Menlo School

One of the great challenges of studying the coupled ionosphere-thermosphere system is the difficulty of making distributed in situ measurements simultaneously. Beginning with Birkeland's (1908) pioneering study of ionospheric currents via equivalent currents, many investigations of the complex, coupled, three-dimensional ionosphere-thermosphere system represent this system as a slab or thin shell either out of expedience or necessity. Almost all existing methods for assimilative reconstruction of ionosphere-thermosphere electrodynamics are based on such thin-shell representations. In this study, we use rocket-based measurements of neutral wind profiles and incoherent scatter radar measurements to directly calculate central height-integrated quantities in IT electrodynamics (perpendicular current, Joule heating, Hall and Pedersen conductance) and compare with estimates based on height-integrated quantities derived from the height-integrated Ohm's law and the expression for height-integrated Joule heating. It is shown that when an appropriate estimate of the neutral wind is included, estimates of these height-integrated quantities lie within ~20% of their true values. When the neutral wind is ignored (i.e., assumed to be zero in Earth's corotating frame of reference) estimates differ from their true values by as much as 100%.

Abstract Adversarial attacks in computer vision exploit the vulnerabilities of machine learning models by introducing subtle perturbations to input data, often leading to incorrect predictions or classifications. These attacks have evolved in sophistication with the advent of deep learning, presenting significant challenges in critical applications, which can be harmful for society. However, there is also a rich line of research from a transformative perspective that leverages adversarial techniques for social good. Specifically, we examine the rise of proactive schemes, methods that encrypt input data using additional signals termed templates, to enhance the performance of deep learning models. By embedding these imperceptible templates into digital media, proactive schemes are applied across various applications, from simple image enhancements to complicated deep learning frameworks to aid performance, as compared to the passive schemes, which don’t change the input data distribution for their framework. The survey delves into the methodologies behind these proactive schemes, the data perturbation and template learning processes, and their application to modern computer vision and natural language processing applications. Additionally, it discusses the challenges, potential vulnerabilities, and future directions for proactive schemes, ultimately highlighting their potential to foster the responsible and secure advancement of deep learning technologies.

Solar tracking systems dynamically adjust photovoltaic panel orientation to maintain optimal sun alignment, potentially enhancing energy harvest compared to fixed installations.This study investigates the design, construction, and performance validation of an automated single-axis solar tracker incorporating light-dependent resistor (LDR) sensors, servo motor actuation, and Arduino-based control.We quantified photovoltaic cell response with three key parameters: angle of incidence, spectral wavelength, and tracking versus static operation.The angle of incidence of incoming light was found to greatly affect the power output of the solar cell, which steadily decreased as angle increased; an orthogonal 0 had the highest output.We characterized the output of the solar cell on the spectral wavelength incident on the setup (630nm, 530nm, and 470nm wavelengths), and found discrepancies in the output power favoring longer wavelengths.A laboratory conducted experiment demonstrated nearly eight-fold power improvement (7.9 mW tracked versus 1.0 mW static).The field experiment that is conducted under California sun conditions over a full day (7:30 AM to 6:00 PM) revealed 53.92% increase to average power output (26.32 mW tracked versus 17.10 mW for static panel matched to altitude angle 36.95).The tracking system maintained superior performance throughout the day graph other than at solar noon, where the power values converged to around the same magnitude as the sunlight reached the point where it was directly incident upon the static solar cell.Temperature control throughout experiments eliminated thermal effects as confounding variables.These findings establish that automated solar tracking provides significant energy harvest enhancement. II.

Shared network testbeds are critical for systems and networking research. However, their shared hardware can introduce variability—like increased jitter or loss—that may impact experiment fidelity or reproducibility.

