Deep learning models deployed on edge devices frequently encounter resource variability, which arises from fluctuating energy levels, timing constraints, or prioritization of other critical tasks within the system. State-of-the-art machine learning pipelines generate resource-agnostic models that are not capable to adapt at runtime. In this work, we introduce Resource-Efficient Deep Subnetworks (REDS) to tackle model adaptation to variable resources. In contrast to the state-of-the-art, REDS leverages structured sparsity constructively by exploiting permutation invariance of neurons, which allows for hardware-specific optimizations. Specifically, REDS achieves computational efficiency by (1) skipping sequential computational blocks identified by a novel iterative knapsack optimizer, and (2) taking advantage of data cache by re-arranging the order of operations in REDS computational graph. REDS supports conventional deep networks frequently deployed on the edge and provides computational benefits even for small and simple networks. We evaluate REDS on eight benchmark architectures trained on the Visual Wake Words, Google Speech Commands, Fashion-MNIST, CIFAR-10 and ImageNet-1K datasets, and test on four off-the-shelf mobile and embedded hardware platforms. We provide a theoretical result and empirical evidence demonstrating REDS' outstanding performance in terms of submodels' test set accuracy, and demonstrate an adaptation time in response to dynamic resource constraints of under 40$\mu$s, utilizing a fully-connected network on Arduino Nano 33 BLE.

We introduce model folding, a novel data-free model compression technique that merges structurally similar neurons across layers, significantly reducing the model size without the need for fine-tuning or access to training data. Unlike existing methods, model folding preserves data statistics during compression by leveraging k-means clustering, and using novel data-free techniques to prevent variance collapse or explosion. Our theoretical framework and experiments across standard benchmarks, including ResNet18 and LLaMA-7B, demonstrate that model folding achieves comparable performance to data-driven compression techniques and outperforms recently proposed data-free methods, especially at high sparsity levels. This approach is particularly effective for compressing large-scale models, making it suitable for deployment in resource-constrained environments.

Careful parametrization of networking protocols is crucial to maximize the performance of low-power wireless systems and ensure that stringent application requirements can be met. This is a non-trivial task involving thorough characterization on testbeds and requiring expert knowledge. Unfortunately, the community still lacks a tool to facilitate parameter exploration while minimizing the necessary experimentation time on testbeds. Such a tool would be invaluable, as exhaustive parameter searches can be time-prohibitive or unfeasible given limited testbed availability, whereas non-exhaustive unguided searches rarely deliver satisfactory results. In this article, we present APEX, a framework enabling an automated and informed parameter exploration for low-power wireless protocols and allowing convergence to the best parameter set within a limited number of testbed trials. We design APEX using Gaussian processes to effectively handle noisy experimental data and estimate the optimality of a parameter combination. After developing a prototype of APEX, we demonstrate its effectiveness by parametrizing two IEEE 802.15.4 protocols across a wide range of application requirements. Our results show that APEX can return the best parameter set with up to 10.6×, 4.5×, 4.3×, and 3.25× less testbed trials than traditional solutions based on exhaustive search, greedy approaches, support vector regression and reinforcement learning, respectively.

Understanding attention is crucial for improving safety in driving scenarios. Detected and classified objects, along with their observation by the driver, are used as a measure of attention. This paper investigates the differences between human and artificial attention in real-world and replay driving scenarios. By analyzing attention patterns from drivers and a vision-language model agent, we identify a number of differences. The results highlight the limitations of current AI attention models and suggest the way forward for developing more context-aware systems.

