Seedling Information

There are a variety of opportunities to engage with IARPA, and a seedling is another option to receive funding for your ideas. Seedlings are typically 9-12 month research efforts intended to “take an idea from disbelief to doubt,” or rather, to help to determine whether a full research program (lasting 3-5 years) is warranted. Most seedlings stem from proposals received through the General Broad Agency Announcements that are released annually.

Submit Your Idea

IARPA Distribution and Evaluation System (IDEAS) is a web-based software tool that enables online submission of unclassified proposals submitted in response to IARPA research solicitations.

If you are interested in learning more or have questions about a particular seedling, please contact (301) 851-7500 or


Examples of Previous Seedling

Algorithms for Learning Using Privileged Information

Learning Using Privileged Information (LUPI) is a novel approach to improving machine learning by supplementing training data with additional data (“privileged information”). In a LUPI model, some of the training examples are augmented with additional comments, comparisons, or other useful information. This additional information has been shown theoretically to speed up the rate of convergence for learning. In a recent study, LUPI-based Support Vector Machines (SVM+) were found to achieve the same or better accuracy as an SVM model with a significant reduction in the amount of training data required. One challenge that this approach faces is that it requires 10 times the computation time as an SVM, which could reduce the practicality of the approach for many applications. The goal of the seedling is to reduce computation time to SVM levels without eliminating the benefits of the SVM+ approach and to further explore the types of privileged information that could benefit the IC.

Performer: Applied Communication Sciences

Resulting Research: FA8650-11-C-7161

Analog CMOS

Transmutational materials change over time, causing a related change in bulk material properties such as electrical and thermal conductance or mass density. Where such processes are reliable and unaffected by physical effects such as electromagnetic and mechanical forces and temperature, the change in bulk properties can be utilized as a reliable measure of time or as a chip attribution and identification scheme. Used with integrated circuits, such materials have potential applications for: autonomous, self-timed fuse, chip attribution and identification, and low-power time measurement. This seedling identified and studied suitable transmutational materials, developed CMOS-compatible processes for use in integrated circuits, and identified, characterized, and demonstrated potential applications.

Performer: NanoMason

Resulting Research: N66001-C-13-2004

Augmented Reality-based Collaborative and Corroborative Collection (ARC3)

The Augmented Reality-based Collaborative and Corroborative Collection (ARC3) seedling is an effort to analyze the requirements for designing, building and demonstrating an augmented-reality system that will work seamlessly with humansP to collaboratively and corroboratively collect information in dynamic real-world environments. As part of this effort, the performer (SRI International) will (1) study key needs, specifications, high-risk areas and potential approaches to building an end-to-end ARC3 system; (2) empirically study the implications of a dynamic information aperture with respect to trade-offs between efficiency, usability and reliability when automated, versus user-driven approaches are followed. This will be accomplished by developing experimental prototype interfaces for AR collaboration and corroboration between an analyst and a collector; and (3) empirically study techniques for automated and user-driven information extraction in the presence of uncertain and unreliable visual localization and scene understanding technologies. An experimental prototype will build user scenarios in which the interplay between audio-visual-text processing and user interaction can be systematically evaluated.

Performer: Johns Hopkins University Applied Physics Laboratory, SRI International

Resulting Research: 2012-120508000010 OR FA8650-14-C-7430

Automated Location Identification

This seedling will explore automatic methods to determine the geographic location of a scene that has been recorded in video and still images. A unique aspect of this work is the utilization of existing publicly-tagged data to create a hierarchical tree model of a region structured by city, neighborhood, and block and leverage that information for search and retrieval. This seedling’s results support the Finder program.

Performer: University of Central Florida

Resulting Research: N10PC20231

Automatic Speech Recognition (ASR)

This research project proposes to lay the groundwork for future research in Automatic Speech Recognition (ASR) technology by quantitatively and qualitatively exploring what is wrong with ASR.
The researchers plan to quantitatively analyze the sources of failure in current acoustic modeling approaches. The focus of the ASR research community has been on improving performance with increasing amounts of matched-condition training data and language modeling. Using a diagnostic approach, their prior research has revealed that many errors are due to mismatch between the data and the statistical assumptions built into the acoustic models. A key idea in the current effort is to use resampling at different levels to systematically modify the data in order to selectively match particular assumptions in the model. This will be used to discern specific error conditions and the extent of their contribution to overall error rate as well as providing a quantitative metric for comparing data sets with regard to factors affecting performance. An important aspect of this research is training/test mismatch (a condition in which current ASR systems fail quite badly). In addition, the research team will ask experts in the field where they believe ASR technology is effective, where it fails, and what its shortcomings are. This survey will include both interviews and a review of the literature. This research effort will produce new quantitative methods and metrics for ASR acoustic model failure analysis, a report on failure modes and lessons learned, and suggestions for future research. The software developed in the research will be made freely available to the research community.

