CGSW 10.0 Student Research Presentation Abstracts

To make heterogeneous multi-robot teams more robust in deployment, we cannot assume agents all have equal performance in sensing, actuation, or communication. Suppose we deploy a multi-robot team to visit specific points in an environment and take various measurements or collect samples. It is very possible a ground robot tasked with collecting samples becomes slowed by mud, or the camera of a drone capturing images becomes faulty. Conversely, some robots may perform better than expected, either with variations in design or different wear over time. Accounting for these individual variations will improve the overall team performance. While prior works accommodate instantaneous variations or differences in physical characteristics, we calculate the reputation, which one can generally think of as a history of performance.

We consider a heterogeneous team of n robots, i = {1, · · · , n}. Each robot has at least one of m capabilities, j = {1, · · · , m}, to serve events in a convex environment Q ⊂ R2, with points in Q denoted q. We invoke a weighted Voronoi-based coverage control approach (Figure 1) to assign regions of Q to robots, such that each robot is responsible for covering the events in their cell(s). This approach partitions the space into m partitions, assigning robot i to partition Pj if robot i has capability j. The probability that an event at point q will require capability j is represented by the density function ϕj (q).
Reputation to Improve Performance. As robots serve events, we assign each robot a reputation based on their ability to meet events. We use this reputation to adjust the weights of each robot cell, ultimately adjusting the amount of area a robot covers. We define the performance pj i of robot i for capability j within a time window as the ratio of the number of events served by robot i of type j, to the total number of events of type j that appear in the robot cell. We then define the reputation memory Mj i for robot i for each of its capabilities j as the set of the N most recent performance metrics pji . We define the reputation rji as the average of the reputation memory Mji. Intuitively, as a robot i serves more events requiring capability j, its reputation associated with capability j increases, and thus the cell size increases.
Estimating Event Probability. The density function ϕj (q) represents the probability an event requiring capability j will appear at any point q. Rather than assuming robots have knowledge of event density functions, we estimate the density functions over time based on current and past event appearances. This allows robots to adapt to evolving, stochastic demand while optimally positioning themselves to serve events. Robots estimate the density function ϕj (q) as events occur using kernel density estimation (KDE) to estimate the probability that events will occur in Q.

We carried out 25 randomized simulations each of our approach (with weights) and the comparison algorithm (without weights). As a performance metric, we recorded the total number of events met by the end of the trial. We observe our weighted approach outperforms the non-weighted approach, with a statistical significance of p < 0.001 in the scenario where m = 2, and p < 0.05 when m = 3. We therefore conclude that, by adjusting the size of a Voronoi cell based on robot reputation, the number of events the team meets as a whole increases.

Recently proposed Guessing Random Additive Noise Decoding (GRAND) has challenged these limitations by introducing a noise-centric approach to decode the data that is entirely independent of the codebook structure. GRAND relies on guessing the noise that affects a transmission rather than using the codebook structure. This feature makes GRAND a universal decoding algorithm that is especially suited for high-rate short-length codes. GRAND eliminates the need of standardization and provides a future-proof decoding solution. In soft-detection channel, in addition to the information bits, the confidence of the receiver in terms of the reliability of bits is provided. In such soft-detection channel scenario, Ordered Reliability Bits Guessing Random Additive Noise Decoding (ORBGRAND) provides the optimal ML decoding. In this work, we will demonstrate ORBGRAND in 40 nm CMOS featuring: 1) ultra-low energy consumption of 0.76 pJ/b and low power consumption of 4.9 mW compared to state-of-the-art softdetection decoders, 2) high fabricated throughput performance of 6.5Gbps while utilizing only a small core area of 0.4 mm2, 3) ability to abandon the sorting of soft information on-the-fly with dynamic clock gating, saving power and energy, while automatically adapting its performance to channel conditions; 4) energyand area-efficient integer partitions architecture for accurate ordered-reliability bit patterns with a Logistic Weight (LW) up to 104 and error correction up to 13-bits; and 5) reconfigurable architecture that supports codeword (CW) lengths between 32 to 256 bits. A demo of the chip will also be presented

Four different sets of assumptions are proposed for these mappings, varying in how stiffness is distributed throughout the arm, as well the linearity and/or hystersis of the actuator torques. We demonstrate how to calibrate each model from experimental data in a soft arm powered by shape memory alloy (SMA) wires that contract via Joule heating. Then, we perform hardware tests to predict the robot’s pose in open-loop, using only actuator temperature, and compare model performance under each simplifying assumption.

