Real-time learning of decay trajectory of Higgs boson using reservoir-in-reservoir architecture
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Real time learning, Higgs boson, reservoir computing, recurrent neural networks, non linear dynamic systems, machine learning, particle decay
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose reservoir-in-reservoir architecture to learn aperiodic non linear dynamic systems such as Higgs boson particle decay in real time.
Abstract: Real-time learning of the decay trajectory in Higgs bosons as they interact in the Higgs Field is the key to understanding and furthering of the mass providing mechanism and particle interaction mechanism beyond the Standard model in particle physics. We propose a novel machine learning architecture called reservoir-in-reservoir, to learn this complex high dimensional weak and electromagnetic interaction model involving a large number of arbitrary parameters whose full understanding remains elusive to physicists, making it harder to handcraft features or represent in a closed-form equation. Reservoir-in-reservoir is a reservoir computing (RC) approach, where we built a large reservoir using a pool of small reservoirs that are individually specialized to learn patterns from discrete time samples of decay trajectory without any prior knowledge. Each small reservoir consists of a paired primary and secondary reservoir of recurrently-connected neurons, known as learner and generator, respectively, with a readout connected to the head. During the training phase, we activate the learner-generator pairs within the pool. Then we excite each learners with an unit impulse and individual time windows of the incoming system. We train the internal recurrent connections and readouts using a recursive least squares-based First-Order and Reduced Control Error (FORCE) algorithm. To enhance adaptability and performance, we implement a time-varying forgetting factor optimization during training. This optimization helps control the fading and adaptation of the covariance matrix based on variations in the incoming decay trajectory and patterns. This comprehensive training strategy aims to guarantee that the entire reservoir pool evolves in harmony with the desired output dynamics. We optimize hyper-parameters such as the number of learner-generator pairs within the pool, their network sizes, batch sizes, and the number of training trials. During testing, we excite the generators in the pool, with only an unit impulse, to mimic the dynamic system. We facilitate real-time learning by re-triggering the training process involving learner-generator pairs whenever the error rate exceeds a predefined threshold. We evaluate our reservoir-in-reservoir architecture using Higgs boson decay trajectories as
detected in the Compact Muon Solenoid (CMS) detector of CERN’s Large Hadron Collider (LHC). The reservoir pool is used to model the dynamics of momentum components (and transverse momentum) as Higgs boson decays into photons and leptons (electrons and muons) with invariant masses between 120-130 GeV. Our results indicate that reservoir-in-reservoir architecture is a well suited machine
learning paradigm in learning dynamical systems such as Higgs boson decay.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6481
Loading