GTC 2017 Can an Artificial-Intelligence Win a Nobel Prize?
Can an Artificial-Intelligence Win a Nobel Prize? Paul Michael Wigley Hush Carlos Cairon Patrick Perumbil Andre Anton van John Ian Nick Joe Kyle Mahassen Gordon Kuhn Quinlivan Everitt Manju Luiten den Hengel Bastian Petersen Robins Hope Hardman Sooriyabandara McDonald
CAN AN ARTIFICIAL- INTELLIGENCE WIN A NOBEL PRIZE?
CAN AN ARTIFICIAL- INTELLIGENCE WIN A NOBEL PRIZE? NO
CAN MACHINE LEARNING BE USED TO OPTIMIZE AN ULTRA- COLD ATOM EXPERIMENT AND REDISCOVER A RESULT THAT PREVIOUSLY WON A NOBEL PRIZE?
CAN MACHINE LEARNING BE USED TO OPTIMIZE AN ULTRA- COLD ATOM EXPERIMENT AND REDISCOVER A RESULT THAT PREVIOUSLY WON A NOBEL PRIZE? PRETTY MUCH
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC P B Wigley et al. Scientific Reports 6 25890 (2016)
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC Computer Control Physics Science P B Wigley et al. Scientific Reports 6 25890 (2016)
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC
Can an Artificial-Intelligence Win a Nobel Prize? What is a Bose Einstein Condensate (BEC)? Absolute Zero
Can an Artificial-Intelligence Win a Nobel Prize? What is a Bose Einstein Condensate (BEC)? kK 293K K 2.73K mK μ K 1.9K nK 100 nK Absolute Zero
Can an Artificial-Intelligence Win a Nobel Prize? Nobel Prize 2001 ▸ BEC proposed in 1924 ▸ BEC created in 1995 ▸ Nobel prize awarded in 2001 Albert Satyendra Einstein Nath Bose Eric Wolfgang Carl Cornell Ketterle Wieman
Can an Artificial-Intelligence Win a Nobel Prize? BEC Precision measurement with a BEC Source ▸ Atoms are sensitive to gravity and magnetic fields. ▸ Geoscience. Phase ▸ BECs are a coherent, narrow Change linewidth source for atomic interferometers. S S Szigeti et al. NJP 14 023009 Measure
Can an Artificial-Intelligence Win a Nobel Prize? Precision measurement with a BEC ▸ Atoms are sensitive to gravity and magnetic fields. ▸ Gravitation precision: 10 -9 Δ g/g ▸ Magnetic field gradient precision: 8 pT/m ▸ First interferometer to measure both. K S Hardman et al. Phys. Rev. Lett. 117 , 138501 (2016)
Can an Artificial-Intelligence Win a Nobel Prize? Evaporative cooling to create a BEC V ρ ρ ρ E E E
Can an Artificial-Intelligence Win a Nobel Prize? Evaporation ramps V(t) (b) X V C(X ) t ▸ Ergodic dynamics, V(t) two-body s-wave interactions, no other loss rates, => Exponential ramps optimal. t
Can an Artificial-Intelligence Win a Nobel Prize? Evaporation ramps V(t) (b) X V C(X ) t ▸ Ergodic dynamics, V(t) two-body s-wave interactions, ? no other loss rates, => Exponential ramps optimal. t
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC
Can an Artificial-Intelligence Win a Nobel Prize? Evaporation as an optimization problem ▸ We can parametrize the ramps:
Can an Artificial-Intelligence Win a Nobel Prize? Evaporation as an optimization problem ▸ We can parametrize the ramps: ▸ 3 ramps, common = 16 parameters
Can an Artificial-Intelligence Win a Nobel Prize? Evaporation as an optimization problem ▸ Condensate number difficult with few measurements Condensed Thermal state state ▸ Use width of image above a threshold. ▸ Cost => . Uncertainty from 2 measurements
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation How do we pick Algorithm Optimization of a BEC what X to test next?
Can an Artificial-Intelligence Win a Nobel Prize? Previous automated optimization experiments ▸ Quantum chemistry ▸ R S Judson et al. PRL 68 , 1500–1503 (1992) ▸ Cold ion quantum computing ▸ Kelly et al. PRL 112 , 240504 (2014) ▸ Cold atoms ▸ I Geisel et al. APL 102 , 214105 (2013)
Can an Artificial-Intelligence Win a Nobel Prize? Previous used automated optimization algorithms ▸ Brute force search X 1 C(X) ▸ 16 parameters, to 10% each experiment 1 min total time ~ 10 17 s ▸ Nelder-Mead ▸ Caught in local minima ▸ Genetic ▸ Chooses new points randomly X 2 ▸ What’s missing?