Distribution shifts are characterized by differences between the training and test data distributions. They can significantly reduce the accuracy of machine learning models deployed in real-world scenarios. This paper explores the distribution shift problem when classifying pollen grains from microscopic images collected in the wild with a low-cost camera sensor. We leverage the domain knowledge that geometric features are highly important for accurate pollen identification and introduce two novel geometric image augmentation techniques to significantly narrow the accuracy gap between the model performance on the train and test datasets. In particular, we show that Tenengrad and ImageToSketch filters are highly effective to balance the shape and texture information while leaving out unimportant details that may confuse the model. Extensive evaluations on various model architectures demonstrate a consistent improvement of the model generalization to field data of up to 14% achieved by the geometric augmentation techniques when compared to a wide range of standard image augmentations. The approach is validated through an ablation study using pollen hydration tests to recover the shape of dry pollen grains. The proposed geometric augmentations also receive the highest scores according to the affinity and diversity measures from the literature.

Multimodal datasets are a critical component in recent breakthroughs such as CLIP, Stable Diffusion and GPT-4, yet their design does not receive the same research attention as model architectures or training algorithms. To address this shortcoming in the machine learning ecosystem, we introduce DataComp, a testbed for dataset experiments centered around a new candidate pool of 12.8 billion image-text pairs from Common Crawl. Participants in our benchmark design new filtering techniques or curate new data sources and then evaluate their new dataset by running our standardized CLIP training code and testing the resulting model on 38 downstream test sets. Our benchmark consists of multiple compute scales spanning four orders of magnitude, which enables the study of scaling trends and makes the benchmark accessible to researchers with varying resources. Our baseline experiments show that the DataComp workflow leads to better training sets. Our best baseline, DataComp-1B, enables training a CLIP ViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, outperforming OpenAI's CLIP ViT-L/14 by 3.7 percentage points while using the same training procedure and compute. We release \datanet and all accompanying code at www.datacomp.ai.

Accurate calibration of low-cost environmental sensors is a prerequisite for their successful use in many monitoring applications. State-of-the-art calibration methods vary from simple linear regression to sophisticated deep models based on LSTMs and GRUs. The latter take past measurements to improve calibration accuracy. In this article, we argue that both recent past and close future measurements help to achieve accurate calibration, whereas accuracy improvements beyond the past come with a delay introduced by the occurrence of the future. We propose a generalized many-to-many calibration scheme called SensorFormer based on the successful Transformer model which takes both past and future raw measurements into account. We show that the proposed approach: 1) outperforms other methods by improving calibration accuracy by 16.5%–20.4% on public data sets and own field data and 2) can efficiently run on low-power microcontrollers with very limited computational and storage capabilities. The latter is achieved by a novel optimization technique based on learnable input subsampling taking advantage of the properties of typical sensor data. We manage to reduce the model size by 20%–33% and minimize the overall floating point operations per second (FLOPs) by 65% while maintaining superior accuracy than state-of-the-art methods.

The COVID-19 related lockdown measures offer a unique opportunity to understand how changes in economic activity and traffic affect ambient air quality, and how much pollution reduction can the society offer through digitalization and mobility-limiting policies. In this work, we estimate pollution reduction over the lockdown period by using the measurements from ground air pollution monitoring networks, training long-term prediction models and comparing their predictions to measured values over the lockdown month. We show that our models achieve state-of-the-art performance and evaluate up to -29.4%, -28.1%, and -52.8%, change in NO2 in Eastern Switzerland, Beijing and Wuhan respectively. Our reduction estimates take local weather into account. What can we learn from pollution emissions during lockdown? The lockdown period was too short to train meaningful models from scratch. We therefore use transfer learning to update only mobility-dependent variables. We show that the obtained models are suitable for the analysis of the post-lockdown periods and capable of estimating the future air pollution reduction potential.

Low-power operation of power-hungry MOX sensors is usually achieved by duty cycling them periodically. However, a difficulty arises if a sensor is operated at irregular time intervals due to intermittent energy availability in energy- neutral or batteryless applications. In this work, we propose a compensation method which re-maps on-demand measurements to virtually duty-cycled readings in the value domain by taking the duration of the last off-time into account. We evaluate our compensation algorithm based on Sensirion’s SGP30 sensor and achieve up to 79% accuracy improvement compared to uncompensated measurements.

