts. A typical example of such real-world tasks is dual-arm manipulation. Learning to naively solve such tasks with reinforcement learning is often unfeasible due to the sample complexity and exploration requirements growing with the dimensionality of the action and state spaces. Instead, we would like to handle such environments as multi-agent systems and have several agents control parts of the whole. However, decentralizing the generation of actions requires coordination across agents through a channel limited to information central to the task. This paper proposes an approach to coordinating multi-robot manipulation through learned latent action spaces that are shared across different agents. We validate our method in simulated multi-robot manipulation tasks and demonstrate improvement over previous baselines in terms of sample efficiency and learning performance.

Supernova spectral time series can be used to reconstruct a spatially resolved explosion model known as supernova tomography. In addition to an observed spectral time series, a supernova tomography requires a radiative transfer model to perform the inverse problem with uncertainty quantification for a reconstruction. The smallest parametrizations of supernova tomography models are roughly a dozen parameters with a realistic one requiring more than 100. Realistic radiative transfer models require tens of CPU minutes for a single evaluation making the problem computationally intractable with traditional means requiring millions of MCMC samples for such a problem. A new method for accelerating simulations known as surrogate models or emulators using machine learning techniques offers a solution for such problems and a way to understand progenitors/explosions from spectral time series. There exist emulators for the TARDIS supernova radiative transfer code but they only perform well on simplistic low-dimensional models (roughly a dozen parameters) with a small number of applications for knowledge gain in the supernova field. In this work, we present a new emulator for the radiative transfer code TARDIS that not only outperforms existing emulators but also provides uncertainties in its prediction. It offers the foundation for a future active-learning-based machinery that will be able to emulate very high dimensional spaces of hundreds of parameters crucial for unraveling urgent questions in supernovae and related fields., Comment: 7 pages, accepted at ICML 2022 Workshop on Machine Learning for Astrophysics

Auto-encoder models that preserve similarities in the data are a popular tool in representation learning. In this paper we introduce several auto-encoder models that preserve local distances when mapping from the data space to the latent space. We use a local distance preserving loss that is based on the continuous k-nearest neighbours graph which is known to capture topological features at all scales simultaneously. To improve training performance, we formulate learning as a constraint optimisation problem with local distance preservation as the main objective and reconstruction accuracy as a constraint. We generalise this approach to hierarchical variational auto-encoders thus learning generative models with geometrically consistent latent and data spaces. Our method provides state-of-the-art performance across several standard datasets and evaluation metrics.

The use of machine learning in artistic music generation leads to controversial discussions of the quality of art, for which objective quantification is nonsensical. We therefore consider a music-generating algorithm as a counterpart to a human musician, in a setting where reciprocal interplay is to lead to new experiences, both for the musician and the audience. To obtain this behaviour, we resort to the framework of recurrent Variational Auto-Encoders (VAE) and learn to generate music, seeded by a human musician. In the learned model, we generate novel musical sequences by interpolation in latent space. Standard VAEs however do not guarantee any form of smoothness in their latent representation. This translates into abrupt changes in the generated music sequences. To overcome these limitations, we regularise the decoder and endow the latent space with a flat Riemannian manifold, i.e., a manifold that is isometric to the Euclidean space. As a result, linearly interpolating in the latent space yields realistic and smooth musical changes that fit the type of machine--musician interactions we aim for. We provide empirical evidence for our method via a set of experiments on music datasets and we deploy our model for an interactive jam session with a professional drummer. The live performance provides qualitative evidence that the latent representation can be intuitively interpreted and exploited by the drummer to drive the interplay. Beyond the musical application, our approach showcases an instance of human-centred design of machine-learning models, driven by interpretability and the interaction with the end user., Comment: 3rd Conference on AI Music Creativity (AIMC 2022)

Understanding the evolution of massive binary stars requires accurate estimates of their masses. This understanding is critically important because massive star evolution can potentially lead to gravitational wave sources such as binary black holes or neutron stars. For Wolf-Rayet stars with optically thick stellar winds, their masses can only be determined with accurate inclination angle estimates from binary systems which have spectroscopic $M \sin i$ measurements. Orbitally-phased polarization signals can encode the inclination angle of binary systems, where the Wolf-Rayet winds act as scattering regions. We investigated four Wolf-Rayet + O star binary systems, WR 42, WR 79, WR 127, and WR 153, with publicly available phased polarization data to estimate their masses. To avoid the biases present in analytic models of polarization while retaining computational expediency, we used a Monte Carlo radiative transfer model accurately emulated by a neural network. We used the emulated model to investigate the posterior distribution of parameters of our four systems. Our mass estimates calculated from the estimated inclination angles put strong constraints on existing mass estimates for three of the systems, and disagrees with the existing mass estimates for WR 153. We recommend a concerted effort to obtain polarization observations that can be used to estimate the masses of Wolf-Rayet binary systems and increase our understanding of their evolutionary paths., Comment: 24 pages, 13 figures, accepted to ApJ

