Officials of the U.S. Defense Advanced Research Projects Agency (DARPA) in Arlington, Va., released a solicitation on Monday (HR001118S0044) for the three-year Learning with Less Labels (LwLL) program.
The goal is to reduce the amount of labeled data in AI systems by six or more orders of magnitude, and reducing the amount of data necessary to adapt models to new environments to tens to hundreds of labeled examples.
In supervised machine learning, the system learns by example to recognize things like objects in images or speech, DARPA officials explain. Humans provide these examples as labeled data; enough labeled data lends itself to accurate pattern recognition.
Training accurate models, however, requires lots of labeled data. Deep neural networks have emerged as the state of the art for tasks like machine translation, speech recognition, or object recognition because of their superior accuracy. Still, deep-learning network models need many examples to achieve good performance.
The commercial world has harvested and created large sets of labeled data for training models. These data sets often are created via crowdsourcing: a cheap and efficient way to create labeled data. Unfortunately, crowdsourcing techniques often are not possible for proprietary or sensitive data. Creating data sets for these sorts of problems can result in 100x higher costs and 50x longer time to label.
To make matters worse, machine learning models are brittle because their performance can degrade severely with small changes in their operating environment. The performance of computer vision systems, for example, degrades when data comes from a new sensor, which requires additional labels after initial training.
For many problems, the labeled data required to adapt models to new environments approaches the amount required to train a new model from scratch.
To reduce the labeled data necessary to train accurate models, the DARPA LwLL program focuses on two areas: developing learning algorithms that learn and adapt efficiently; and characterizing machine learning problems to prove the limits of learning and adaptation.
The learning algorithms portion, technical areal 1, seeks to develop learning algorithms that reduce the amount of labeled data necessary to build a model from scratch by at least a factor of 106; and to adapt to new environments with hundreds of labeled examples.
It's likely that new methods will be necessary to focus on aspects of input data while reducing nuisance variation, and to use implicit or indirect supervision to exploit unlabeled data. DARPA anticipates advances in methods such as meta-learning, automated transfer learning, reinforcement learning, active learning, unsupervised or semi-supervised learning, and k-shot learning.
Proving the limits of learning and adaptation, technical area 2, focuses on limiting the amount of labeled data necessary to solve a machine learning problem. This will require methods to characterize decision difficulty and complexity of the data that helps make decisions.
Characterizations should prove limits on training and adaptation for different classes of machine learning problems, different models, and different kinds of data. DARPA seeks theories that prove tight bounds on learning in the presence of transfer and meta-transfer learning.
Companies interested should submit abstracts no later than 21 Aug. 2018, and full proposals no later than 2 Oct. 2018 to the DARPA BAA Website at https://baa.darpa.mil.
More information is online at https://www.fbo.gov/spg/ODA/DARPA/CMO/HR001118S0044/listing.html.
Ready to make a purchase? Search the Military & Aerospace Electronics Buyer's Guide for companies, new products, press releases, and videos
