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: SymbolNet: neural symbolic regression with adaptive dynamic pruning for compression
Abstract Compact symbolic expressions have been shown to be more efficient than neural network (NN) models in terms of resource consumption and inference speed when implemented on custom hardware such as field-programmable gate arrays (FPGAs), while maintaining comparable accuracy (Tsoiet al2024EPJ Web Conf.29509036). These capabilities are highly valuable in environments with stringent computational resource constraints, such as high-energy physics experiments at the CERN Large Hadron Collider. However, finding compact expressions for high-dimensional datasets remains challenging due to the inherent limitations of genetic programming (GP), the search algorithm of most symbolic regression (SR) methods. Contrary to GP, the NN approach to SR offers scalability to high-dimensional inputs and leverages gradient methods for faster equation searching. Common ways of constraining expression complexity often involve multistage pruning with fine-tuning, which can result in significant performance loss. In this work, we propose S y m b o l N e t , a NN approach to SR specifically designed as a model compression technique, aimed at enabling low-latency inference for high-dimensional inputs on custom hardware such as FPGAs. This framework allows dynamic pruning of model weights, input features, and mathematical operators in a single training process, where both training loss and expression complexity are optimized simultaneously. We introduce a sparsity regularization term for each pruning type, which can adaptively adjust its strength, leading to convergence at a target sparsity ratio. Unlike most existing SR methods that struggle with datasets containing more than O ( 10 ) inputs, we demonstrate the effectiveness of our model on the LHC jet tagging task (16 inputs), MNIST (784 inputs), and SVHN (3072 inputs).  more » « less
Award ID(s):
2019786
PAR ID:
10588324
Author(s) / Creator(s):
; ; ;
Publisher / Repository:
IOP Publishing Ltd
Date Published:
Journal Name:
Machine Learning: Science and Technology
Volume:
6
Issue:
1
ISSN:
2632-2153
Page Range / eLocation ID:
015021
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract We generalize a magnetogram-matching Biot–Savart law (BSl) from planar to spherical geometry. For a given coronal current densityJ, this law determines the magnetic field B ˜ whose radial component vanishes at the surface. The superposition of B ˜ with a potential field defined by a given surface radial field,Br, provides the entire configuration whereBrremains unchanged by the currents. Using this approach, we (1) upgrade our regularized BSls for constructing coronal magnetic flux ropes (MFRs) and (2) propose a new method for decomposing a measured photospheric magnetic field as B = B pot + B T + B S ˜ , where the potential,Bpot, toroidal,BT, and poloidal, B S ˜ , fields are determined byBr,Jr, and the surface divergence ofB–Bpot, respectively, all derived from magnetic data. OurBTis identical to the one in the alternative Gaussian decomposition by P. W. Schuck et al., whileBpotand B S ˜ are different from their poloidal fields B P < and B P > , which arepotentialin the infinitesimal proximity to the upper and lower side of the surface, respectively. In contrast, our B S ˜ has no such constraints and, asBpotandBT, refers to thesameupper side of the surface. In spite of these differences, for a continuousJdistribution across the surface,Bpotand B S ˜ are linear combinations of B P < and B P > . We demonstrate that, similar to the Gaussian method, our decomposition allows one to identify the footprints and projected surface-location of MFRs in the solar corona, as well as the direction and connectivity of their currents. 