We introduce a method for real-time navigation and tracking with differentiably rendered world models. Learning models for control has led to impressive results in robotics and computer games, but this success has yet to be extended to vision-based navigation. To address this, we transfer advances in the emergent field of differentiable rendering to model-based control. We do this by planning in a learned 3D spatial world model, combined with a pose estimation algorithm previously used in the context of TSDF fusion, but now tailored to our setting and improved to incorporate agent dynamics. We evaluate over six simulated environments based on complex human-designed floor plans and provide quantitative results. We achieve up to 92% navigation success rate at a frequency of 15 Hz using only image and depth observations under stochastic, continuous dynamics.

Manual fits to spectral times series of Type Ia supernovae have provided a method of reconstructing the explosion from a parametric model but due to lack of information about model uncertainties or parameter degeneracies direct comparison between theory and observation is difficult. In order to mitigate this important problem we present a new way to probabilistically reconstruct the outer ejecta of the normal Type Ia supernova SN 2002bo. A single epoch spectrum, taken 10 days before maximum light, is fit by a 13-parameter model describing the elemental composition of the ejecta and the explosion physics (density, temperature, velocity, and explosion epoch). Model evaluation is performed through the application of a novel rapid spectral synthesis technique in which the radiative transfer code, TARDIS, is accelerated by a machine-learning framework. Analysis of the posterior distribution reveals a complex and degenerate parameter space and allows direct comparison to various hydrodynamic models. Our analysis favors detonation over deflagration scenarios and we find that our technique offers a novel way to compare simulation to observation., Comment: 12 Pages, 4 Figures, submitted to AAS Journals, comments and constructive criticism welcome

Placing robots outside controlled conditions requires versatile movement representations that allow robots to learn new tasks and adapt them to environmental changes. The introduction of obstacles or the placement of additional robots in the workspace, the modification of the joint range due to faults or range-of-motion constraints are typical cases where the adaptation capabilities play a key role for safely performing the robot's task. Probabilistic movement primitives (ProMPs) have been proposed for representing adaptable movement skills, which are modelled as Gaussian distributions over trajectories. These are analytically tractable and can be learned from a small number of demonstrations. However, both the original ProMP formulation and the subsequent approaches only provide solutions to specific movement adaptation problems, e.g., obstacle avoidance, and a generic, unifying, probabilistic approach to adaptation is missing. In this paper we develop a generic probabilistic framework for adapting ProMPs. We unify previous adaptation techniques, for example, various types of obstacle avoidance, via-points, mutual avoidance, in one single framework and combine them to solve complex robotic problems. Additionally, we derive novel adaptation techniques such as temporally unbound via-points and mutual avoidance. We formulate adaptation as a constrained optimisation problem where we minimise the Kullback-Leibler divergence between the adapted distribution and the distribution of the original primitive while we constrain the probability mass associated with undesired trajectories to be low. We demonstrate our approach on several adaptation problems on simulated planar robot arms and 7-DOF Franka-Emika robots in a dual robot arm setting., Comment: There is a supplementary video accompanying the paper. It can be found at https://youtu.be/7UI6QX-eZ3I

Amortised inference enables scalable learning of sequential latent-variable models (LVMs) with the evidence lower bound (ELBO). In this setting, variational posteriors are often only partially conditioned. While the true posteriors depend, e.g., on the entire sequence of observations, approximate posteriors are only informed by past observations. This mimics the Bayesian filter -- a mixture of smoothing posteriors. Yet, we show that the ELBO objective forces partially-conditioned amortised posteriors to approximate products of smoothing posteriors instead. Consequently, the learned generative model is compromised. We demonstrate these theoretical findings in three scenarios: traffic flow, handwritten digits, and aerial vehicle dynamics. Using fully-conditioned approximate posteriors, performance improves in terms of generative modelling and multi-step prediction., Comment: Published as a conference paper at ICLR 2021 (Poster)