Popular low-cost air quality sensors embedded into IoT and mobile devices are based on metal oxides (MOX) that change their electrical resistance in response to ambient pollutants emitted as gases. Operating MOX sensors continuously is expensive, since it requires to heat up and maintain a hotplate at several hundred degrees. To save energy, sensors are commonly duty cycled with short on-times and long off-times. However, doing so adversely affects the sensor's chemical reactions, which have slower transients as the off-time increases. As a result, sensor sensitivity to various gases deviates from a continuously powered sensor. In this paper, we show that it is possible to recover accurate continuous-sensor measurements from transient responses obtained from a duty cycled sensor and compensate for an altered multi-gas cross-sensitivity profile using machine learning methods. On a test set, we achieve a mean absolute error (MAE) of 24ppb between continuous ground-truth measurements and obtained model predictions of tVOC. This results in estimating 86.6% of Indoor Air Quality (IAQ) levels correctly compared to 68.1% if no correction is used. Our models are invariant to minor baseline shifts and work for both tVOC and CO2-eq signals provided by the sensor. Thanks to our models, 98.5% of the energy consumption can be reduced while maintaining high accuracy. This optimization enables energy-harvesting-based operation of IAQ sensors in indoor IoT scenarios.

Low-cost sensors are extensively used in numerous Internet of Things (IoT) applications to measure relevant physical processes. Today, processing context data is increasingly done by proprietary algorithms tuned to a specific use-case, e.g., a sensor measuring activity intensity of a cow. Readings from these sensors may be subject to data distribution shifts, which challenge robustness of models using these sensor readings. In this paper, we propose a new sensor data processing framework, which leverages a co-dependency between data quality and model robustness to detect performance issues of data-driven predictive models in the field. We show how distribution shifts in the input data impact the quality of the model, which relies on application-specific sensors, and present indicators capable of detecting such shifts in the wild. The proposed framework used in the context of precision cattle farming allows improving the quality of cow lameness predictive models on the field data by up to 62%.

In this work, we show that sensor data may leak a sensor's location even if the latter is not explicitly included in the data set. The sensors are localized by linking sensor data, in particular activity data, with publicly available environmental data such as weather data. We show that using a linkage attack a cow can be localized within an entire country with an average accuracy of up to 32.6 km solely from activity traces recorded with a tracker in the cow's stomach.

Network control in microgrids is an active research area driven by a steady increase in energy demand, the necessity to minimize the environmental footprint, yet achieve socioeconomic benefits and ensure sustainability. Reducing deviation of the predicted energy consumption from the actual one, softening peaks in demand and filling in the troughs, especially at times when power is more affordable and clean, present challenges for the demand-side response. In this paper, we present a hierarchical energy system architecture with embedded control. This architecture pushes prediction models to edge devices and executes local control loops to address the challenge of managing demand-side response locally. We employ a two-step approach: At an upper level of hierarchy, we adopt a conventional machine learning pipeline to build load prediction models using automated domain-specific feature extraction and selection. Given historical data, these models are then used to label prediction failure events that force the operator to use backup energy sources to stabilize the network. On a lower level of hierarchy, computed labels are used to train impact models realized by LSTM networks running on edge devices to infer the probability that the power consumption of the player contributes to the upper level prediction failure event. The system is evaluated on clustered and aggregated energy traces from a public data set of academic buildings. The results show the benefits of the proposed hierarchical energy system architecture in terms of impact prediction with 55% accuracy. This allows minimizing the number of prediction failure events by 11.69 % by executing targeted local control.

In this paper, we present the design of our prototype of an automated real-time and affordable pollen sensing system. The design consists of three main subsystems: (1) a trap with automatic filtering, (2) a particle concentration system, and (3) a digital microscope with autofocus. The prototype shows effective particle gathering, filtering and concentration in a tiny sized area. As a result, we reduce particle loss and improve image quality taken by the optical system when searching and autofocusing on pollen grains. Our first prototype collects raw time-stamped data and transmits these to the backend server where we plan to run the detection and classification algorithms to extract accurate pollen counts from microscopic images. The key advantage of processing images at the backend is that we let the experts undertake corrective actions and help the system learn to detect and classify pollen using state-of-the-art interactive imitation learning algorithms. The final model can then run locally on embedded hardware.