Performer: SRI International

Resulting Research: FA8650-15-C-9101

Collapse Immune Collaboration Environment for Research Oversight (CICERO)

Use computational models to explore the interaction dynamics of (simulated) groups of collaborating intelligence analysts. The goal is to use the models to examine under what conditions groups are more likely to succumb to "cognitive collapse," and what interventions if any might be used to steer groups away from this undesirable state. "Cognitive collapse" refers to a state in which a group becomes fixated on a single hypothesis, failing to take into account contradictory evidence or to consider alternative hypotheses. The seedling completed at the end of 2QFY2011, with results demonstrating that the simulation-based tools provide a viable approach to exploring this type of group behaviors.

Performer: Jacobs Strategic Solutions Group, Inc.

Resulting Research: W911NF-10-C-0108

Data Engine for an Analyst's Workbench

Develop a high-performance and general-purpose Data Engine that will enable a workbench to operate on very large amounts of uncertain data, and that will provide the Inference Engine with just the data it needs in a usable form. The proposed Data Engine has three key features. First, the engine will store and manage uncertain data, so uncertainty will be an integral part of its data model and algorithms. Second, the engine will be general purpose, meaning that the engine can be used in a variety of workbench settings, by many different inferencing algorithms, and by different applications. Thus, the engine will support a data model, query language, and interface that are appropriate for a wide spectrum of needs. Third, the engine will be high performance and it will be able to efficiently process vast amounts of data, often residing on disk or slow memory. The user will have the option of accepting approximate results in order to improve running times further.
Language Computer Corporation is exploring the automatic identification and characterization of emerging events as reported in textual information feeds, such as news reports. Complex events of the sort of interest to the IC are made up of multiple component events and contextual influences which over time can converge into an action or event of interest. By identifying component events and tracking them over time, event structures can emerge, supporting indicators and warnings to analysts as an input to their task prioritization process.

Performer: Stanford University

Resulting Research: N10PC20259

Dynamic Deep Displays

In 2011, the Dynamic Deep Displays seedling successfully demonstrated feasibility of photo-refractive materials for update-able holographic displays, and over its 12-month period increased brightness, image speed and energy efficiency all by a factor of 10. Because of its compelling implications (no glasses required) to the future of 3D displays, Innovation Magazine identified this technology in the top three developments of 2010, right behind the iPad. The seedling permitted characterization of performance requirements for holographic displays allowing collaborative analysis with no head-gear, and the quantification of many metrics that formed the Synthetic Holographic Observation program. This is a good example of a seedling being used in the creation of a program.

Performer: Sandia National Lab

Resulting Research: N/A


Language Computer Corporation is exploring the automatic identification and characterization of emerging events as reported in textual information feeds, such as news reports. Complex events of the sort of interest to the IC are made up of multiple component events and contextual influences which over time can converge into an action or event of interest. By identifying component events and tracking them over time, event structures can emerge, supporting indicators and warnings to analysts as an input to their task prioritization process.

Performer: Language Computer Corporation

Resulting Research: W911NF-10-C-0096

Electrically Small Antenna Development

The Electrically Small Antenna Development seedling is developing non-Foster broadband impedance matching circuits using negative resistors. Current non-Foster matching networks use negative capacitors and negative inductors, for which it is a significant challenge to maintain linearity and stability. Using negative resistors should provide fewer parasitic effects, lower current and voltage magnitudes, and improve linearity. The project includes analysis, simulation, design, building, and testing of broadband matching networks for receive and transmit antenna matching. The testing will compare the performance without matching networks and with active matching circuits.

Performer: Northrop Grumman Corporation

Resulting Research: N66001-12-C-2015

Embodied Common Sense in Vision and Language

eCortex, Inc. is exploring the hypothesis that when a computer vision system patterned after the visual system of the mammalian brain is trained on realistic inputs (e.g., 3D imagery and video data) as opposed to the simpler inputs typically used (e.g., static 2D images), it will automatically learn useful “common sense” rules for recognizing objects in cluttered scenes. Indeed, the results showed that, after the system learned the depth cues from realistic 3D imagery, performance increased for 2D imagery where the system successfully applied the 3D depth cues absent in the 2D images. These insights may guide the development of significantly more sophisticated automated image analysis systems.

Performer: eCortex, Inc.