Supernova spectral time series contain a wealth of information about the progenitor and explosion process of these energetic events. The modeling of these data requires the exploration of very high dimensional posterior probabilities with expensive radiative transfer codes. Even modest parametrizations of supernovae contain more than ten parameters and a detailed exploration demands at least several million function evaluations. Physically realistic models require at least tens of CPU minutes per evaluation putting a detailed reconstruction of the explosion out of reach of traditional methodology. The advent of widely available libraries for the training of neural networks combined with their ability to approximate almost arbitrary functions with high precision allows for a new approach to this problem. Instead of evaluating the radiative transfer model itself, one can build a neural network proxy trained on the simulations but evaluating orders of magnitude faster. Such a framework is called an emulator or surrogate model. In this work, we present an emulator for the TARDIS supernova radiative transfer code applied to Type Ia supernova spectra. We show that we can train an emulator for this problem given a modest training set of a hundred thousand spectra (easily calculable on modern supercomputers). The results show an accuracy on the percent level (that are dominated by the Monte Carlo nature of TARDIS and not the emulator) with a speedup of several orders of magnitude. This method has a much broader set of applications and is not limited to the presented problem., Comment: 6 pages;5 figures submitted to AAS Journals. Constructive Criticism invited

With the increased focus on quantum circuit learning for near-term applications on quantum devices, in conjunction with unique challenges presented by cost function landscapes of parametrized quantum circuits, strategies for effective training are becoming increasingly important. In order to ameliorate some of these challenges, we investigate a layerwise learning strategy for parametrized quantum circuits. The circuit depth is incrementally grown during optimization, and only subsets of parameters are updated in each training step. We show that when considering sampling noise, this strategy can help avoid the problem of barren plateaus of the error surface due to the low depth of circuits, low number of parameters trained in one step, and larger magnitude of gradients compared to training the full circuit. These properties make our algorithm preferable for execution on noisy intermediate-scale quantum devices. We demonstrate our approach on an image-classification task on handwritten digits, and show that layerwise learning attains an 8% lower generalization error on average in comparison to standard learning schemes for training quantum circuits of the same size. Additionally, the percentage of runs that reach lower test errors is up to 40% larger compared to training the full circuit, which is susceptible to creeping onto a plateau during training., Comment: 11 pages, 7 figures

We solve the problem of 6-DoF localisation and 3D dense reconstruction in spatial environments as approximate Bayesian inference in a deep state-space model. Our approach leverages both learning and domain knowledge from multiple-view geometry and rigid-body dynamics. This results in an expressive predictive model of the world, often missing in current state-of-the-art visual SLAM solutions. The combination of variational inference, neural networks and a differentiable raycaster ensures that our model is amenable to end-to-end gradient-based optimisation. We evaluate our approach on realistic unmanned aerial vehicle flight data, nearing the performance of state-of-the-art visual-inertial odometry systems. We demonstrate the applicability of the model to generative prediction and planning., Comment: Update for ICLR2021

Learning to control robots without requiring engineered models has been a long-term goal, promising diverse and novel applications. Yet, reinforcement learning has only achieved limited impact on real-time robot control due to its high demand of real-world interactions. In this work, by leveraging a learnt probabilistic model of drone dynamics, we learn a thrust-attitude controller for a quadrotor through model-based reinforcement learning. No prior knowledge of the flight dynamics is assumed; instead, a sequential latent variable model, used generatively and as an online filter, is learnt from raw sensory input. The controller and value function are optimised entirely by propagating stochastic analytic gradients through generated latent trajectories. We show that "learning to fly" can be achieved with less than 30 minutes of experience with a single drone, and can be deployed solely using onboard computational resources and sensors, on a self-built drone.

Measuring the similarity between data points often requires domain knowledge, which can in parts be compensated by relying on unsupervised methods such as latent-variable models, where similarity/distance is estimated in a more compact latent space. Prevalent is the use of the Euclidean metric, which has the drawback of ignoring information about similarity of data stored in the decoder, as captured by the framework of Riemannian geometry. We propose an extension to the framework of variational auto-encoders allows learning flat latent manifolds, where the Euclidean metric is a proxy for the similarity between data points. This is achieved by defining the latent space as a Riemannian manifold and by regularising the metric tensor to be a scaled identity matrix. Additionally, we replace the compact prior typically used in variational auto-encoders with a recently presented, more expressive hierarchical one---and formulate the learning problem as a constrained optimisation problem. We evaluate our method on a range of data-sets, including a video-tracking benchmark, where the performance of our unsupervised approach nears that of state-of-the-art supervised approaches, while retaining the computational efficiency of straight-line-based approaches., Comment: Thirty-seventh International Conference on Machine Learning (ICML) 2020

Learning a model of dynamics from high-dimensional images can be a core ingredient for success in many applications across different domains, especially in sequential decision making. However, currently prevailing methods based on latent-variable models are limited to working with low resolution images only. In this work, we show that some of the issues with using high-dimensional observations arise from the discrepancy between the dimensionality of the latent and observable space, and propose solutions to overcome them.