Great public speakers are made, not born. Practicing a presentation in front of colleagues is common practice and results in a set of subjective judgements what could be improved. In this paper we describe the design and implementation of a mobile app which estimates the quality of speaker's delivery in real time in a fair, repeatable and privacy-preserving way. Quantle estimates the speaker's pace in terms of the number of syllables, words and clauses, computes pitch and duration of pauses. The basic parameters are then used to estimate the talk complexity based on readability scores from the literature to help the speaker adjust his delivery to the target audience. In contrast to speech-to-text-based methods used to implement a digital presentation coach, Quantle does processing locally in real time and works in the flight mode. This design has three implications: (1) Quantle does not interfere with the surrounding hardware, (2) it is power-aware, since 95.2% of the energy used by the app on iPhone 6 is spent to operate the built-in microphone and the screen, and (3) audio data and processing results are not shared with a third party therewith preserving speaker's privacy.

Sensing tongue movements enables various applications in hands-free interaction and alternative communication. We propose BARTON, a BARometer based low-power and robust TONgue movement sensing system. Using a low sampling rate of below 50 Hz, and only extracting simple temporal features from in-ear pressure signals, we demonstrate that it is plausible to distinguish important tongue gestures (left, right, forward) at low power consumption. We prototype BARTON with commodity earpieces integrated with COTS barometers for in-ear pressure sensing and an ARM micro-controller for signal processing. Evaluations show that BARTON yields 94% classification accuracy and 8.4 mW power consumption, which achieves comparable accuracy, but consumes 44 times lower energy than the state-of-the-art microphone-based solutions. BARTON is also robust to head movements and operates with music played directly from earphones.

The availability of novel, high-resolution pollution maps enables a wide range of new application scenarios, which were not possible before. In this paper, we combine high-resolution pollution maps available for the city of Zurich, Switzerland, with road network data to analyze how much urban dwellers can reduce their exposure to air pollution by not taking the shortest path between origin and destination but a healthier and slightly longer alternative route. We introduce a new weight function to assess the exposure on each street segment and evaluate the benefits of the healthier path. Finally, we efficiently implement the algorithm as stand-alone application for iOS and Android devices. The app helps city residents to understand and reduce their exposure to air pollutants.

Frequent sensor calibration is essential in sensor networks with low-cost sensors. We exploit the fact that temporally and spatially close measurements of different sensors measuring the same phenomenon are similar. Hence, when calibrating a sensor, we adjust its calibration parameters to minimize the differences between co-located measurements of previously calibrated sensors. In turn, freshly calibrated sensors can now be used to calibrate other sensors in the network, referred to as multi-hop calibration.

Up-to-date information on urban air pollution is of great importance for health protection agencies to assess air quality and provide advice to the general public in a timely manner. In particular, ultrafine particles (UFPs) are widely spread in urban environments and may have a severe impact on human health. However, the lack of knowledge about the spatio-temporal distribution of UFPs hampers profound evaluation of these effects. In this paper, we analyze one of the largest spatially resolved UFP data set publicly available today containing over 25 million measurements. We collected the measurements throughout more than a year using mobile sensor nodes installed on top of public transport vehicles in the city of Zurich, Switzerland. Based on these data, we develop land-use regression models to create pollution maps with a high spatial resolution of 100m × 100 m. We compare the accuracy of the derived models across various time scales and observe a rapid drop in accuracy for maps with subweekly temporal resolution. To address this problem, we propose a novel modeling approach that incorporates past measurements annotated with metadata into the modeling process. In this way, we achieve a 26% reduction in the root-mean-square error-a standard metric to evaluate the accuracy of air quality models-of pollution maps with semi-daily temporal resolution. We believe that our findings can help epidemiologists to better understand the adverse health effects related to UFPs and serve as a stepping stone towards detailed real-time pollution assessment.