Resulting Research: W911NF-10-C-0084

Measuring Dyadic Interactions

The Charles Stark Draper Laboratory is conducting the Measuring Dyadic Interactions seedling to investigate novel approaches for analyzing physiological data collected during near-real world, face-to-face dyadic interactions. This research builds on a body of research that suggests there are indicators of behavioral and even physiological synchrony that emerge during such interactions and can provide clues to the quality of the exchange and eventual outcomes. Application of this research remains limited because it has primarily used subjective measures of behavior, relied on methods that have low temporal precision, and often treats each member of a dyad as an individual, rather than analyzing the interaction as a pair. Draper will seek to advance this research by applying linear and non-linear methods to analyze physiological and non-verbal data on a millisecond time frame (such as changes in the heart rate of the interviewer and interviewee relative to one another) as each person responds to and interacts with their dyadic partner.

Performer: Charles Stark Draper Laboratory, Inc.

Resulting Research: FA8659-12-C-7210

Meme Epidemiology

Develop a framework for understanding how ideas, particularly extremist ideas, are spread throughout a population.  The framework will be based on epidemiological methods used in understanding how infectious diseases are spread throughout a population.  The goal is to develop proof-of-concept software tools that can identify the characteristics of “infectious” concepts of extremism/radicalization and how they morph through Internet blogs.  The intent is to model how infectious ideas spread and how they evolve over time.  Support vector machines will be trained to recognize concepts through the use of words and expressions and then analyze additional textual information (blogs) to track how the concept is replicated.

Performer: SRI International

Resulting Research: FA8650-11-C-7119

Memristor Processors

Hewlett-Packard Laboratories (HP) has proposed to demonstrate unconventional hybrid computation engines employing memristor crossbars. These so-called dot-product-engines are estimated to enhance speed and power efficiency by orders of magnitude, for many algorithms important to the IC. These algorithms rely on matrix and vector operations, and include but are not limited to pattern recognition in image and video analysis, Fast Fourier Transform, Monte Carlo sampling and optimization, polynomial convolution, streaming data classification, and matrix rank calculation.
The goal of this IARPA seedling is to produce specialized, high-power performance analog and digital hardware for performing real vector-matrix multiplication. As is usually the case with analog devices, precision is a key challenge. This effort does not intend to produce devices that perform high-precision computation. Hence the utility of proposed hardware is limited to algorithms where exactness of the computation is not the primary concern.

Performer: Hewlett Packard

Resulting Research: 2014-14080800008

Meta-learning Strategies

Dartmouth College is exploring the potential of automated selection and parameterization of machine learning techniques to dramatically improve system performance across diverse problem areas. If successful, the research could lay the groundwork for a larger program to automate the management of machine learning algorithms, resulting in dramatic reductions in the need for human experts when working with large, diverse data sets.

Performer: Dartmouth College

Resulting Research: N10PC20221

Photon-Photon Gate

The Photon-Photon Gate seedling researches high speed, optical switching as one approaches the quantum-mechanical limit. In principle, a photon-photon gate offers the ability to switch a data bit at high operating speeds with minimal expended energy. The University of California, San Diego research team is seeking to create an optical parametric mixer that exhibits near-isotropic photon gate response. This mixer will be constructed by measuring the fluctuations along the length of an optical fiber with sub-nanometer accuracy and using this information to synthesize a waveguide with near-ideal response. The seedling project culminates with the demonstration of an all-optical gate that is capable of single-bit switching with fewer than 20 photons at a rate of 100Gbps.

Performer: University of California, San Diego

Resulting Research: 2013-13070200004

Reading Between the Lines

Lymba Corporation is exploring the automatic identification and characterization of emerging events as reported in textual information feeds, such as news reports. Complex events of the sort of interest to the IC are made up of multiple component events and contextual influences which over time can converge into an action or event of interest. By identifying component events and tracking them over time, event structures can emerge, supporting indicators and warnings to analysts as an input to their task prioritization process.

Performer: Lymba Corporation

Resulting Research: W911NF-10-C-0101

Reliable Inference with Missing, Masked, Malfunctioning, or Malicious Sensors

Reliable Inference with Missing, Masked, Malfunctioning or Malicious Sensors (RIM4S) is developing a unified approach to obtain reliable inferences from sensor networks when data reported may have missing, masked, or inaccurate data. These errors may be unintentional (e.g., malfunctioning sensors) or intentional (e.g., maliciously tampered with, or deleted or obscured for privacy preservation). RIM4S represents multiple error types using a probabilistic graphical model, and then uses belief propagation to efficiently evaluate candidate hypotheses. Current methods to compensate for these types of errors are to threshold and reject data or sensors believed to be inaccurate or unreliable. The proposed approach promises to be able to use degraded information, enabling inference with fewer sensors or measurements.