We address tracking and prediction of multiple moving objects in visual data streams as inference and sampling in a disentangled latent state-space model. By encoding objects separately and including explicit position information in the latent state space, we perform tracking via amortized variational Bayesian inference of the respective latent positions. Inference is implemented in a modular neural framework tailored towards our disentangled latent space. Generative and inference model are jointly learned from observations only. Comparing to related prior work, we empirically show that our Markovian state-space assumption enables faithful and much improved long-term prediction well beyond the training horizon. Further, our inference model correctly decomposes frames into objects, even in the presence of occlusions. Tracking performance is increased significantly over prior art.

The study of dexterous manipulation has provided important insights in humans sensorimotor control as well as inspiration for manipulation strategies in robotic hands. Previous work focused on experimental environment with restrictions. Here we describe a method using the deformation and color distribution of the fingernail and its surrounding skin, to estimate the fingertip forces, torques and contact surface curvatures for various objects, including the shape and material of the contact surfaces and the weight of the objects. The proposed method circumvents limitations associated with sensorized objects, gloves or fixed contact surface type. In addition, compared with previous single finger estimation in an experimental environment, we extend the approach to multiple finger force estimation, which can be used for applications such as human grasping analysis. Four algorithms are used, c.q., Gaussian process (GP), Convolutional Neural Networks (CNN), Neural Networks with Fast Dropout (NN-FD) and Recurrent Neural Networks with Fast Dropout (RNN-FD), to model a mapping from images to the corresponding labels. The results further show that the proposed method has high accuracy to predict force, torque and contact surface., Comment: Robotica

We address the problem of one-to-many mappings in supervised learning, where a single instance has many different solutions of possibly equal cost. The framework of conditional variational autoencoders describes a class of methods to tackle such structured-prediction tasks by means of latent variables. We propose to incentivise informative latent representations for increasing the generalisation capacity of conditional variational autoencoders. To this end, we modify the latent variable model by defining the likelihood as a function of the latent variable only and introduce an expressive multimodal prior to enable the model for capturing semantically meaningful features of the data. To validate our approach, we train our model on the Cornell Robot Grasping dataset, and modified versions of MNIST and Fashion-MNIST obtaining results that show a significantly higher generalisation capability.

System identification of complex and nonlinear systems is a central problem for model predictive control and model-based reinforcement learning. Despite their complexity, such systems can often be approximated well by a set of linear dynamical systems if broken into appropriate subsequences. This mechanism not only helps us find good approximations of dynamics, but also gives us deeper insight into the underlying system. Leveraging Bayesian inference, Variational Autoencoders and Concrete relaxations, we show how to learn a richer and more meaningful state space, e.g. encoding joint constraints and collisions with walls in a maze, from partial and high-dimensional observations. This representation translates into a gain of accuracy of learned dynamics showcased on various simulated tasks., Comment: Appears in Proceedings of the 36th International Conference on Machine Learning (ICML)

We propose to learn a hierarchical prior in the context of variational autoencoders to avoid the over-regularisation resulting from a standard normal prior distribution. To incentivise an informative latent representation of the data, we formulate the learning problem as a constrained optimisation problem by extending the Taming VAEs framework to two-level hierarchical models. We introduce a graph-based interpolation method, which shows that the topology of the learned latent representation corresponds to the topology of the data manifold---and present several examples, where desired properties of latent representation such as smoothness and simple explanatory factors are learned by the prior., Comment: Published at NeurIPS 2019 (spotlight)

Recently, it has been shown that many functions on sets can be represented by sum decompositions. These decompositons easily lend themselves to neural approximations, extending the applicability of neural nets to set-valued inputs---Deep Set learning. This work investigates a core component of Deep Set architecture: aggregation functions. We suggest and examine alternatives to commonly used aggregation functions, including learnable recurrent aggregation functions. Empirically, we show that the Deep Set networks are highly sensitive to the choice of aggregation functions: beyond improved performance, we find that learnable aggregations lower hyper-parameter sensitivity and generalize better to out-of-distribution input size.

