-
SpecXAI -- Spectral interpretability of Deep Learning Models
Authors:
Stefan Druc,
Peter Wooldridge,
Adarsh Krishnamurthy,
Soumik Sarkar,
Aditya Balu
Abstract:
Deep learning is becoming increasingly adopted in business and industry due to its ability to transform large quantities of data into high-performing models. These models, however, are generally regarded as black boxes, which, in spite of their performance, could prevent their use. In this context, the field of eXplainable AI attempts to develop techniques that temper the impenetrable nature of th…
▽ More
Deep learning is becoming increasingly adopted in business and industry due to its ability to transform large quantities of data into high-performing models. These models, however, are generally regarded as black boxes, which, in spite of their performance, could prevent their use. In this context, the field of eXplainable AI attempts to develop techniques that temper the impenetrable nature of the models and promote a level of understanding of their behavior. Here we present our contribution to XAI methods in the form of a framework that we term SpecXAI, which is based on the spectral characterization of the entire network. We show how this framework can be used to not only understand the network but also manipulate it into a linear interpretable symbolic representation.
△ Less
Submitted 20 February, 2023;
originally announced February 2023.
-
Concept Activation Vectors for Generating User-Defined 3D Shapes
Authors:
Stefan Druc,
Aditya Balu,
Peter Wooldridge,
Adarsh Krishnamurthy,
Soumik Sarkar
Abstract:
We explore the interpretability of 3D geometric deep learning models in the context of Computer-Aided Design (CAD). The field of parametric CAD can be limited by the difficulty of expressing high-level design concepts in terms of a few numeric parameters. In this paper, we use a deep learning architectures to encode high dimensional 3D shapes into a vectorized latent representation that can be use…
▽ More
We explore the interpretability of 3D geometric deep learning models in the context of Computer-Aided Design (CAD). The field of parametric CAD can be limited by the difficulty of expressing high-level design concepts in terms of a few numeric parameters. In this paper, we use a deep learning architectures to encode high dimensional 3D shapes into a vectorized latent representation that can be used to describe arbitrary concepts. Specifically, we train a simple auto-encoder to parameterize a dataset of complex shapes. To understand the latent encoded space, we use the idea of Concept Activation Vectors (CAV) to reinterpret the latent space in terms of user-defined concepts. This allows modification of a reference design to exhibit more or fewer characteristics of a chosen concept or group of concepts. We also test the statistical significance of the identified concepts and determine the sensitivity of a physical quantity of interest across the dataset.
△ Less
Submitted 29 April, 2022;
originally announced May 2022.
-
All-order amplitudes at any multiplicity in the multi-Regge limit
Authors:
V. Del Duca,
S. Druc,
J. M. Drummond,
C. Duhr,
F. Dulat,
R. Marzucca,
G. Papathanasiou,
B. Verbeek
Abstract:
We propose an all-loop expression for scattering amplitudes in planar N=4 super Yang-Mills theory in multi-Regge kinematics valid for all multiplicities, all helicity configurations and arbitrary logarithmic accuracy. Our expression is arrived at from comparing explicit perturbative results with general expectations from the integrable structure of a closely related collinear limit. A crucial ingr…
▽ More
We propose an all-loop expression for scattering amplitudes in planar N=4 super Yang-Mills theory in multi-Regge kinematics valid for all multiplicities, all helicity configurations and arbitrary logarithmic accuracy. Our expression is arrived at from comparing explicit perturbative results with general expectations from the integrable structure of a closely related collinear limit. A crucial ingredient of the analysis is an all-order extension for the central emission vertex that we recently computed at next-to-leading logarithmic accuracy. As an application, we use our all-order formula to prove that all amplitudes in this theory in multi-Regge kinematics are single-valued multiple polylogarithms of uniform transcendental weight.
△ Less
Submitted 14 December, 2020; v1 submitted 30 November, 2019;
originally announced December 2019.
-
Amplitudes in the Multi-Regge Limit of N=4 SYM
Authors:
Vittorio Del Duca,
Stefan Druc,
James Drummond,
Claude Duhr,
Falko Dulat,
Robin Marzucca,
Georgios Papathanasiou,
Bram Verbeek
Abstract:
A novel way of computing high-order amplitudes in the multi-Regge limit of planar maximally supersymmetric Yang-Mills theory is presented. In this framework, we are able to obtain high-loop and high-leg results by an easy operation on known amplitudes with fewer loops and lower multiplicity. This mechanism will be reviewed, along with an ensuing factorisation which allows us to determine leading l…
▽ More
A novel way of computing high-order amplitudes in the multi-Regge limit of planar maximally supersymmetric Yang-Mills theory is presented. In this framework, we are able to obtain high-loop and high-leg results by an easy operation on known amplitudes with fewer loops and lower multiplicity. This mechanism will be reviewed, along with an ensuing factorisation which allows us to determine leading logarithmic MHV results for any number of legs at a fixed loop order.