Performer: University of California, Los Angeles

Resulting Research: N66001-12-C-2014

Sensors for Predicting Enhanced Language Learning

The Intelligence Community has a growing need for individuals across all career paths that have advanced foreign language skills. While the IC invests heavily in training advanced language skills, research has demonstrated that language proficiency in adult language learners often does not achieve advanced levels of proficiency and even with the same training proficiency will vary greatly across different adult students. Recent research suggests that general cognitive abilities, brain structure and function may predict language learning outcomes; however, this has never been tested in healthy adults.
The Sensors for Predicting Enhanced Language Learning seedling will seek to build upon previous findings to determine if there are cognitive and/or neurophysiological predictors of individuals who are more likely to be successful in learning and retaining a tonal-based novel language. The Massachusetts Institute of Technology research team will assess if there are pre-training features that predict the level of language acquisition throughout the language training, directly following completion, and at a 3 month follow-up.

Performer: Massachusetts Institute of Technology

Resulting Research: FA8650-12-C-7265

Silicon Microresonator-Based Frequency Combs for High-Performance Spectroscopic Devices in the Mid-Infrared

The goal of the Silicon Microresonator-Based Frequency Combs for High-Performance Spectroscopic Devices in the Mid-Infrared (SMFC) seedling is to investigate and develop a compact, integrated mid-infrared (MIR) frequency comb source in the 4-8 micron spectral region based on MIR parametric comb generation in microresonators. A frequency comb is a coherent light source whose spectrum consists of a series of discrete, equally spaced elements. The SMFC seedling is investigating air-clad, etchless silicon microresonators for comb generation in the MIR. It is also developing fabrication techniques to reduce coupling and propagation losses and optimizing the microresonator dispersion to enable generation of ultra-broadband combs. Potential applications enabled by the SMFC technology include high-resolution spectroscopy, high-speed detection and ultrashort pulse generation.

Performer: Cornell University

Resulting Research: FA8650-15-C-9100

Smart Devices Technical Study

The study will report on: forecasts of Smart Device technologies, characterization and overview of technical domains associated with networks of Smart Devices, current state-of-the-art, and intended operational use by developers and end-users, and opportunities and challenges for the IC.

Performer: Mitre Corporation

Resulting Research: W911NF-13-C-0085

Spectral Holographic Optical Processing

The S2 Corporation technology employs spatial-spectral holograms in cryogenically cooled crystals, and provides an all-optical means of implementing a cross-correlation algorithm to search for patterns of interest in high-speed (> 100 GHz) optical data.
Efforts in the first year served to identify limitations of single, linear hologram templates working with serial streaming data. This led to the proposal of angular multiplexing to massively scale the technology to search for millions of sub-patterns of interest on >100 GHz data. By redesigning the S2 system to employ spatial + angular multiplexing, together with novel template superposition, it is anticipated the number of programmable sub-patterns can be increased upwards of 1000-fold with 10-100x improvement in power usage per physical channel. By month 15, S2 demonstrated multiplexing of streaming data across 3 angles, with substantial reduction of noise floor and related false positive and negative rates.

Performer: S2 Corporation

Resulting Research: 2014-14020500006

Study: High-Speed Massive Dynamic Graph Analysis

Three studies: Very Large Graphs (VLG2), VLG2, and VLG-Hardware.  VLG1 effort focused on a proof-of-concept demonstration implementing time-dependent graph theory and methods for analyzing very large scale graphs that include temporal dependencies.  VLG2 extended the research by studying the impact of uncertainty mechanisms in data on anomalous subgraph detection and to understand and characterize uncertainty in real datasets.  VLG-H investigated the advanced hardware system components for processing massive network data two to four orders of magnitude beyond existing capabilities.  Specific technical improvements to be explored are data dependent optimization, a high bandwidth communication network, cache-less memory, specialized processing for sparse matrices, low-power circuitry, a sparse array instruction set, and semantic processing.

Performer: MIT-Lincoln Laboratory

Resulting Research: FA8721-05-C-002

Study: Neurally Integrated Imagery Databases

The goal is to develop algorithms for decoding patterns of brain activity (brain signals) that can then query large databases of labeled/indexed images.  The idea is to develop a system whereby a cooperative subject would be placed in a non-invasive device (e.g., fMRI) and simply think about an image.  The system would then reconstruct the image, identify appropriate attributes, and then search large image databases for similar images.  The deliverables would be reports, analyses, and research data (state-of-the-art, program formulation, experimental design, maybe some experimental results).  Intelligence analysts would benefit from simply thinking of an image and having a system retrieve similar images rather than entering in image attributes.  There would be a significant reduction in image search and retrieval times.

Performer: Johns Hopkins University Applied Physics Laboratory

Resulting Research: N00024-03-D-6606

Text-Driven Forecasting

Carnegie Mellon University is exploring the potential of purely statistical models that estimate the future influence of scientific publications from the textual content of those publications. If successful, the research could lay the groundwork for a larger program to generate predictive models from a wide range of document types (e.g., news articles, speeches, financial reports).

Performer: Carnegie Mellon University

Resulting Research: N10PC20222