In this paper we propose a Bayesian method for estimating architectural parameters of neural networks, namely layer size and network depth. We do this by learning concrete distributions over these parameters. Our results show that regular networks with a learnt structure can generalise better on small datasets, while fully stochastic networks can be more robust to parameter initialisation. The proposed method relies on standard neural variational learning and, unlike randomised architecture search, does not require a retraining of the model, thus keeping the computational overhead at minimum., Comment: The 22nd International Conference on Artificial Intelligence and Statistics (AISTATS 2019)

The length of the geodesic between two data points along a Riemannian manifold, induced by a deep generative model, yields a principled measure of similarity. Current approaches are limited to low-dimensional latent spaces, due to the computational complexity of solving a non-convex optimisation problem. We propose finding shortest paths in a finite graph of samples from the aggregate approximate posterior, that can be solved exactly, at greatly reduced runtime, and without a notable loss in quality. Our approach, therefore, is hence applicable to high-dimensional problems, e.g., in the visual domain. We validate our approach empirically on a series of experiments using variational autoencoders applied to image data, including the Chair, FashionMNIST, and human movement data sets., Comment: 28th International Conference on Artificial Neural Networks, 2019

Learning from multiple sources of information is an important problem in machine-learning research. The key challenges are learning representations and formulating inference methods that take into account the complementarity and redundancy of various information sources. In this paper we formulate a variational autoencoder based multi-source learning framework in which each encoder is conditioned on a different information source. This allows us to relate the sources via the shared latent variables by computing divergence measures between individual source's posterior approximations. We explore a variety of options to learn these encoders and to integrate the beliefs they compute into a consistent posterior approximation. We visualise learned beliefs on a toy dataset and evaluate our methods for learning shared representations and structured output prediction, showing trade-offs of learning separate encoders for each information source. Furthermore, we demonstrate how conflict detection and redundancy can increase robustness of inference in a multi-source setting., Comment: AAAI 2019, Association for the Advancement of Artificial Intelligence (AAAI) 2019

Robots can rapidly acquire new skills from demonstrations. However, during generalisation of skills or transitioning across fundamentally different skills, it is unclear whether the robot has the necessary knowledge to perform the task. Failing to detect missing information often leads to abrupt movements or to collisions with the environment. Active learning can quantify the uncertainty of performing the task and, in general, locate regions of missing information. We introduce a novel algorithm for active learning and demonstrate its utility for generating smooth trajectories. Our approach is based on deep generative models and metric learning in latent spaces. It relies on the Jacobian of the likelihood to detect non-smooth transitions in the latent space, i.e., transitions that lead to abrupt changes in the movement of the robot. When non-smooth transitions are detected, our algorithm asks for an additional demonstration from that specific region. The newly acquired knowledge modifies the data manifold and allows for learning a latent representation for generating smooth movements. We demonstrate the efficacy of our approach on generalising elementary skills, transitioning across different skills, and implicitly avoiding collisions with the environment. For our experiments, we use a simulated pendulum where we observe its motion from images and a 7-DoF anthropomorphic arm., Comment: Published on 2018 IEEE/RSJ International Conference on Intelligent Robots and System

Model-based approaches bear great promise for decision making of agents interacting with the physical world. In the context of spatial environments, different types of problems such as localisation, mapping, navigation or autonomous exploration are typically adressed with specialised methods, often relying on detailed knowledge of the system at hand. We express these tasks as probabilistic inference and planning under the umbrella of deep sequential generative models. Using the frameworks of variational inference and neural networks, our method inherits favourable properties such as flexibility, scalability and the ability to learn from data. The method performs comparably to specialised state-of-the-art methodology in two distinct simulated environments., Comment: Preprint of publication at RSS 2019

We propose stochastic, non-parametric activation functions that are fully learnable and individual to each neuron. Complexity and the risk of overfitting are controlled by placing a Gaussian process prior over these functions. The result is the Gaussian process neuron, a probabilistic unit that can be used as the basic building block for probabilistic graphical models that resemble the structure of neural networks. The proposed model can intrinsically handle uncertainties in its inputs and self-estimate the confidence of its predictions. Using variational Bayesian inference and the central limit theorem, a fully deterministic loss function is derived, allowing it to be trained as efficiently as a conventional neural network using mini-batch gradient descent. The posterior distribution of activation functions is inferred from the training data alongside the weights of the network. The proposed model favorably compares to deep Gaussian processes, both in model complexity and efficiency of inference. It can be directly applied to recurrent or convolutional network structures, allowing its use in audio and image processing tasks. As an preliminary empirical evaluation we present experiments on regression and classification tasks, in which our model achieves performance comparable to or better than a Dropout regularized neural network with a fixed activation function. Experiments are ongoing and results will be added as they become available.