Through our experimentation, we discovered that we cannot assume the tested soft system can be described with a homogeneous bending stiffness even though the material of the system is homogeneous. Further testing demonstrated that calculating different torsional spring constants for different discretized links resulted in the most accurate estimations. Experimentation also demonstrated the hysteric behavior of the SMA’s and their dependency on temperature and torsional spring constants.

Future work plans to adapt the system from 2D to 3D. Furthermore, this model will be integrated with closed loop feedback control to enable real time responses to human and environmental contact. Using only temperature sensing to estimate the robot pose allows other forms of sensing, such as the bending angle, to be used to discern deviations from the predicted pose. This improves the detection of not just environmental contact, but also provides information about the nature of the contact, which in turn can inform higher level decision making and path planning.

To overcome this limitation, this work introduces a novel method to resolve up to 4-photon coincidences in single-photon detector arrays with row-column readouts. By utilizing unambiguous measurements to estimate probabilities of detection at each pixel, ambiguous multiphoton counts are redistributed among the candidate spatial locations to achieve a reduction in the mean-squared error (MSE) of the reconstruction compared to conventional techniques. With our method, we show that arrays can be operated at high incident photon fluxes, while simultaneously achieving an increase in the peak signal-to-noise ratio (PSNR) between 4-10 dB compared to previous methods. The application of this method to imaging natural scenes is demonstrated using Monte Carlo experiments in Fig 1.

The naïve reconstruction introduces horizontal and vertical streaks due to multiphoton coincidences. While the single-photon reconstruction is free of these artifacts, the reconstructed image is noisy. The multiphoton reconstruction (our method) achieves a PSNR improvement between 6-10 dB compared to the naïve reconstruction and 4-6 dB compared to the singlephoton reconstruction and is free from multiphoton artifacts. Note that these results were obtained at the same mean photons per frame (PPF) value of 3. The change in the performance of these methods with variation in incident photon flux is shown in Fig 2. The multiphoton
estimator achieves the lowest MSE at a PPF of ~1.4. In contrast, the lowest MSE for the singlephoton and naïve estimators are achieved at PPF values of ~1.1 and 0.35 respectively. This shows that in addition to producing better reconstructions, our method allows imaging at higher
incident photon fluxes as compared to conventional reconstruction techniques.

STL inference is the process of providing formal descriptions of system behaviors from observed data in the form of STL formulae. We present a neural-symbolic framework designed for the inference and prediction of time-series data. The neurons of the network can be translated into logic propositions, allowing the entire network to be transferred into an STL formula. The output of the neural network is a robustness degree of the STL formula over the input signal. This degree quantifies the satisfaction or violation of the signal concerning the STL formula translated by the neural network. The training process involves optimizing the neural network parameters to achieve a robustness degree that satisfies certain predefined requirements. However, the computation of the robustness degree includes multiple non-differentiable operations. To address the challenge of non-differentiability within STL, we propose various approximation methods to align with gradient based methods. Importantly, this adaptation does not compromise the qualitative guarantees. The proposed neural network is a template-free model that can dynamically adapt to diverse structures without fixing specific elements of the temporal logic formula. Our overarching goal is to offer a versatile tool that can not only serve as an interpretable neural network but also integrate seamlessly with other machine learning techniques. The interpretable nature is crucial for applications where understanding the decision-making process is essential, such as in autonomous vehicles.

This work has achieved great success in classification problems, including binary classification and attribute-based multi-class classification. For binary classification, our neural network classifies desired and undesired behaviors such that the desired behaviors satisfy the inferred STL formula, while undesired ones violate it. For multi-class classification, our neural network extends its capabilities to classify signals based on specified attributes. These attributes denote shared features among signals across various classes. Furthermore, it can also contribute to the design of controllers by inferring temporal logic formulae from observed system trajectories. This inference process aims to discern specifications governing the system’s behavior. For future work, our research can be extended to predict trajectories that adhere to the identified temporal logic specifications.

Keywords— Machine Learning, Neural Network, Formal Methods, Temporal Logic, Classification

Continuing the macroscale analogy above, if the gymnast doubles their force, the beam will bend twice as much. This is a response of a linear system. However, nano-beams and NEMS generally do not behave linearly. Due to many contributing factors in real-world situations, such as geometric, material and dissipation-related nonlinear effects, nanoscale sensors are governed by more complicated, nonlinear laws that remain elusive. However, they are important for enabling precise control and characterization of these devices in order to make them a reliable medical device or a viable commercial product.

