Attention:The NSF Public Access Repository (PAR) system and access will be unavailable from 5:00 PM ET until 8:00 PM ET on Friday, September 11 due to maintenance. We apologize for the inconvenience.

Explore Research Products in the PAR
It may take a few hours for recently added research products to appear in PAR search results.


Title: Sum-of-Parts: Self-Attributing Neural Networks with End-to-End Learning of Feature Groups
Self-attributing neural networks (SANNs) present a potential path towards interpretable models for high-dimensional problems, but often face significant trade-offs in performance. In this work, we formally prove a lower bound on errors of per-feature SANNs, whereas group-based SANNs can achieve zero error and thus high performance. Motivated by these insights, we propose Sum-of-Parts (SOP), a framework that transforms any differentiable model into a group-based SANN, where feature groups are learned end-to-end without group supervision. SOP achieves state-of-the-art performance for SANNs on vision and language tasks, and we validate that the groups are interpretable on a range of quantitative and semantic metrics. We further validate the utility of SOP explanations in model debugging and cosmological scientific discovery.  more » « less
Award ID(s):
2442421
PAR ID:
10675368
Author(s) / Creator(s):
; ; ; ; ; ;
Publisher / Repository:
PMLR
Date Published:
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. ; ; ; ; (, SPIE Proceedings Volume PC13909)
    Dynamic photonics enables applications in LiDAR, reconfigurable optics, and computational imaging. Phase-change materials (PCMs), such as volatile VO₂ and non-volatile GeSe₃, are promising for multifunctional devices due to their large stimulus-driven optical changes. Yet, designing free-space metaphotonics across multiple material states remains challenging for conventional, library-based methods. Device performance depends on coupled electromagnetic and thermal physics, creating a high-dimensional design space. We present a deep learning–based inverse design framework that integrates coupled electro-thermal modeling and fabrication constraints directly into the optimization loop, ensuring physically accurate and manufacturable outcomes. This approach captures the coupled nature of the problem and enables performance-optimized devices beyond proof-of-concept. We validate our methodology through theory, fabrication, and characterization of a reconfigurable metaphotonic imaging application, providing an end-to-end pathway for advancing dynamic flat optics. 
    more » « less
  2. ; ; ; ; ; (, MLSys 2020)
    null (Ed.)
    Systems for ML inference are widely deployed today, but they typically optimize ML inference workloads using techniques designed for conventional data serving workloads and miss critical opportunities to leverage the statistical nature of ML. In this paper, we present WILLUMP, an optimizer for ML inference that introduces two statistically-motivated optimizations targeting ML applications whose performance bottleneck is feature computation. First, WILLUMP automatically cascades feature computation for classification queries: WILLUMP classifies most data inputs using only high-value, low-cost features selected through empirical observations of ML model performance, improving query performance by up to 5× without statistically significant accuracy loss. Second, WILLUMP accurately approximates ML top-K queries, discarding low-scoring inputs with an automatically constructed approximate model and then ranking the remainder with a more powerful model, improving query performance by up to 10× with minimal accuracy loss. WILLUMP automatically tunes these optimizations’ parameters to maximize query performance while meeting an accuracy target. Moreover, WILLUMP complements these statistical optimizations with compiler optimizations to automatically generate fast inference code for ML applications. We show that WILLUMP improves the end-to-end performance of real-world ML inference pipelines curated from major data science competitions by up to 16× without statistically significant loss of accuracy. 
    more » « less
  3. ; ; ; ; ; ; ; ; ; ; et al (, Materials Chemistry Frontiers)
    Newly developed fused-ring electron acceptors (FREAs) have proven to be an effective class of materials for extending the absorption window and boosting the efficiency of organic photovoltaics (OPVs). While numerous acceptors have been developed, there is surprisingly little structural diversity among high performance FREAs in literature. Of the high efficiency electron acceptors reported, the vast majority utilize derivatives of 2-(3-oxo-2,3-dihydroinden-1-ylidene)malononitrile (INCN) as the acceptor moiety. It has been postulated that the high electron mobility exhibited by FREA molecules with INCN end groups is a result of close π–π stacking between the neighboring planar INCN groups, forming an effective charge transport pathway between molecules. To explore this as a design rationale for electron acceptors, we synthesized a new fused-ring electron acceptor, IDTCF, which has methyl substituents out of plane to the conjugated acceptor backbone. These methyl groups hinder packing and expand the π–π stacking distance by ∼1 Å, but have little impact on the optical or electrochemical properties of the individual FREA molecule. The extra steric hindrance from the out of plane methyl substituents restricts packing and results in large amounts of geminate recombination, thus degrading the device performance. Our results show that intermolecular interactions (especially π–π stacking between end groups) play a crucial role in performance of FREAs. We demonstrated that the planarity of the acceptor unit is of paramount importance as even minor deviations in end group distance are enough to disrupt crystallinity and cripple device performance. 
    more » « less
  4. ; ; ; ; (, USENIX Symposium on Networked Systems Design and Implementation (NSDI))
    Machine learning has shown tremendous potential for improving the capabilities of network traffic analysis applications, often outperforming simpler rule-based heuristics. However, ML-based solutions remain difficult to deploy in practice. Many existing approaches only optimize the predictive performance of their models, overlooking the practical challenges of running them against network traffic in real time. This is especially problematic in the domain of traffic analysis, where the efficiency of the serving pipeline is a critical factor in determining the usability of a model. In this work, we introduce CATO, a framework that addresses this problem by jointly optimizing the predictive performance and the associated systems costs of the serving pipeline. CATO leverages recent advances in multi-objective Bayesian optimization to efficiently identify Pareto-optimal configurations, and automatically compiles end-to-end optimized serving pipelines that can be deployed in real networks. Our evaluations show that compared to popular feature optimization techniques, CATO can provide up to 3600× lower inference latency and 3.7× higher zero-loss throughput while simultaneously achieving better model performance. 
    more » « less
  5. ; (, PLOS Computational Biology)
    Mac_Gabhann, Feilim (Ed.)
    Deep learning-based methods for identifying and tracking cells within microscopy images have revolutionized the speed and throughput of data analysis. These methods for analyzing biological and medical data have capitalized on advances from the broader computer vision field. However, cell tracking can present unique challenges, with frequent cell division events and the need to track many objects with similar visual appearances complicating analysis. Existing architectures developed for cell tracking based on convolutional neural networks (CNNs) have tended to fall short in managing the spatial and global contextual dependencies that are crucial for tracking cells. To overcome these limitations, we introduce Cell-TRACTR (Transformer withAttention forCellTracking andRecognition), a novel deep learning model that uses a transformer-based architecture. Cell-TRACTR operates in an end-to-end manner, simultaneously segmenting and tracking cells without the need for post-processing. Alongside this model, we introduce the Cell-HOTA metric, an extension of the Higher Order Tracking Accuracy (HOTA) metric that we adapted to assess cell division. Cell-HOTA differs from standard cell tracking metrics by offering a balanced and easily interpretable assessment of detection, association, and division accuracy. We test our Cell-TRACTR model on datasets of bacteria growing within a defined microfluidic geometry and mammalian cells growing freely in two dimensions. Our results demonstrate that Cell-TRACTR exhibits strong performance in tracking and division accuracy compared to state-of-the-art algorithms, while also meeting traditional benchmarks in detection accuracy. This work establishes a new framework for employing transformer-based models in cell segmentation and tracking. 
    more » « less
  • Citation Formats
  • MLA
  • APA
  • Chicago
  • BibTeX
  • Text Encoded Conversion
  • XML
  • Save / Share this Record
  • Send to Email