Kjolstad et. al. proposed a tensor algebra compiler. It takes expressions that define a tensor element-wise, such as $f_{ij}(a,b,c,d) = \exp\left[-\sum_{k=0}^4 \left((a_{ik}+b_{jk})^2\, c_{ii} + d_{i+k}^3 \right) \right]$, and generates the corresponding compute kernel code. For machine learning, especially deep learning, it is often necessary to compute the gradient of a loss function $l(a,b,c,d)=l(f(a,b,c,d))$ with respect to parameters $a,b,c,d$. If tensor compilers are to be applied in this field, it is necessary to derive expressions for the derivatives of element-wise defined tensors, i.e. expressions for $(da)_{ik}=\partial l/\partial a_{ik}$. When the mapping between function indices and argument indices is not 1:1, special attention is required. For the function $f_{ij} (x) = x_i^2$, the derivative of the loss is $(dx)_i=\partial l/\partial x_i=\sum_j (df)_{ij}2x_i$; the sum is necessary because index $j$ does not appear in the indices of $f$. Another example is $f_{i}(x)=x_{ii}^2$, where $x$ is a matrix; here we have $(dx)_{ij}=\delta_{ij}(df)_i2x_{ii}$; the Kronecker delta is necessary because the derivative is zero for off-diagonal elements. Another indexing scheme is used by $f_{ij}(x)=\exp x_{i+j}$; here the correct derivative is $(dx)_{k}=\sum_i (df)_{i,k-i} \exp x_{k}$, where the range of the sum must be chosen appropriately. In this publication we present an algorithm that can handle any case in which the indices of an argument are an arbitrary linear combination of the indices of the function, thus all the above examples can be handled. Sums (and their ranges) and Kronecker deltas are automatically inserted into the derivatives as necessary. Additionally, the indices are transformed, if required (as in the last example). The algorithm outputs a symbolic expression that can be subsequently fed into a tensor algebra compiler. Source code is provided., Comment: Technical Report

Neural samplers such as variational autoencoders (VAEs) or generative adversarial networks (GANs) approximate distributions by transforming samples from a simple random source---the latent space---to samples from a more complex distribution represented by a dataset. While the manifold hypothesis implies that the density induced by a dataset contains large regions of low density, the training criterions of VAEs and GANs will make the latent space densely covered. Consequently points that are separated by low-density regions in observation space will be pushed together in latent space, making stationary distances poor proxies for similarity. We transfer ideas from Riemannian geometry to this setting, letting the distance between two points be the shortest path on a Riemannian manifold induced by the transformation. The method yields a principled distance measure, provides a tool for visual inspection of deep generative models, and an alternative to linear interpolation in latent space. In addition, it can be applied for robot movement generalization using previously learned skills. The method is evaluated on a synthetic dataset with known ground truth; on a simulated robot arm dataset; on human motion capture data; and on a generative model of handwritten digits., Comment: Published on the 21st International Conference on Artificial Intelligence and Statistics (AISTATS), 2018

We introduce a methodology for efficiently computing a lower bound to empowerment, allowing it to be used as an unsupervised cost function for policy learning in real-time control. Empowerment, being the channel capacity between actions and states, maximises the influence of an agent on its near future. It has been shown to be a good model of biological behaviour in the absence of an extrinsic goal. But empowerment is also prohibitively hard to compute, especially in nonlinear continuous spaces. We introduce an efficient, amortised method for learning empowerment-maximising policies. We demonstrate that our algorithm can reliably handle continuous dynamical systems using system dynamics learned from raw data. The resulting policies consistently drive the agents into states where they can use their full potential.

The recognition of actions from video sequences has many applications in health monitoring, assisted living, surveillance, and smart homes. Despite advances in sensing, in particular related to 3D video, the methodologies to process the data are still subject to research. We demonstrate superior results by a system which combines recurrent neural networks with convolutional neural networks in a voting approach. The gated-recurrent-unit-based neural networks are particularly well-suited to distinguish actions based on long-term information from optical tracking data; the 3D-CNNs focus more on detailed, recent information from video data. The resulting features are merged in an SVM which then classifies the movement. In this architecture, our method improves recognition rates of state-of-the-art methods by 14% on standard data sets., Comment: Published in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)