Being grateful & giving a conference

Being grateful & giving a conference
The lecture I give at HelioSPIR on Deep Learning for Near-Infrared - https://www.heliospir.net/

This post is going to deviate a bit from previous topics. I am happy to have my PhD project, as it is truly a journey, and I am grateful to be paid to learn, explore, and contribute to the advancement of knowledge. It also allows me to fulfill some dreams, such as organizing a conference, which is the topic of this post. In June 2023, currently in a few weeks, I will give a small conference in the south of France at "HelioSPIR" about my research and more particularly on "Deep Learning for NIR spectroscopy: overview & thoughts". If you're curious, here is the abstract.


Keywords: deep learning, chemometrics, near-infrared spectroscopy, machine learning


Summary:

Near-Infrared spectroscopy (NIR) combined with chemometrics has been used for years in various fields to predict properties, product concentration, classify samples, etc. At IFPEN, NIR and chemometrics are daily used since more than 30 years, both in the laboratory and on-line on pilot plant units, to predict properties of interest of petroleum cuts, and more recently, for biomass characterization and plastic recycling processes. A major task of this approach is to determine a consensus on which signal processing method applied to the spectra and which optimal settings to use for the chosen method (regression, classification, etc.). The frequent practice is a time-consuming trial-and-error experimentation that must be reiterated when changing the spectrometer, for a change in acquisition conditions, to implement in real-time for process monitoring, etc.

Deep Learning approach has a different philosophy than chemometrics since it is interested in representation learning, which builds internal abstract representation to perform its tasks. Deep learning designs a model automatically extracting knowledge from data and simplify or remove human crafted operations on datasets. Additionally, this sub-field is able not only to extract linear information from data, like traditional chemometrics learning algorithm, but also non-linearity, which helps the model extract more complex information from data. Based on previous experience, especially recently published studies that demonstrate deep networks can learn critical patterns from raw data [1], it is possible to consider Deep Learning as a challenger for Chemometrics traditional practices to simplify its processes.

The current state-of-the-art of Deep Learning applied to NIR analysis is in its infancy compared to computer vision and natural language processing. It is composed of multiple directions, from Artificial Neural Network (ANN) to Convolutional Neural Network (CNN), and other architectures such as Recurrent Neural Network (RNN) like Long Short-Term Memory (LSTM) network, or generative network like Autoencoders (AE), Transformers, and Generative Adversarial Networks (GAN).

ANNs are the simplest neural networks and have been used on NIR data mostly between 2008 and 2011 [2,3]. Their overall performance were interesting, but the absence of interpretability of these networks indicates a potential drawback. Furthermore, their difficulty of interpretability leads to a limited applicability in tasks, not only in research but also on commercial products. In the data-science community, it is well known that the strong usage of other networks such as RNNs, CNNs and more recently transformers, tends to lead the way and indicates a low trust in ANNs. Also, considering the latter do not have a clear gain in performance over Partial Least Squares, yet add a complexity in relation to tweaking, the usage of these networks has decreased over time.

Currently, the most popular networks on NIR analysis are the CNN based architectures, which is an evolution of ANN with reduced risks of overfitting. These deep networks rely on convolution operations that, when stacked in multiple layers stacks, are able to extract meaningful neighborhood features from the input data. The CNNs architectures used in NIR analysis can be associated in multiple philosophies or groups. A first group, close to traditional chemometrics, uses digital signal processing then feeds a shallow CNN. A second group, close to modern deep convolutional neural networks, uses deep architecture to delegate the manual tasks of variables selection and digital signal processing. Many studies show these networks perform well [4–6], yet these diverse workflows with various philosophies lead to the following question: how to use deep learning for NIR analysis, which regularly involves to use small datasets. Furthermore, which architectures and learning strategy should we explore to satisfy this constraint?

The literature shown beyond CNNs, diverse architectures can manage different tasks in a more efficient and effective manner, like RNNs and AEs for qualitative/quantitative analysis, and GANs for simulating spectra. These networks and proposed usage are following:

  • LSTMs are networks that excel in handling sequential and time-series data. They use memory cells and gating mechanisms to selectively retain or discard information at each step of the sequence, this approach can extract long-term dependencies present in an input spectrum. This approach has shown great performance for food industry on quantitative analysis applied to manure [7].
  • AEs are networks that build new representation of data with less noise. They learn to build a compressed latent version of a spectrum, then to rebuild a decompressed spectrum without noise. This approach has shown impressive performance for chemical quantitative analysis applied to aging of product with NIR analysis [8].
  • GANs are a combination of two networks to generate new representation of data, especially for improving a dataset with new samples. A first network learns to generate a representation of data by adding noise to an existing dataset, then another network learns to discriminate by deciding if the representation is good enough or to discard it and try differently. This approach has shown success for augmenting a small dataset by generating fake samples to help training PLS models and improving their robustness [9].

