Preventing undesirable behavior of intelligent machines.

Intelligent machines using machine learning algorithms are ubiquitous, ranging from simple data analysis and pattern recognition tools to complex systems that achieve superhuman performance on various tasks. Ensuring that they do not exhibit undesirable behavior-that they do not, for example, cause harm to humans-is therefore a pressing problem. We propose a general and flexible framework for designing machine learning algorithms. This framework simplifies the problem of specifying and regulating undesirable behavior. To show the viability of this framework, we used it to create machine learning algorithms that precluded the dangerous behavior caused by standard machine learning algorithms in our experiments. Our framework for designing machine learning algorithms simplifies the safe and responsible application of machine learning.

A Fully Customized Baseline Removal Framework for Spectroscopic Applications.

The task of proper baseline or continuum removal is common to nearly all types of spectroscopy. Its goal is to remove any portion of a signal that is irrelevant to features of interest while preserving any predictive information. Despite the importance of baseline removal, median or guessed default parameters are commonly employed, often using commercially available software supplied with instruments. Several published baseline removal algorithms have been shown to be useful for particular spectroscopic applications but their generalizability is ambiguous. The new Custom Baseline Removal (Custom BLR) method presented here generalizes the problem of baseline removal by combining operations from previously proposed methods to synthesize new correction algorithms. It creates novel methods for each technique, application, and training set, discovering new algorithms that maximize the predictive accuracy of the resulting spectroscopic models. In most cases, these learned methods either match or improve on the performance of the best alternative. Examples of these advantages are shown for three different scenarios: quantification of components in near-infrared spectra of corn and laser-induced breakdown spectroscopy data of rocks, and classification/matching of minerals using Raman spectroscopy. Software to implement this optimization is available from the authors. By removing subjectivity from this commonly encountered task, Custom BLR is a significant step toward completely automatic and general baseline removal in spectroscopic and other applications.

Matrix Effects in Quantitative Analysis of Laser-Induced Breakdown Spectroscopy (LIBS) of Rock Powders Doped with Cr, Mn, Ni, Zn, and Co.

Obtaining quantitative chemical information using laser-induced breakdown spectroscopy is challenging due to the variability in the bulk composition of geological materials. Chemical matrix effects caused by this variability produce changes in the peak area that are not proportional to the changes in minor element concentration. Therefore the use of univariate calibrations to predict trace element concentrations in geological samples is plagued by a high degree of uncertainty. This work evaluated the accuracy of univariate minor element predictions as a function of the composition of the major element matrices of the samples and examined the factors that limit the prediction accuracy of univariate calibrations. Five different sample matrices were doped with 10-85 000 ppm Cr, Mn, Ni, Zn, and Co and then independently measured in 175 mixtures by X-ray fluorescence, inductively coupled plasma atomic emission spectrometry, and laser-induced breakdown spectroscopy, the latter at three different laser energies (1.9, 2.8, and 3.7 mJ). Univariate prediction models for minor element concentrations were created using varying combinations of dopants, matrices, normalization/no normalization, and energy density; the model accuracies were evaluated using root mean square prediction errors and leave-one-out cross-validation. The results showed the superiority of using normalization for predictions of minor elements when the predicted sample and those in the training set had matrices with similar SiO2 contents. Normalization also mitigates differences in spectra arising from laser/sample coupling effects and the use of different energy densities. Prediction of minor elements in matrices that are dissimilar to those in the training set can increase the uncertainty of prediction by an order of magnitude. Overall, the quality of a univariate calibration is primarily determined by the availability of a persistent, measurable peak with a favorable transition probability that has little to no interference from neighboring peaks in the spectra of both the unknown and those used to train it.