    more » « less
  2. Abstract We present13CO(J= 1 → 0) observations for the EDGE-CALIFA survey, which is a mapping survey of 126 nearby galaxies at a typical spatial resolution of 1.5 kpc. Using detected12CO emission as a prior, we detect13CO in 41 galaxies via integrated line flux over the entire galaxy and in 30 galaxies via integrated line intensity in resolved synthesized beams. Incorporating our CO observations and optical IFU spectroscopy, we perform a systematic comparison between the line ratio 12 / 13 I [ 12 CO ( J = 1 0 ) ] / I [ 13 CO ( J = 1 0 ) ] and the properties of the stars and ionized gas. Higher 12 / 13 values are found in interacting galaxies compared to those in noninteracting galaxies. The global 12 / 13 slightly increases with infrared colorF60/F100but appears insensitive to other host-galaxy properties such as morphology, stellar mass, or galaxy size. We also present azimuthally averaged 12 / 13 profiles for our sample up to a galactocentric radius of 0.4r25(∼6 kpc), taking into account the13CO nondetections by spectral stacking. The radial profiles of 12 / 13 are quite flat across our sample. Within galactocentric distances of 0.2r25, the azimuthally averaged 12 / 13 increases with the star formation rate. However, Spearman rank correlation tests show the azimuthally averaged 12 / 13 does not strongly correlate with any other gas or stellar properties in general, especially beyond 0.2r25from the galaxy centers. Our findings suggest that in the complex environments in galaxy disks, 12 / 13 is not a sensitive tracer for ISM properties. Dynamical disturbances, like galaxy interactions or the presence of a bar, also have an overall impact on 12 / 13 , which further complicates the interpretations of 12 / 13 variations. 
    more » « less
  3. Abstract This paper investigates uniqueness results for perturbed periodic Schrödinger operators on Z d . Specifically, we consider operators of the form H = Δ + V + v , where Δ is the discrete Laplacian, V : Z d R is a periodic potential, and v : Z d C represents a decaying impurity. We establish quantitative conditions under which the equation Δ u + V u + v u = λ u , for λ C , admits only the trivial solution u 0 . Key applications include the absence of embedded eigenvalues for operators with impurities decaying faster than any exponential function and the determination of sharp decay rates for eigenfunctions. Our findings extend previous works by providing precise decay conditions for impurities and analyzing different spectral regimes ofλ. 
    more » « less
  4. Abstract Polyatomic molecules have been identified as sensitive probes of charge-parity violating and parity violating physics beyond the Standard Model (BSM). For example, many linear triatomic molecules are both laser-coolable and have parity doublets in the ground electronic X ˜ 2 Σ + ( 010 ) state arising from the bending vibration, both features that can greatly aid BSM searches. Understanding the X ˜ 2 Σ + ( 010 ) state is a crucial prerequisite to precision measurements with linear polyatomic molecules. Here, we characterize the fundamental bending vibration of 174 YbOH using high-resolution optical spectroscopy on the nominally forbidden X ˜ 2 Σ + ( 010 ) A ˜ 2 Π 1 / 2 ( 000 ) transition at 588 nm. We assign 39 transitions originating from the lowest rotational levels of the X ˜ 2 Σ + ( 010 ) state, and accurately model the state’s structure with an effective Hamiltonian using best-fit parameters. Additionally, we perform Stark and Zeeman spectroscopy on the X ˜ 2 Σ + ( 010 ) state and fit the molecule-frame dipole moment to D m o l = 2.16 ( 1 ) Dand the effective electrong-factor to g S = 2.07 ( 2 ) . Further, we use an empirical model to explain observed anomalous line intensities in terms of interference from spin–orbit and vibronic perturbations in the excited A ˜ 2 Π 1 / 2 ( 000 ) state. Our work is an essential step toward searches for BSM physics in YbOH and other linear polyatomic molecules. 
    more » « less
  5. Abstract Entanglement is an intrinsic property of quantum mechanics and is predicted to be exhibited in the particles produced at the Large Hadron Collider. A measurement of the extent of entanglement in top quark-antiquark ( t t ¯ ) events produced in proton–proton collisions at a center-of-mass energy of 13 TeV is performed with the data recorded by the CMS experiment at the CERN LHC in 2016, and corresponding to an integrated luminosity of 36.3 fb−1. The events are selected based on the presence of two leptons with opposite charges and high transverse momentum. An entanglement-sensitive observableDis derived from the top quark spin-dependent parts of the t t ¯ production density matrix and measured in the region of the t t ¯ production threshold. Values of D < 1 / 3 are evidence of entanglement andDis observed (expected) to be 0.480 0.029 + 0.026 ( 0.467 0.029 + 0.026 ) at the parton level. With an observed significance of 5.1 standard deviations with respect to the non-entangled hypothesis, this provides observation of quantum mechanical entanglement within t t ¯ pairs in this phase space. This measurement provides a new probe of quantum mechanics at the highest energies ever produced. 
    more » « less