Considering GANs can be useful to enhance a given dataset and the networks interested in qualitative and quantitative analysis can have different types of uses and thus be combined to extract different characteristics present within the data, it is possible to consider the following hypothesis. The CNN, AE and RNN networks, each of which is interested in extracting different types of knowledge within the data, can be combined to understand neighborhood behaviors, minimizing noise within an internal representation as well as finding the long-term relationships present within the spectra, coupled with larger dataset thanks to the help of GANs. Obviously, combining these architectures is a complex and ambitious task, but it is necessary beforehand to better understand them separately. Lastly, other usage could be explored and will lead to a better understanding of how to use these technologies to better serve our domain.

Based on the overview provided, many topics needs to be both explored and discussed. Deep learning can certainly bring a lot to NIR data analysis. Further research is necessary to explore these topics. For its improvement of performance and robustness, compared to PLS, but how much and in which context is it necessary to apply both digital signal processing and variable selection? How to combine these multiple approaches to solve more complex tasks? How to use others architecture and tweak their associate hyperparameters to solve efficiently NIR issues? How does model interpretability work for such model, we know it is possible for images and natural language but what is the state of knowledge for vibrational spectral analysis?


Bibliography

[1] Yang J., Xu J., Zhang X., Wu C., Lin T., Ying Y. Deep learning for vibrational spectral analysis: Recent progress and a practical guide, Analytica chimica acta, 2019, 1081, 6-17.
DOI: 10.1016/j.aca.2019.06.012.

[2] Balabin R.M., Safieva R.Z., Lomakina E.I. Wavelet neural network (WNN) approach for calibration model building based on gasoline near infrared (NIR) spectra, Chemometrics and Intelligent Laboratory Systems, 2008, 93, 1, 58-62. DOI: 10.1016/j.chemolab.2008.04.003.

[3] Balabin R.M., Lomakina E.I., Safieva R.Z. Neural network (ANN) approach to biodiesel analysis: Analysis of biodiesel density, kinematic viscosity, methanol and water contents using near infrared (NIR) spectroscopy, Fuel, 2011, 90, 5, 2007-2015. DOI: 10.1016/j.fuel.2010.11.038.

[4] Mishra P., Passos D. A synergistic use of chemometrics and deep learning improved the predictive performance of near-infrared spectroscopy models for dry matter prediction in mango fruit, Chemometrics and Intelligent Laboratory Systems, 2021, 212, 104287.
DOI: 10.1016/j.chemolab.2021.104287.

[5] Zhang X., Lin T., Xu J., Luo X., Ying Y. DeepSpectra: An end-to-end deep learning approach for quantitative spectral analysis, Analytica chimica acta, 2019, 1058, 48-57.
DOI: 10.1016/j.aca.2019.01.002.

[6] Acquarelli J., van Laarhoven T., Gerretzen J., Tran T.N., Buydens L.M. C., Marchiori E. Convolutional neural networks for vibrational spectroscopic data analysis, Analytica chimica acta, 2017, 954, 22-31. DOI: 10.1016/j.aca.2016.12.010.

[7] Tan A., Wang Y., Zhao Y., Wang B., Li X., Wang A.X. Near infrared spectroscopy quantification based on Bi-LSTM and transfer learning for new scenarios, Spectrochimica acta. Part A, Molecular and biomolecular spectroscopy, 2022, 283, 121759. DOI: 10.1016/j.saa.2022.121759.

[8] Grossutti M., D'Amico J., Quintal J., MacFarlane H., Quirk A., Dutcher J.R. Deep Learning and Infrared Spectroscopy: Representation Learning with a β-Variational Autoencoder, The journal of physical chemistry letters, 2022, 13, 25, 5787-5793. DOI: 10.1021/acs.jpclett.2c01328.

[9] He K., Liu J., Li Z. Application of GAN for prediction of Gasoline Properties, 2020.