Wireless sensor networks are increasingly used in application scenarios where a high data quality is inevitable, e.g., the control of industrial production areas. Nevertheless, many deployments must live with strict constraints regarding the sensing hardware and may not employ newest sensing technologies, e.g., due to limited energy budget, size, and bandwidth. Additionally, many applications would benefit from not only gathering absolute sensor readings but also knowing the quality of their low-cost sensor measurements. In this paper, we introduce a model-driven approach that (i) provides reliable accuracy bounds for individual noisy sensor readings and (ii) detects systematic and transient sensor errors. We apply our method to static and mobile real-world deployments of noisy and unstable low-cost sensors by analyzing large sets of urban temperature and ozone measurements. We find that the proposed algorithm successfully calculates precise accuracy bounds. We compare them to measurements of high-quality instruments and show that up to 96 % of the reference measurements are inside the computed accuracy bounds in the static scenario and up to 94 % in the mobile scenario. This is surprisingly high for the used low-cost sensors. By analyzing data from our static longterm deployment, we reveal that the ozone sensor's reliability is dependent on seasonal weather conditions.

Up-to-date information on urban air pollution, such as reliable pollution maps, is of great importance for health protection agencies to timely assess the air quality situation and provide advice to the general public. Ultrafine particles (UFPs) are widely spread in urban environments and believed to have severe impact on the human health. However, the lack of spatially resolved data hampers profound evaluation of these effects. In this work, we introduce one of the largest spatially resolved UFP data set available today, with over 25 million measurements to build high-resolution pollution maps for an urban area of 100 km2. The data is collected throughout more than one year using mobile sensor nodes, which are installed on top of public transport vehicles in the city of Zurich, Switzerland. We develop land-use regression models to create pollution maps with a high spatial resolution and study their temporal resolution limit.

We present the design of a reliable nurse call system based on wireless embedded devices and multi-hop protocols. Our work is motivated by the need for such system during annual summer camps for people with muscular dystrophy and the lack of suitable alternative solutions. We describe how our prototype meets the reliability and real-time requirements of such system, and report on results from a two-week deployment during a camp with 13 affected boys in July 2013.

Monitoring and managing urban air pollution is a significant challenge for the sustainability of our environment. We quickly survey the air pollution modeling problem,introduce a new dataset of mobile air quality measurements in Zurich, and discuss the challenges of making sense of these data.

Logging and tracing are important methods to gain insight into the behavior of sensor network applications. Existing generic solutions are often limited to nodes with a direct serial connection and do not provide the required efficiency for network-wide logging. Instead, this is often realized by application-specific subsystems developed for custom logging statements. In this paper, we present TinyLTS - a generic and efficient Logging and Tracing System for TinyOS. TinyLTS consists of a compiler extension that separates dynamic from static information at compile time, a declarative solution for inserting logging statements, an extensible framework for flexible storing and transmitting of logging data and a frontend for recombining dynamic and static information. Our system provides concise yet expressive programming abstractions for the developer combined with efficiency comparable to custom solutions.

Cycle-accurate emulation of sensor networks allows a detailed analysis of platform target code for development and evaluation. However, the high overhead incurred by providing the necessary fidelity limits the size of the emulated networks considerably. The use of multiple cores provided by modern hardware can significantly improve the speed of emulation but requires synchronization algorithms to preserve causality. Based on the well-known multithreaded event-driven emulator Avrora, we investigate a number of synchronization methods including an algorithm that does not require any locks to improve the performance. We show that both the speed and the scalability can be significantly improved without sacrificing correctness. Additionally, we evaluate the impact of modern CPU technologies such as simultaneous multithreading on emulation performance.