△ Less
Submitted 26 November, 2018;
originally announced November 2018.
-
Multi-Loop Amplitudes in the High-Energy Limit in N=4 SYM
Authors:
Vittorio Del Duca,
Stefan Druc,
James Drummond,
Claude Duhr,
Falko Dulat,
Robin Marzucca,
Georgios Papathanasiou,
Bram Verbeek
Abstract:
We introduce a novel way to perform high-order computations in multi-Regge-kinematics in planar N=4 supersymmetric Yang-Mills theory and generalize the existing factorization into building blocks at two loops to all loop orders. Afterwards, we will explain how this framework can be used to easily obtain higher-loop amplitudes from existing amplitudes and how to relate them to amplitudes with highe…
▽ More
We introduce a novel way to perform high-order computations in multi-Regge-kinematics in planar N=4 supersymmetric Yang-Mills theory and generalize the existing factorization into building blocks at two loops to all loop orders. Afterwards, we will explain how this framework can be used to easily obtain higher-loop amplitudes from existing amplitudes and how to relate them to amplitudes with higher number of legs.
△ Less
Submitted 20 July, 2018;
originally announced July 2018.
-
The seven-gluon amplitude in multi-Regge kinematics beyond leading logarithmic accuracy
Authors:
Vittorio Del Duca,
Stefan Druc,
James Drummond,
Claude Duhr,
Falko Dulat,
Robin Marzucca,
Georgios Papathanasiou,
Bram Verbeek
Abstract:
We present an all-loop dispersion integral, well-defined to arbitrary logarithmic accuracy, describing the multi-Regge limit of the 2->5 amplitude in planar N=4 super Yang-Mills theory. It follows from factorization, dual conformal symmetry and consistency with soft limits, and specifically holds in the region where the energies of all produced particles have been analytically continued. After pro…
▽ More
We present an all-loop dispersion integral, well-defined to arbitrary logarithmic accuracy, describing the multi-Regge limit of the 2->5 amplitude in planar N=4 super Yang-Mills theory. It follows from factorization, dual conformal symmetry and consistency with soft limits, and specifically holds in the region where the energies of all produced particles have been analytically continued. After promoting the known symbol of the 2-loop N-particle MHV amplitude in this region to a function, we specialize to N=7, and extract from it the next-to-leading order (NLO) correction to the BFKL central emission vertex, namely the building block of the dispersion integral that had not yet appeared in the well-studied six-gluon case. As an application of our results, we explicitly compute the seven-gluon amplitude at next-to-leading logarithmic accuracy through 5 loops for the MHV case, and through 3 and 4 loops for the two independent NMHV helicity configurations, respectively.
△ Less
Submitted 14 November, 2018; v1 submitted 31 January, 2018;
originally announced January 2018.
-
Multi-Regge kinematics and the moduli space of Riemann spheres with marked points
Authors:
Vittorio Del Duca,
Stefan Druc,
James Drummond,
Claude Duhr,
Falko Dulat,
Robin Marzucca,
Georgios Papathanasiou,
Bram Verbeek
Abstract:
We show that scattering amplitudes in planar N = 4 Super Yang-Mills in multi-Regge kinematics can naturally be expressed in terms of single-valued iterated integrals on the moduli space of Riemann spheres with marked points. As a consequence, scattering amplitudes in this limit can be expressed as convolutions that can easily be computed using Stokes' theorem. We apply this framework to MHV amplit…
▽ More
We show that scattering amplitudes in planar N = 4 Super Yang-Mills in multi-Regge kinematics can naturally be expressed in terms of single-valued iterated integrals on the moduli space of Riemann spheres with marked points. As a consequence, scattering amplitudes in this limit can be expressed as convolutions that can easily be computed using Stokes' theorem. We apply this framework to MHV amplitudes to leading-logarithmic accuracy (LLA), and we prove that at L loops all MHV amplitudes are determined by amplitudes with up to L + 4 external legs. We also investigate non-MHV amplitudes, and we show that they can be obtained by convoluting the MHV results with a certain helicity flip kernel. We classify all leading singularities that appear at LLA in the Regge limit for arbitrary helicity configurations and any number of external legs. Finally, we use our new framework to obtain explicit analytic results at LLA for all MHV amplitudes up to five loops and all non-MHV amplitudes with up to eight external legs and four loops.
△ Less
Submitted 28 June, 2016;
originally announced June 2016.
