skip to main content
US FlagAn official website of the United States government
dot gov icon
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
https lock icon
Secure .gov websites use HTTPS
A lock ( lock ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.


Title: Towards Minimal Targeted Updates of Language Models with Targeted Negative Training
Award ID(s):
2145542
PAR ID:
10519172
Author(s) / Creator(s):
; ;
Publisher / Repository:
Transactions in Machine Learning Research
Date Published:
Journal Name:
Transactions on machine learning research
ISSN:
2835-8856
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Cell type-specific interfaces within living animals will be invaluable for achieving communication with identifiable cells over the long term, enabling applications across many scientific and medical fields. However, biological tissues exhibit complex and dynamic organization properties that pose serious challenges for chronic cell-specific interfacing. A new technology, combining chemistry and molecular biology, has emerged to address this challenge: genetically targeted chemical assembly (GTCA), in which specific cells are genetically programmed (even in wild-type or non-transgenic animals, including mammals) to chemically construct non-biological structures. Here, we discuss recent progress in genetically targeted construction of materials and outline opportunities that may expand the GTCA toolbox, including specific chemical processes involving novel monomers, catalysts and reaction regimes both de cellula (from the cell) and ad cellula (towards the cell); different GTCA-compatible reaction conditions with a focus on light-based patterning; and potential applications of GTCA in research and clinical settings. 
    more » « less
  2. Background: Millions of people are affected yearly by “runner’s knee” and osteoarthritis, which is thought to be related to impact force. Millions are also affected by chronic falling, who are usually both difficult to identify and train. While at first glance, these topics seem to be entirely disconnected, there appears to be a need for a device that would address both issues. This paper proposes and investigates the use of the Variable Stiffness Treadmill (VST) as a targeted training device for the different populations described above. Materials and Methods: The VST is the authors’ unique robotic split-belt treadmill that can reduce the vertical ground stiffness of the left belt, while the right belt remains rigid. In this work, heart rate and energy expenditure are measured for healthy subjects in the challenging asymmetric environment created by the VST and compared to a traditional treadmill setting. Results: This study shows that this asymmetric environment results in an increase in heart rate and energy expenditure, an increase in activity in the muscles about the hip and knee, and a decrease in impact force at heel strike. Conclusions: Compliant environments, like those created on the VST, may be a beneficial tool as they can: reduce high-impact forces during running and walking, significantly engage the muscles surrounding the hip and knee allowing for targeted training and rehabilitation, and assist in identifying and training high fall-risk individuals. 
    more » « less
  3. Lossy trapdoor functions, introduced by Peikert and Waters (STOC ’08), can be initialized in one of two indistinguishable modes: in injective mode, the function preserves all information about its input, and can be efficiently inverted given a trapdoor, while in lossy mode, the function loses some information about its input. Such functions have found countless applications in cryptography, and can be constructed from a variety of Cryptomania assumptions. In this work, we introduce targeted lossy functions (TLFs), which relax lossy trapdoor functions along two orthogonal dimensions. Firstly, they do not require an inversion trapdoor in injective mode. Secondly, the lossy mode of the function is initialized with some target input, and the function is only required to lose information about this particular target. The injective and lossy modes should be indistinguishable even given the target. We construct TLFs from Minicrypt assumptions, namely, injective pseudorandom generators, or even one-way functions under a natural relaxation of injectivity. We then generalize TLFs to incorporate branches, and construct all-injective-but-one and all-lossy-but-one variants. We show a wide variety of applications of targeted lossy functions. In several cases, we get the first Minicrypt constructions of primitives that were previously only known under Cryptomania assumptions. Our applications include: Pseudo-entropy functions from one-way functions. Deterministic leakage-resilient message-authentication codes and improved leakage-resilient symmetric-key encryption from one-way functions. Extractors for extractor-dependent sources from one-way functions. Selective-opening secure symmetric-key encryption from one-way functions. A new construction of CCA PKE from (exponentially secure) trapdoor functions and injective pseudorandom generators. We also discuss a fascinating connection to distributed point functions. 
    more » « less