In this paper, we systemically disentangle BHG, a graph deep learning framework for daily users' ads recommendations. BHG mainly relies on two pillars: (1) graph tokenization to convert the input temporal heterogeneous graph into sequences of tokens, and (2) graph MLP-Mixer neural architecture to learn node representations on sequences of tokens via a mini-batch manner. In general, BHG embraces three advantages: (1) flexibility, i.e., BHG can be seamlessly integrated with any existing industrial recommendation model by treating the learned node embeddings as additional features that encode interactions, (2) efficiency, i.e., the graph tokenization allows sampling the neighborhood both locally and globally, and reduces the number of nodes considered for aggregations, and (3) model simplicity, i.e., the graph MLP-Mixer does not require self-attention for aggregating nodes and hence enjoys the simplicity. We demonstrate the superior performance of the proposed BHG on two internal datasets and one public dataset. We hope this paper can share insights and explain large-scale graph deep learning deployments for researchers, engineers, and practitioners.

Training high-quality deep learning models is a challenging task due to computational and technical requirements. A growing number of individuals, institutions, and companies increasingly rely on pre-trained, third-party models made available in public repositories. These models are often used directly or integrated in product pipelines with no particular precautions, since they are effectively just data in tensor form and considered safe. In this paper, we raise awareness of a new machine learning supply chain threat targeting neural networks. We introduce MaleficNet 2.0, a novel technique to embed self-extracting, self-executing malware in neural networks. MaleficNet 2.0 uses spread-spectrum channel coding combined with error correction techniques to inject malicious payloads in the parameters of deep neural networks. MaleficNet 2.0 injection technique is stealthy, does not degrade the performance of the model, and is robust against removal techniques. We design our approach to work both in traditional and distributed learning settings such as Federated Learning, and demonstrate that it is effective even when a reduced number of bits is used for the model parameters. Finally, we implement a proof-of-concept self-extracting neural network malware using MaleficNet 2.0, demonstrating the practicality of the attack against a widely adopted machine learning framework. Our aim with this work is to raise awareness against these new, dangerous attacks both in the research community and industry, and we hope to encourage further research in mitigation techniques against such threats.

Human beings perceive the real world through a spectrum of sensory modalities, encompassing auditory, visual, and linguistic faculties. This work proposes an industry-level omni-modal large language model (LLM) pipeline that integrates auditory, visual, and linguistic modalities to overcome challenges such as limited tri-modal datasets, high computational costs, and complex feature alignments. Our pipeline consists of three main components: First, a modular, end-to-end framework enabling flexible configuration of various encoder-LLM-decoder architectures. Second, a lightweight training strategy that pre-trains audio-language alignment on the state-of-the-art vision-language model Qwen2.5-VL, thus avoiding the costly pre-training of vision-specific modalities. Third, an audio synthesis pipeline that generates high-quality audio-text data from diverse real-world scenarios, supporting applications such as Automatic Speech Recognition and Speech-to-Speech chat. To this end, we introduce an industry-level omni-modal LLM, NEXUS-O. Extensive experiments validate the efficacy of our pipeline, yielding the following key findings: (1) In the visual understanding task, NEXUSO exhibits superior performance compared with its backbone model - Qwen2.5-VL-7B, validating the efficiency of our training strategy. (2) Within the English Spoken Question-Answering task, the model achieves better accuracy than the same-period competitor (i.e, MiniCPM-o2.6-7B) in the LLaMA Q. benchmark. (3) In our realworld ASR testset, NEXUS-O achieves outstanding performance, indicating its robustness in real scenarios. (4) In the Speech-to-Text Translation task, our model outperforms Qwen2-Audio-Instruct-7B. (5) In the Text-to-Speech task, based on pretrained vocoder (e.g., Fishspeech1.4 or CosyVoice2.0), NEXUS-O is comparable to its backbone vocoder on Seed-TTS benchmark. (6) An in-depth analysis of tri-modal alignment reveals that incorporating the audio modality enhances representational alignment between vision and language.

