Spectroscopy is the study of the absorption and emission of light and other radiation by matter. Different instruments exist which use different light regions, like mid-infrared spectroscopy or near-infrared spectroscopy. These monthly meetings focus on presenting the novelties of this sector, from the utilization of novel statistical methods, to the equations to predict novel traits. In particular this group focuses on latest research in agriculture, from the analyses of soil, to food, passing through grass, and milk and animal traits predictions.

The meetings are online, on the last Tuesday of each month, from 14.00 to 15.00 GMT.

Upcoming meetings

Next Meeting:

May 30th 2pm 

Speaker: Szymon Urbas, Postdoc at Insight Centre for data analytics, Dublin.

Title: Accurate and reliable predictions from spectral data using Bayesian partial least squares with agri-food applications

Abstract:

Partial least squares (PLS) regression is a popular method for making predictions from spectral data in the agri-food sector. For example, traits such as protein content or heat stability can be predicted from high-dimensional mid-infrared spectral data of milk samples. The PLS regression approach is based on maximising the covariance between the predictor and response vectors through a latent–space decomposition. However, the method does not arise from the statistical inference of any particular probabilistic model. Thus it cannot account for parameter uncertainty, and it does not easily lend itself to modifications. Furthermore, the choice of the dimension, “Q”, of the latent space can be sensitive to the training dataset and the choice process can be computationally expensive. We introduce a Bayesian latent-variable model which emulates the desirable properties of PLS. To eschew the need to choose the latent dimension, we employ a multiplicative gamma process prior which enforces shrinkage on the latent variables. The modelling framework is very flexible and allows for modifications inspired by sparse regression approaches such as spike-and-slab or Bayesian LASSO.

The proposed Bayesian partial least squares framework is applied to three different spectral data problems from the agri-food sector, and we show how the point predictions are at least as accurate as the industry-standard PLS methods. In particular, the regularisation provided by the use of sparsity-inducing priors makes the method particularly well-suited to “small N, large P” problems. In addition to accurate point predictions, and unlike standard PLS, the Bayesian partial least squares approach provides prediction intervals with the correct coverage.

Our meetings will commence again after summer on the 26th September 2pm

Past meetings