neural-networks: a tool for gamma-spectroscopy
TRANSCRIPT
Neural-networks: What are they ?
Gabriel T. – GSD
Software 2.0
Building block for Industry 4.0
Artificial intelligence is the new electricity
Just another tool in your machine learning toolbox
Is deep learning already hitting its limitations?
Companies will spend $50 billion on artificial intelligence this year
with little to show for it
Amount of data
Per
form
ance
Neural-networks: How do they work ?
Gabriel T. – GSD
neural network
Input Output
Features: - model independent- support for high-dimensional data
Neural-networks: How do they work ?
Gabriel T. – GSD
feed-forward
Input Output
connectionsneurons
w02
w11
w01
w12
Neural-networks: How do they work ?
Gabriel T. – GSD
feed-forward
back-propagation
Input Output
connectionsneurons
w02
w11
w01
w12
Neural-networks: What are they good for ?
Gabriel T. – GSD
Material sample
Zeff
= ?
Classical approach: dual-R Neural network approach
𝐼=𝐼 0 ∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
R≈ln (Attlow)
ln (Att high)≈1𝑍 eff
NNAttlowAtthigh
Z eff
G. V. Turturica, V. Iancu, et al., EPJ Plus 135, 140 (2020)
Neural-networks: What are they good for ?
Gabriel T. – GSD
Material sample
Zeff
= ?
𝐼=𝐼 0 ∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
Simulated Experimental0
0.01
0.02
0.03
0.04
0.05
0.06
Evaluation RMSE
Dual-R ANN
24%48%
G. V. Turturica, V. Iancu, et al., EPJ Plus 135, 140 (2020)
Neural-networks: What are they good for ?
Gabriel T. – GSD
Material sample
Zeff
= ?
Classical approach: dual-R Neural network approach
𝐼=𝐼 0 ∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
R≈ln [Integral (Att low)]
ln [ Integral (Att high)]≈1𝑍 eff
NN Z eff
Atthigh
G. V. Turturica, V. Iancu, C. A. Ur, under review @ NIM A
high-energy
low-energy
high-energy
low-energy
Neural-networks: What are they good for ?
Gabriel T. – GSD
G. V. Turturica, V. Iancu, C. A. Ur, under review @ NIM A
Neural network: discrimination improvement, ½ dose reduction, ½ time reduction
Material sample
Zeff
= ?
𝐼=𝐼 0∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
high-energy
low-energy
high-energy
low-energy
Neural-networks: What are they good for ?
Gabriel T. – GSD
Classical approach: dual-R Neural network approach
R≈ln (Attlow)
ln(Atthigh)NN
Z effAttlow
AtthighG. V. Turturica, V. Iancu, et al., EPJ Plus 135, 140 (2020)
Neural-networks: What are they good for ?
Gabriel T. – GSD
G. V. Turturica, V. Iancu, et al., EPJ Plus 135, 140 (2020)
Neural-networks: What are their limits ?
Gabriel T. – GSD
Material sample
𝐼=𝐼 0∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
Experiment: 19 points
Geant4:378 points
Simulated Experimental0
0.01
0.02
0.03
0.04
0.05
0.06
Evaluation RMSE
Dual-R ANN
24%48%
Neural-networks: What are their limits ?
Gabriel T. – GSD
Material sample
𝐼=𝐼 0∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
Experiment: 19 points
Geant4:378 points
Simulated Experimental0
0.01
0.02
0.03
0.04
0.05
0.06
Evaluation RMSE
Dual-R ANN
24%48%
Pretraining was essential !
Pretraining data
Neural-networks: What are their limits ?
Gabriel T. – GSD
Material sample
𝐼=𝐼 0∙𝑒−𝜇(𝐸 ,𝑍 )𝜌 𝑙
Dual-R NN – experimental dataPretraining
NN – experimental data
Neural-networks: What are their limits ?
Gabriel T. – GSD
Ghost imaging setup:
Compressed sensing:
2000 images
[123, 401, …, 212]
- high complexity input- complex network architecture- large sample size- hard to simulate samples
Neural network:
Neural-networks: What are their limits ?
Gabriel T. – GSD
Ghost imaging setup:
Compressed sensing:
2000 images
[123, 401, …, 212]
- high complexity input- complex network architecture- large sample size- hard to simulate samples
no convergence was achieved
Neural network:
Neural-networks: What we hope they can do ?
Gabriel T. – GSD
Robust automatic spectral analysis:
Automatic geometric and isotopic object identification:
NN
isotopic: 0.73
geometric: 0.99
authentication network
unique item
Neural-networks: A tool for gamma-spectroscopy ?
Gabriel T. – GSD
Advantages Limitations
- limited performance on small sized datasets- hard to train on high complexity input
- can outperform classical algorithms on regression tasks. - relatively easy to generate accurate pretraining data using Monte Carlo simulations