United States

How best to leverage small datasets for machine learning tasks is an active area of research. Synthetic data has the potential to alleviate object rarity and long-tail distributions, provided the synthetic data introduces more signal than noise into the system. The IQT Labs Synthesizing Robustness project aimed to explore whether domain adaptation strategies (attempting to make the synthetic data look more “realistic”) can improve the efficacy of the synthetic data when it comes to localizing rare objects. In this project we explored generative adversarial network (GAN) techniques for domain adaptation on the publicly available RarePlanes satellite imagery dataset. 

Over a number of experiments we quantified the detection performance for rare aircraft as a function of training size per object class. We also note that strategically augmenting the real dataset is nearly as effective as adding synthetic data in the quest to improve the detection or rare object classes. Intriguingly, in certain cases, confusion between similar-appearing object classes is a more important factor than training dataset size in determining detection performance. Finally, we share some of the lessons learned from this project; for example: fully extracting the utility of synthetic data is a tricky and nuanced process, requiring effort and creativity.

NAML 2022

Synthesizing Robustness: On The Nuances of Extracting Utility from Synthetic Data

satellite imagery

synthetic data

computer vision

**Welcome to EMNLP 2023!**

On behalf of the EMNLP 2023 Organizing Committee, I extend a warm and heartfelt welcome to all of you to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). It is with immense pleasure and excitement that we gather here in Singapore, a vibrant hub of innovation and technological advancement.

The conference program is packed with insightful presentations, thought-provoking workshops, and engaging networking opportunities. In addition to the technical sessions, we have also planned several social events that will provide you with the opportunity to connect with your colleagues.

The conference is held in person at the Resorts World Convention Centre in Singapore, and available on-line with the help of Underline.

For those in Singapore, I hope that you will find time to explore the exciting Sentosa Island, Gardens by the Bay, Singapore Botanic Gardens and the many other attractions unique to Singapore.

Of course, we are grateful to our sponsors and partners for their generous support of EMNLP 2023, whose contributions make it possible for us to host this world-class event.

Yuji Matsumoto (RIKEN AIP)<br>
EMNLP 2023 General Chair<br><br><br><br>

To access the **EMNLP 2023** event page on Underline, you need to register for the Conference.<br>
Please follow **[this link](https://2023.emnlp.org/registration/)** for more details. 

EMNLP 2023

EMNLP 2023 took place in Singapore from Dec 6th to Dec 10th, 2023.

technical paper

Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.

The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team.

The goals of NAML are:

To spread awareness of current machine learning research relevant to Naval applications

To connect machine learning researchers with experts in military requirements

To build and strengthen collaborations within the defense research community


<figure class="video_container">
  <iframe src="https://s3.amazonaws.com/pf-user-files-01/u-59356/uploads/2022-03-22/hk13mxo/NAML_ReceptionWithtitle.mp4" width=826  height=526  frameborder="0" allowfullscreen="true"> </iframe>
</figure>

Please log in to explore this event. You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

Registration for this workshop has now closed. We appreciate your interest and hope to welcome you at a future event. If you’d like to hear about similar upcoming events, please check our website or join our mailing list.

Registration for this workshop has now closed.

Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team.

The goals of NAML are:

To spread awareness of current machine learning research relevant to Naval applications

To connect machine learning researchers with experts in military requirements

To build and strengthen collaborations within the defense research community



In this talk, we tackle a generic n-dimensional numerical computing problem and explore performance and tradeoffs between popular frameworks based on open source jupyter notebook examples. Most data science practitioners perform their work in Python because of its high level abstraction and rich set of numerical computing libraries. But the choice of library and methodology is driven by complexity-impacting constraints like problem size, latency, memory, SWAP, hardware, and others. To that end, we demonstrate that a wide selection of GPU-accelerated libraries (RAPIDS, CuPy, Numba, Dask, cuNumeric), including the development of hand tuned CUDA kernels, are accessible to data scientists without ever leaving Python.

Evaluating Your Options for Accelerated Numerical Computing in Pure Python

The work presented describes how real-time geospatial image processing can be accomplished on a SWaP (size, weight, and power)-constrained edge system. Traditionally, the data processing requirements for these workflows end-to-end are compute-intensive, and therefore challenging to deploy to power-constrained edge-based hardware. The edge platform we use is a self-hosted BF2 DPU paired with an A100 GPU, allowing for a low-power, high-performance edge solution without a host CPU. We'll demonstrate how common geospatial processing workloads such as orthorectification can be GPU-accelerated and launched from our edge BF2 system.

High Performance Geospatial Image Processing at the Edge

It is critical to always have an accurate estimate of the position of aircraft on the carrier deck. Currently, determining the pose of aircraft is done via visual inspection by several sailors who then update a digital interface with any changes to the position or status of aircraft. This task is labor intensive, requires focus for extended periods of time, and requires multiple personnel stationed during all flight operations. PATRIOT (Panoramic Asset Tracking of Real-time Information for the Ouija Tabletop) is a research endeavor to apply computer vision techniques to existing camera feeds from the carrier to identify and track aircraft, personnel, and vehicles. This work focuses on two branches used in the overall PATRIOT pipeline that estimate the pose of aircraft from synthetic and real videos: keypoint-based techniques and bounding box-based techniques. Several state-of-the-art keypoint-based methods (High-Resolution Net, Higher High-Resolution Net, and OpenPifPaf) and accompanying pose estimation algorithms are compared with a bounding box-based method using Faster R-CNN and an encoder-decoder model.

