in a nutshell...in a nutshell exploring the applicability of deep learning methods in mid-infrared...
TRANSCRIPT
![Page 1: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/1.jpg)
IN A NUTSHELL
EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES
Franck Albinet, Amelia Lee Zhi Yi, Petra Schmitter, Romina Torres Astorga, and Gerd Dercon
![Page 2: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/2.jpg)
Mid-Infrared spectroscopy allows for high-throughput prediction of soil properties.
![Page 3: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/3.jpg)
Partial Least Square (PLS) is the mainstream approach*.
* Including ad hoc pre-processing + features engineering such as wavelength selections.
![Page 4: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/4.jpg)
But yet fails in some ways*. How?
* E.g. in prediction of Potassium or exhibiting poor reproducibility
![Page 5: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/5.jpg)
Mid-Infrared spectroscopy data are high-dimensional.
![Page 6: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/6.jpg)
It suffers from Curse of Dimensionality a.k.a models require billions* of data
* Mid-Infrared spectra are a scarce resource.
![Page 7: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/7.jpg)
In reality, data is concentrated in a much smaller latent region.
![Page 8: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/8.jpg)
The quest is to identify this region of lower dimension containing the information.
![Page 9: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/9.jpg)
Dimensionality reduction + a priori information is the standard.
![Page 10: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/10.jpg)
PLS is exactly doing so but overly drastically*.
* therefore losing the ability to predict difficult analytes (e.g Potassium).
![Page 11: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/11.jpg)
How can Deep Learning be a part of the solution?*
* although Deep Neural Networks are notoriously data intensive!
![Page 12: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/12.jpg)
By understanding why it works so well in so many areas!
![Page 13: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/13.jpg)
Mathematicians begin to understand how*!
* Understanding deep convolutional networks, S.Mallat
![Page 14: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/14.jpg)
… and show that not everything needs to be learned!
* Understanding deep convolutional networks, S.Mallat
![Page 15: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/15.jpg)
For instance by using Wavelet Scattering Networks
[1]
* http://mathsdl-spring20.willwhitney.com/assets/documents/ScatteringTransform.pdf
![Page 16: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/16.jpg)
Which yield parsimonious* representation by capturing spectra singularities.
*drastic dimensionality reduction is achieved because the selected spectra have very few of the non-zero features.
![Page 17: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/17.jpg)
These features as inputs to DL | ML algorithms = Hybrid*
* In small data regime as dimensionality have been drastically reduced
![Page 18: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/18.jpg)
Higher prediction power?In small data regime?Higher reproducibility?Higher interpretability?
![Page 19: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/19.jpg)
Well, that’s the plan!
Our research agenda for the coming year.
![Page 20: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/20.jpg)
Context of these research activities
https://bit.ly/3d805vJ
![Page 21: IN A NUTSHELL...IN A NUTSHELL EXPLORING THE APPLICABILITY OF DEEP LEARNING METHODS IN MID-INFRARED SPECTROSCOPY FOR SOIL PREDICTIONS PROPERTIES Franck Albinet, Amelia Lee Zhi Yi, Petra](https://reader034.vdocuments.us/reader034/viewer/2022042812/5faa1a2c506fe8285330dd6e/html5/thumbnails/21.jpg)
Contact details● [email protected]
● Gerd Dercon, Head of Soil and Water Management & Crop Nutrition Laboratory, International Atomic Energy Agency