Abstract Based on experimentation, this paper outlines the efficiency and accuracy improvements that can be achieved using a cutting-edge generative AI chat enabled research assistant to facilitate well planning exercises. The technology innovations required to achieve those gains are also outlined. With the goal of measuring these improvements, during the summers of 2024 and 2025, the research team tested an experimental, domain aware agentic research assistant which could answer a question, cite the sources and show the researcher the original material it had used to determine the answer. Seventy-five (75) volunteers from thirty (30) countries participated in a simple well planning exercise answering a series of questions requiring analysis and interpolation of information contained in Equinor's Volve data set which consists of approximately 30,000 documents of mostly unstructured data (PDFs, etc.). To assist in answering the questions, the system had access to a good portion of those documents including well design documents, well logs, geological and stratigraphical data and operational reports (e.g. daily drilling reports) as well as SPE's PetroWiki. The system was benchmarked against a more standard but still state of the art RAG (Retrieval Augmented Generation) solution. By all measures, the experimental system outperformed the state-of-the-art RAG system, improving the accuracy of the answers by 152% and reducing the time required to answer by 86%. This paper explores the efficiency gained when a researcher, especially one with less industry experience, has access to a capable virtual research assistant which, in turn, has access to a broad set of documents covering the domain. The paper describes the limits of a state-of-the-art RAG system, the impact of building domain awareness into a novel system and the technology areas that make a difference in that system's usefulness when presented with the volumes of material common in the industry. While the technology continued to be improved in general, the next round of research focused on compiling answers across a large set of documents and finding information within plots and graphs. Similar testing, this time with testers who did not have oil and gas experience, showed that subjects achieved 86% correct answers with the improved solution vs 27% with an updated baseline RAG system. Finally, we performed ablation testing to understand the efficacy of various search approaches in achieving these results for various types of questions.

Abstract IntroductionIn this study we performed a retrospective chart review of alloimmunized pregnancies undergoing monitoring for hemolytic disease of the fetus and newborn (HDFN) in a cohort with known fetal antigen (FA) status. The objective was to compare fetal monitoring with titers and middle cerebral artery peak systolic velocity Doppler (MCA‐PSV Doppler) in FA‐negative and FA‐positive pregnancies and to understand downstream impacts of screening with titers in FA‐negative pregnancies.MethodsRetrospective chart review of alloimmunized pregnant patients who underwent FA cell‐free DNA (cfDNA) analysis and participated in a registry in which neonatal genotyping was performed to confirm FA cfDNA results. Patients were divided into FA‐positive or FA‐negative cohorts based on the FA genotypes. Medical records were reviewed to characterize monitoring for anemia, titer results, MCA‐PSV Doppler results, and whether interventions were performed. Fischer's exact or chi‐square tests were performed as appropriate to ascertain differences between the cohorts.ResultsSixty‐nine alloimmunized pregnant individuals were included. Forty pregnancies (58%) were FA positive. Of 29 FA‐positive pregnancies who had titers more than once, 20 (69%) had titers that rose over two or more time points. There were 29 FA‐negative pregnancies (42%). Of 20 FA‐negative pregnancies that underwent titers more than once, 10 (50%) had titers that rose over two or more time points. The FA‐positive and FA‐negative cohorts had similar proportions of pregnancies that reached critical titers—70% (28/40) and 69% (20/29), respectively. In the FA‐positive cohort 14 of 34 participants who underwent MCA‐PSV Doppler reached greater than 1.5 multiples of the median (MoM), suggestive of fetal anemia. One patient in the FA‐negative cohort reached 1.5 MoM.ConclusionsTiters frequently yield elevations that, while reflective of maternal alloimmunization, are not indicative of positive FA status or risk for HDFN in antigen‐negative pregnancies. Such results lead to unnecessary monitoring, including MCA‐PSV Doppler as well as invasive interventions. These outcomes can be avoided using methods including FA cfDNA, which has been shown to be highly accurate. These results suggest that monitoring for fetal anemia should be discontinued if FA cfDNA has determined the FA status is negative, as this will reduce downstream burdens and risks to alloimmunized patients who are not at risk for HDFN.