A5086188300- Stochastic Gradient Optimization Techniques
- Advanced Thermodynamics and Statistical Mechanics
- Neural Networks and Applications
- Markov Chains and Monte Carlo Methods
- Statistical Mechanics and Entropy
- Machine Learning and Data Classification
- Theoretical and Computational Physics
- Machine Learning and Algorithms
- Thermal Radiation and Cooling Technologies
- Gaussian Processes and Bayesian Inference
- stochastic dynamics and bifurcation
- Quantum many-body systems
- Receptor Mechanisms and Signaling
- Quantum Information and Cryptography
- Protein Structure and Dynamics
- Adversarial Robustness in Machine Learning
- Quantum and electron transport phenomena
- Quantum Computing Algorithms and Architecture
- Neural dynamics and brain function
- Computational Drug Discovery Methods
- Quantum Mechanics and Applications
Xihua University
2024
The University of Tokyo
2020-2021
The thermodynamic uncertainty relation (TUR) describes a trade-off between nonequilibrium currents and entropy production serves as fundamental principle of thermodynamics. However, currently known TURs presuppose either specific initial states or an infinite-time average, which severely limits the range applicability. Here we derive finite-time TUR valid for arbitrary from Cram\'er-Rao inequality. We find that variance accumulated current is bounded by instantaneous at final time, suggests...
The thermodynamics and kinetics of a nonequilibrium classical system fundamentally constrain the precision an observable regarding celebrated thermodynamic uncertainty relation (TUR) kinetic (KUR). They have been extended to open quantum systems obeying Lindblad master equation where unique effects are identified. Recently, new set principles that further incorporate response external perturbation discovered, named TUR KUR (R-KUR). In this work, we generalize R-KUR using Cram\'er-Rao bound...
Entering the era of post-quantum supremacy has given one ability to precisely control noisy intermediate-scale quantum (NISQ) processors with multiqubits and extract valuable many-body correlation resources for many distinct applications. We here construct thermalized states on a 62-qubit superconducting processor use them demonstrate principle Maxwell's demon. further direct effect caused by demon charging process battery (QB). depicted nonequilibrium transportation in our QB through...
The noise in stochastic gradient descent (SGD), caused by minibatch sampling, is poorly understood despite its practical importance deep learning. This work presents the first systematic study of SGD and fluctuations close to a local minimum. We analyze linear regression detail then derive general formula for approximating different types minima. For application, our results (1) provide insight into stability training neural network, (2) suggest that large learning rate can help...
In the vanishing learning rate regime, stochastic gradient descent (SGD) is now relatively well understood. this work, we propose to study basic properties of SGD and its variants in non-vanishing regime. The focus on deriving exactly solvable results discussing their implications. main contributions work are derive stationary distribution for discrete-time a quadratic loss function with without momentum; particular, one implication our result that fluctuation caused by dynamics takes...
The exploration of far-from-equilibrium systems has been at the forefront nonequilibrium thermodynamics, with a particular focus on understanding fluctuations and response thermodynamic to external perturbations. In this study, we introduce universal kinetic uncertainty relation, which provides fundamental trade-off between precision for generic observables dynamical activity in Markovian systems. We demonstrate practical applicability tightness derived bound through illustrative examples....
Stochastic gradient descent (SGD) undergoes complicated multiplicative noise for the mean-square loss. We use this property of SGD to derive a stochastic differential equation (SDE) with simpler additive by performing random time change. Using formalism, we show that log loss barrier $\Delta\log L=\log[L(\theta^s)/L(\theta^*)]$ between local minimum $\theta^*$ and saddle $\theta^s$ determines escape rate from minimum, contrary previous results borrowing physics linear $\Delta...
Maxwell's demon can be utilized to construct quantum information engines. While most of the existing engines harness thermal fluctuations, that utilize fluctuations have recently been discussed. We propose a new type genuinely engine harnesses achieve cumulative storage useful work and unidirectional transport particle. Our scheme does not require thermalization, which eliminates ambiguity in evaluating power velocity our proposed contrast other find tradeoff relationship between maximum...