Can an Artificial-Intelligence Win a Nobel Prize? Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC
Can an Artificial-Intelligence Win a Nobel Prize Machine learning: The problem ▸ Regression
Can an Artificial-Intelligence Win a Nobel Prize Machine learning: Gaussian process fit ▸ Kriging: Geoscience (a) ▸ Assumes data is samples from C(X ) a set of gaussian process. ▸ Produces estimate of the mean and uncertainty. X Correlation length ▸ Requires a set of hyperparameters: correlation Short length for each dimension. Medium ▸ Hyperparameters fit from data. Long
Can an Artificial-Intelligence Win a Nobel Prize Machine learning: Strategy
Can an Artificial-Intelligence Win a Nobel Prize Machine learning: Strategy (A) (B) (C) ▸ Where would you do the next experiment?
Can an Artificial-Intelligence Win a Nobel Prize Choosing the next point to test ▸ Minimize: ▸ When b=0 learner acts like a “scientist” ▸ When b=1 learner acts like an “engineer” ▸ We swept b between 0 to 1. ▸ Used randomized gradient solver to find minimum.
Can an Artificial-Intelligence Win a Nobel Prize Overview (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t Machine Learning Automated Evaporation Algorithm Optimization of a BEC
Can an Artificial-Intelligence Win a Nobel Prize Results I Common Machine Nelder-Mead 5 x 10 5 atoms Training Learning 133 experiments Data 10 experiments
Can an Artificial-Intelligence Win a Nobel Prize Using the model ▸ Machine learning algorithms also produces a model ▸ Correlation lengths! ▸ For small data sets we found: ▸ Learner identified important (short correlation length) parameters correctly. ▸ Learner did not consistently identify unimportant (long correlation length).
Can an Artificial-Intelligence Win a Nobel Prize Maximum likelihood for hyperparameters ▸ Problem: with small data sets multiple high likelihood correlation lengths ▸ Solution: Use multiple Gaussian process and weight them. ▸ Akin to particle filters ▸ More computational time, less parameters.
Can an Artificial-Intelligence Win a Nobel Prize Test of machine learning model 6 parameters for start and end of ramps (a) (b) X X C(X ) (a) (b) V C(X ) X C(X) t + 1 useless parameter=> ▸ Can the machine learner identify the useless parameter?
Can an Artificial-Intelligence Win a Nobel Prize Results II 0.0 ● ML (6p) ▲ ■ ■ ■ ■ ■ ■ ● ● ■ ■ ■ ■ ▲ ■ ▲ ■ ▲ ▲ ■ ■ ▲ ▲ ■ ■ ▲ ■ ■ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ NM (7p) ▲ ▲ -0.5 ▲ ▲ ■ ■ ■ ▲ ▲ ▲ ▲ ■ ▲ ■ ■ ■ ▲ ■ ● ▲ ■ ▲ ● ● ● ● ● ■ ▲ ■ ML (7p) ▲ ▲ ■ ▲ ▲ ▲ ■ ● ■ ▲ ● ■ ▲ ▲ ▲ ▲ ■ ■ ▲ ■ ▲ ▲ ● ▲ ▲ ▲ -1.0 ● ■ ■ ■ ■ ▲ ▲ ● ▲ ▲ ■ ▲ ▲ ▲ ▲ ● ▲ ▲ ▲ ▲ cost ▲ ▲ ▲ ▲ ■ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ● ▲ ▲ ● ● ▲ ● ▲ -1.5 ▲ ▲ ▲ ▲ ■ ▲ ● ▲ ■ ■ ▲ ■ ■ ▲ ▲ ▲ ■ ▲ ▲ ■ ■ ■ ■ ■ ■ ■ ▲ ▲ ■ ▲ ● -2.0 ▲ ▲ ▲ ▲ ■ ▲ ▲ ● ▲ ▲ ■ ▲ ▲ ■ ■ ■ ▲ ■ ■ ▲ ▲ ▲ ▲ ■ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ▲ ● ▲ ▲ ▲ ▲ ▲ ▲ ● ▲ ● ● -2.5 80 100 0 20 40 60 120 evaluation -1.0 ▸ Machine learner correctly identified -1.2 normalised parameters the useless parameter. -1.4 -1.6 cost -1.8 Thermal ▸ Performed much better after -2.0 BEC -2.2 parameter was removed. -2.4 - 0.2 - 0.1 0.0 0.1 0.2 normalised parameters
Can an Artificial-Intelligence Win a Nobel Prize Conclusions ▸ Automated optimization of quantum experiments works even for high dimensions: 10 16 vs 30 (ML). ▸ Machine learner => Faster than NM ▸ Produces model, predicts importance of parameters. ▸ Can take into account uncertainty in measurements. ▸ Can pick points with most uncertainty or minimum cost (or something in between) ▸ Uses fast gradient methods on predicted model to find optimum.
Recommend
More recommend