Boundary recognition is an important and challenging issue in wireless sensor networks when no coordinates or distances are available. The distinction between inner and boundary nodes of the network can provide valuable knowledge to a broad spectrum of algorithms. This article tackles the challenge of providing a scalable and range-free solution for boundary recognition that does not require a high node density. We explain the challenges of accurately defining the boundary of a wireless sensor network with and without node positions and provide a new definition of network boundary in the discrete domain. Our solution for boundary recognition approximates the boundary of the sensor network by determining the majority of inner nodes using geometric constructions, which guarantee that for a given d , a node lies inside of the construction for a d -quasi unit disk graph model of the wireless sensor network. Moreover, such geometric constructions make it possible to compute a guaranteed distance from a node to the boundary. We present a fully distributed algorithm for boundary recognition based on these concepts and perform a detailed complexity analysis. We provide a thorough evaluation of our approach and show that it is applicable to dense as well as sparse deployments.

The availability of secret keys is a precondition for the use of many security solutions and protocols. However, securely assigning such keys to nodes is a challenging task in the context of wireless sensor networks. In this paper we present a novel solution for a secure key assignment in wireless sensor networks that can be used during the initial configuration of nodes or for an ad-hoc key assignment by mobile nodes. The idea is to transmit the key information over a side channel using a controllable light source as the sender and the light sensors available on wireless sensor nodes as receivers. We demonstrate that our solution fulfills the relevant security requirements while at the same time being cost effective and easy to use.

Wireless sensor networks play an essential role in many pervasive computing scenarios as providers of context data. However, interacting with sensor nodes and selecting specific nodes for an interaction is difficult due to the constraints of the sensor node hardware. In this paper, we present and discuss three different approaches for this node interaction problem based on gestures, on light signals and on information provided by the sensor nodes using their LEDs. We demonstrate in our evaluation that these mechanisms solve the interaction problem for a variety of scenarios and effectively support the user in the interaction process. At the same time, our solutions set only low requirements on the wireless sensor node hardware.

Boundary recognition is an important and challenging issue in wireless sensor networks when no coordinates or distances are available. The distinction between inner and boundary nodes of the network can provide valuable knowledge to a broad spectrum of algorithms. This paper tackles the challenge of providing a scalable and range-free solution for boundary recognition that does not require a high node density. Our solution approximates the boundary of the sensor network by determining the inner nodes using geometric constructions that guarantee that, for a given d, a node lies inside of the construction for a d-quasi unit disk graph model of the wireless sensor network. Moreover, such geometric constructions make it possible to compute a guaranteed distance from a node to the boundary. We provide a thorough evaluation of our approach and show that it is applicable to dense as well as sparse deployments.

The lifetime requirements on wireless sensor networks often require the redundant deployment of sensor nodes with appropriate management mechanisms based on node clustering. Yet, existing clustering approaches do not take the primary task of sensor networks into account: performing relevant measurements. They usually form 'arbitrary' clusters, e.g., using connectivity information, and thus, the resulting measurements are often of only limited use to the applications. This problem can be avoided by considering application-specific semantics. For indoor applications, the notion of a room provides a natural unit of clustering since walls are constructed deliberately to ensure locality. This paper shows that it is feasible to automatically create clusters that reflect boundaries between rooms by analyzing the measurements of inexpensive, broadly available sensors. The paper first analyzes the applicability of statistical clustering methods and based on this analysis, it proposes and evaluates a lightweight approach to determine clusters in real deployments.

Virtual memory has been successfully used in different domains to extend the amount of memory available to applications. We have adapted this mechanism to sensor networks, where, traditionally, RAM is a severely constrained resource. In this paper we show that the overhead of virtual memory can be significantly reduced with compile-time optimizations to make it usable in practice, even with the resource limitations present in sensor networks.

In the area of wireless sensor networks, cross-layer interactions are often preferred to strictly layered architectures. However, architectural properties such as modularity and the reusability of components suffer from such optimizations. In this paper we present TinyXXL that provides programming abstractions for data exchange, a form of cross-layer interaction with a large potential for optimizations. Our approach decouples components providing and using data, and it allows for automatic optimizations of applications composed of reusable components. Its runtime representation is efficient regarding memory consumption and processing overhead