To discover these nonlinear governing laws, I implement a machine-learning approach called sparse identification of nonlinear dynamics (SINDy), which is based on the thresholded linear regression model that is able to recover a symbolic form of a time series data. Briefly, the algorithm works as follows. I first build a Michelson interferometer that allows me to take ultrasensitive measurements of a NEMS device under known perturbations in the time domain. This is an experimental step. Then, I define a search space of potential candidates for a system of nonlinear governing equations. I sparsify the system while maintaining fit accuracy and eventually converge to an interpretable system of nonlinear differential equations that gives rise to my experimental data set. This is the machine learning step. To confirm the results, I introduce Pareto front analysis for leveraging model simplicity and fit accuracy, and utilize
established analytical expressions to guide the machine learning process.

In a broader perspective, this cross-disciplinary, physics-informed approach is potentially applicable to many nonlinear systems beyond nanoscale systems, such as climate change, disease modeling and even fluctuations in stock markets.

Motivated by introspection frameworks focusing on virtual machine (VM) settings and ML methods for software discovery, we design PraxiPaaS as a framework to inspect PaaS container images with a highly scalable ML inference pipeline by scanning file changes during package installations. Our ML model adopts a discovery-byexample approach, utilizing container layer file system changes as package fingerprints, matching them with trained ones collected from test containers. The ML pipeline includes a bagging of word2vec encoders and a corresponding bagging of ML models, i.e., XGBoost. Specifically, we modify the bagging method to train each submodel to recognize a fixed number of packages. This design allows optimizing the balance between incremental training time and F1 score by configuring the number of packages attributed to each submodel. Our evaluation shows the bagging of models provides an exponential drop in incremental training time from 30 minutes to 0.5s with 32 CPU cores while maintaining an F1 score of 0.85 using 10 package labels per model across different datasets, compared with the previously used single ML model design. We deploy a prototype of PraxiPaaS in the New England Research Cloud (NERC) OpenShift test cluster as well.

Keywords— Software Discovery, PaaS, XGBoost, Bagging, Incremental Training

ALBADross is a novel active learning-based framework that diagnoses previously encountered performance anomalies in HPC systems using significantly fewer labeled samples compared to state-of-the-art ML-based frameworks. Our framework combines an active learning-based query strategy and a supervised classifier to minimize the number of labeled samples required to achieve a target anomaly diagnosis score. We evaluate our framework on a production HPC system and a testbed HPC cluster using real and proxy applications. We show that our framework, ALBADross, achieves a 0.95 F1-score using 28x fewer labeled samples compared to a supervised approach with equal F1-score, even when there are previously unseen applications and application inputs in the test dataset.

To further reduce the reliance on labeled anomalous data and demonstrate practical deployment, we introduce Prodigy, a variational autoencoder-based anomaly detection framework. Prodigy outperforms the state-of-the-art alternatives by achieving a 0.95 F1-score when detecting performance anomalies. We also provide a real system implementation of Prodigy that enables seamless integration with monitoring frameworks and facilitates swift deployment in real-world settings. We deploy Prodigy on a production HPC system and demonstrate 88% accuracy in detecting anomalies. Prodigy involves an interface to provide job- and node-level analysis and explanations for anomaly predictions.

IOMMU protection comes at the cost of reduced throughput in IO-intensive workloads, mainly due to the high IOTLB invalidation latency. The Linux OS eliminates this bottleneck by deferring the IOTLB invalidation requests to a later time. This opens a vulnerability window during which a memory region is unmapped but the relevant IOTLB entry remains. In this work, we present a proof-of-concept exploit, empirically demonstrating that a malicious DMA-capable device can use this vulnerability window to leak data used by other IO devices. To address the need for secure, low-overhead DMA operations for high-throughput IO workloads (e.g., multi-threaded writes to NVMe storage, 200 Gbps network), we propose a hardware-assisted mitigation for the deferred invalidation vulnerability. The proposed mitigation improves the overall IO throughput compared to strict invalidation while providing the same security guarantee by preventing the reuse of stale IOTLB entries until the IOTLB is invalidated. The proposed mitigation for deferred invalidation vulnerability works at the IOMMU hardware level, making it compatible with any DMA operation.

We demonstrate the proof-of-concept exploit on an x86 machine emulated by QEMU. The emulated machine contains two IO devices — a storage device and an Intel 82574L NIC — and an Intel IOMMU. We show that the malicious NIC can take advantage of the deferred invalidation of its unmapped memory to access data that was read from the storage device. We also implement and evaluate the proposed mitigation within the QEMU Intel IOMMU implementation using the same QEMU setup we used for the exploit. The proposed mitigation achieves 12.7% higher throughput compared to the strict invalidation mode while providing the same security guarantees. We chose QEMU for the implementation and evaluation because support for an IOMMU is not available in open-source hardware. Additionally, we demonstrate that the variations in the average network throughput of QEMU in different IOMMU modes follow the same trend as that of a real hardware system.

Keywords: IOMMU, DMA attacks, IOTLB, deferred invalidation