Pose Estimation Techniques for PATRIOT (Panoramic Asset Tracking of Real-time Information for the Ouija Tabletop)

The intensity of atmospheric turbulence has a profound effect on the performance of laser weapon systems. Proper configuration and operation of these systems requires nearly real-time knowledge of the atmosphere’s turbulent intensity. Determination of this characteristics is, however, difficult in operational settings. Models exist based on macro-meteorological measurements but these have been shown to have questionable accuracy, especially in the maritime and near-maritime environment. Current techniques for directly determining atmospheric turbulent intensity are also impractical in an operational setting, typically requiring two-sided systems with both a transmitter and receiver, one at the source and one at the target. For these reasons alternative techniques for determining the intensity of atmospheric intensity are desirable. Specifically, a one-sided system, only requiring information acquisition from the source side would be particularly attractive.
The disruption of laser weapon performance due to atmospheric turbulence affects other phenomenon based on the propagation of light, namely imaging. As turbulent intensity increases, the quality of images taken through the atmosphere (a one-sided technique) decreases. Consequently, images blurred by atmospheric turbulence carry information about its intensity. As atmospheric turbulence increases, image quality decreases. The current work uses convolutional neural networks (CNN), in combination with a synthetically generated data set of turbulently blurred images, to predict turbulent intensity from only blurred images. Specifically, a training and validation data set is created using a pristine image which is subsequently blurred using software that simulates the effect of the turbulent atmosphere. The data is then used to train CNNs to predict the turbulent intensity of “unknown” images in the validation set. The results are encouraging and demonstrate the ability of CNNs to predict the intensity of atmospheric turbulence using only a blurred image.

Determining the Intensity of Atmospheric Turbulence Using Blurred Images and a CNN

Many — if not most — applications of artificial intelligence (AI) involve accurately tracking and identifying objects in the physical world. For example, shipboard systems could autonomously recognize and track nearby vessels, self-driving vehicles could identify and track cars and pedestrians, and search-and-rescue systems could automatically locate and track the lost or wounded. However, current tracking techniques have significant weaknesses that limit their applicability and robustness. Most machine learning systems require large-scale training datasets that are difficult or impossible to collect in many applications. These systems are also easily confused by novel scenarios and are vulnerable to adversarial attacks. To overcome these limitations, we draw inspiration from biological visual systems, which are remarkably robust in their ability to track and identify objects, even those that have not been seen before. In this talk, we will describe our current progress building biologically-inspired tracking systems. Our ultimate goal is to lay the groundwork to develop potentially revolutionary visual systems that are more robust and generalizable than today’s techniques.

This work was sponsored by a grant from Naval Surface Warfare Center Crane Division.

Building biologically-inspired object tracking systems

In the current rapidly developing environment of both technologies and techniques in electronic warfare (EW), it is essential for modeling and simulation (M&S) frameworks to keep up with the pace. When it comes to creating an accurate M&S environment, data is king. Collecting the necessary data through flight tests is not feasible due to the cost, availability of assets, and need for rapid turnaround. 
There are many tools that allow for quick creation of RF test data for the M&S environment, but the difference between testing on lab generated data and the real-world data can be the difference between life and death in the field. This is especially true for future cognitive EW systems which, as with all machine learning (ML) systems, will perform worse on noisy real-world data. The ML technique called Style Transfer is a powerful tool that can be leveraged in the EW M&S domain to improve the testing and evaluation (T&E) of the different systems involved in the kill chain.
Style transfer was developed as a tool for capturing the art styles of different prominent artists and transferring that style to other paintings and photographs. This model can show you what the Mona Lisa would look like if painted by Picasso, van Gogh, or Monet. This same technique can be used to modify lab generated, digitized radio frequency (RF) data to look more realistic. The styles that can be captured in this domain include the emitters, environmental noise, and receivers. Multiple styles can be applied to a single signal, which would allow a “clean” lab signal to be modified to look as if it were generated from a specific emitter, traveled through a noisy environment, and was received by a specific receiver. This technique can be applied to any digitized RF data, so it can be used to improve existing data sets. Also, with the rapid improvement of hardware, a model using this technique may be able to act as a real-time filter on generated data in M&S frameworks.
Style transfer provides a way to improve the realism of lab-generated RF data to improve the confidence in the results of T&E events. The technique is useful not only for current systems, but will continue to be relevant as new systems enter the battlefield, including cognitive systems.

Generating Realistic RF Test Data Through Style Transfer

Deep learning algorithms are powerful and present an opportunity to achieve high performance in computer vision applications. Unfortunately, these algorithms also require large quantities of data that can be difficult to assemble, particularly in the case of Defense applications. Obtaining conventional imagery that is relevant to defense applications is expensive and time-consuming. Synthetic data, therefore, has emerged not only as a viable option, but as a necessary component of defense algorithm development strategy. The U.S. Army Futures Command DEVCOM C5ISR Center, Research and Technology Integration Directorate, Modeling and Simulation Division is developing a variety of synthetic data generation capabilities. One of these approaches utilizes fully synthetic imagery generated by physics engines with high-fidelity inputs and models. In particular, we have created an IR image generation pipeline using the UnrealEngine (UE4) and an IR plugin. In order to evaluate the suitability of this synthetic imagery for algorithm training, we used it to train a version of YOLO to detect and classify real targets in field-collected imagery. This presentation outlines the image generation pipeline, the performance of various iterations of the UE4 imagery when evaluated against real data, and future steps for synthetic imagery development.

Premium content

Downloads

Next from NAML 2022

Evaluating Your Options for Accelerated Numerical Computing in Pure Python

Similar lecture

Synthetic Data Generation with Large Language Models for Text Classification: Potential and Limitations