A5087765289- Cloud Computing and Resource Management
- Time Series Analysis and Forecasting
- Advanced Chemical Sensor Technologies
- Distributed and Parallel Computing Systems
- Target Tracking and Data Fusion in Sensor Networks
- Advanced Data Storage Technologies
- Fault Detection and Control Systems
- Scientific Computing and Data Management
- Data Stream Mining Techniques
- Mass Spectrometry Techniques and Applications
- Music Technology and Sound Studies
- Simulation Techniques and Applications
- Big Data and Business Intelligence
- Music and Audio Processing
- Neural Networks and Applications
- Anomaly Detection Techniques and Applications
- Biosensors and Analytical Detection
- Parallel Computing and Optimization Techniques
- Gas Sensing Nanomaterials and Sensors
- Software System Performance and Reliability
- Acoustic Wave Phenomena Research
- Water Systems and Optimization
- Speech and Audio Processing
- Data Quality and Management
- IoT and Edge/Fog Computing
MIT Lincoln Laboratory
2022
Worcester Polytechnic Institute
2018-2021
Moscow Institute of Thermal Technology
2021
Production high-performance computing (HPC) systems are adopting and integrating GPUs into their design to accommodate artificial intelligence (AI), machine learning, data visualization workloads. To aid with the operations of new existing GPU-based large-scale systems, we provide a detailed characterization system operations, job characteristics, user behavior, trends on contemporary GPU-accelerated production HPC system. Our insights indicate that pre-mature phases in modern AI workflow...
Artificial intelligence (AI) and Machine learning (ML) workloads are an increasingly larger share of the compute in traditional High-Performance Computing (HPC) centers commercial cloud systems. This has led to changes deployment approaches HPC clusters cloud, as well a new focus on optimized resource usage, allocations AI frameworks, capabilities such Jupyter notebooks enable rapid prototyping deployment. With these changes, there is need better understand cluster/datacenter operations with...
Chemical sensors play an important role in a variety of civilian and military domains. In these contexts, the ability to accurately quickly identify chemical agents is utmost importance. practice, constraints on physical footprint, power consumption, ease use, time required for accurate detection often restrict utility sensors, particularly remote isolated regions. One solution address this problem engineering advanced signal processing techniques, which decrease detection. This allows...
This paper considers the problem of tracking and predicting state a dynamic system with stochastic dynamics multiple modes operation. A well-known approach to this is "interacting model" (IMM) estimator, which uses knowledge different operation update bank Kalman Filters (each optimal for given mode operation). The IMM combines estimates according posterior probability modes. Despite their popularity, IMMs are known sometimes be slow detect switching, however, can result in large estimation...
Filtering is the process of recovering a signal, x(t), from noisy measurements z(t). One common filter Kalman Filter, which proven to be conditional minimum variance estimator x(t) when are Gaussian random processes. However, in practice (a) not necessarily and (b) an estimation measurement covariance problematic often tuned using cross-validation domain knowledge. In order address Filter's suboptimal performance situations where non-Gaussian noise present, we train deep autoencoder learn...
Given a noisy signal, it is often of interest to estimate its noise-free state. One the more common state estimation techniques Kalman Filter, which optimal under certain conditions, one that measurements are Gaussian random process with known covariance R. However, in practical applications may not be or noise Gaussian. Accordingly, here we employ method called Autoencoder-Kalman Filter (AEKF) learn mapping from inputs for Filter. Training AEKF uses technique domain randomization and has...
High-Performance Computing (HPC) centers and cloud providers support an increasingly diverse set of applications on heterogenous hardware. As Artificial Intelligence (AI) Machine Learning (ML) workloads have become larger share the compute workloads, new approaches to optimized resource usage, allocation, deployment AI frameworks are needed. By identifying their utilization characteristics, HPC systems may be able better match available resources with application demand. leveraging...
American leadership in AI. These broad strategy documents have influenced organizations such as the United States Department of Air Force (DAF). The DAF-MIT AI Accelerator is an initiative between DAF and MIT to bridge gap researchers mission requirements. Several projects supported by are developing public challenge problems that address numerous Federal research priorities. challenges target priorities making large, AI-ready datasets publicly available, incentivizing open-source solutions,...
Chemical and biological agents remain a persistent threat to both Warfighters civilian personnel in combat non-combat environments. While many useful detection technologies exist provide early warning, of them suffer from similar drawbacks. Most common are unacceptable false positive response rates inability discriminate target analytes complex or obscurant-laden A typical approach this problem is develop material solutions for more specific sensitive sensors. useful, these often too...
Through a series of federal initiatives and orders, the U.S. Government has been making concerted effort to ensure American leadership in AI. These broad strategy documents have influenced organizations such as United States Department Air Force (DAF). The DAF-MIT AI Accelerator is an initiative between DAF MIT bridge gap researchers mission requirements. Several projects supported by are developing public challenge problems that address numerous Federal research priorities. challenges...
In this paper we address the application of pre-processing techniques to multi-channel time series data with varying lengths, which refer as alignment problem, for downstream machine learning. The misalignment may occur a variety reasons, such missing data, sampling rates, or inconsistent collection times. We consider collected from MIT SuperCloud High Performance Computing (HPC) center, where different job start times and run HPC jobs result in misaligned data. This makes it challenging...
In this paper we address the application of pre-processing techniques to multi-channel time series data with varying lengths, which refer as alignment problem, for downstream machine learning. The misalignment may occur a variety reasons, such missing data, sampling rates, or inconsistent collection times. We consider collected from MIT SuperCloud High Performance Computing (HPC) center, where different job start times and run HPC jobs result in misaligned data. This makes it challenging...
Chemical agents remain a persistent threat to both Warfighters and civilian personnel in combat non-combat environments. While many useful detection technologies exist provide early warning, of them suffer from similar drawbacks. Most common are unacceptable false positive response rates inability discriminate target analytes complex or obscurant-laden A typical approach this problem is develop material solutions for more specific sensitive sensors. useful, these often too expensive, not...
Artificial intelligence (AI) and Machine learning (ML) workloads are an increasingly larger share of the compute in traditional High-Performance Computing (HPC) centers commercial cloud systems. This has led to changes deployment approaches HPC clusters cloud, as well a new focus on optimized resource usage, allocations AI frame- works, capabilities such Jupyter notebooks enable rapid prototyping deployment. With these changes, there is need better understand cluster/datacenter operations...
Traditional frequency based projection filters, or operators (PO), separate signal and noise through a series of transformations which remove frequencies where is present. However, this technique relies on priori knowledge what contain that these do not overlap, difficult to achieve in practice. To address issues, we introduce PO-neural network hybrid model, the Pseudo Projection Operator (PPO), leverages neural perform selection. We compare filtering capabilities PPO, PO, denoising...