The poster session at CAC attracted many participants, well over 100 posters were presented. I always look for posters (and talks) that utilize our software. It’s always fun and often enlightening to see what researchers are doing with our software, what methods they are using and what types of data they are working with. I list below the 15 posters I found which utilized PLS_Toolbox, MIA_Toolbox or Solo. Click on the links for copies of the posters.
That’s quite a range of data types and methods and some really excellent work! And yes, those last two posters are mine. I always enjoy presenting posters because it gives you a chance to talk to people one-on-one and have more interaction. And it’s an especially great format for new researchers to present and discuss their work.
Eigenvector Research President Barry M. Wise enjoying the poster session at CAC-2026 in Tarragona.
Thanks and congratulations to the organizers of CAC including Chairs Ricard Bouqué and Anna de Juan and the rest of the local committee for putting on a great conference!
Here at Eigenvector we’re working hard on our next releases of PLS_Toolbox (and Solo). Given the changes to MATLAB over the past year we thought you’d appreciate some information about the roadmap to our next updates.
Prior to the release of MATLAB R2025a, PLS_Toolbox was always compatible with the current release of MATLAB and an approximately 5 year window of past releases. However, as we pointed out in a January 2025 blog post, with The MathWorks’ (TMW) release of R2025a this was no longer possible. When MATLAB dropped support for Java it broke our most critical user interfaces (UI). To make matters worse, the R2025a uihtml interface tools did not include several key features that we needed to replicate previous PLS_Toolbox functionality. In order to provide the best experience for our users we had added lots of UI features missing from MATLAB, such as drag-and-drop, via Java.
Fortunately, the UI changes did not affect our command line functions through which all of our computations are done. So command line users and those who incorporated our functions into their own scripts were unaffected.
Throughout 2025 we kept in close contact with TMW and monitored the development of additional html interface tools. As of MATLAB R2026a some, but not all, of the missing functionality has been replaced. TMW continues to add and improve user interface functionality. We have, for instance, partially restored drag-and-drop with a uihtml component.
It is now our goal to have a version of PLS_Toolbox out that is compatible with R2026a+ by the end of 2026. This will require some changes to the work flow, but all essential functionality will be maintained. And the interfaces will have a more modern appearance (see below). Prior to that, by summer’s end, we will release our last version compatible with MATLAB R2024b and previous: PLS_Toolbox 9.6, This version will include new functionality including tools for lean chemometrics (IOT, etc.), conversions to volume fractions and improvements in Diviner. For our Solo users, we will build 9.6 with MATLAB 2024b, which will assure compatibility with current operating systems and improve performance and ease of installation.
New Model Optimizer interface in Dark Mode
MATLAB supports only a limited span of the newest Python versions. This is problematic because this also requires that associated libraries like scikit-learn be updated. In order for us to curate Python tools we need to be able to stick with versions that we’ve validated. We are adding our own Python integration techniques so that we can maintain support for the most stable Python library releases, not the most current (with open-source software the most current version is almost never the most stable version).
We also expect that MATLAB’s html interface tools will change from version to version for some time. We are unsure about how MATLAB’s shift to web technologies will impact our tools. This means that future versions of PLS_Toolbox will likely be more closely tied to specific versions of MATLAB—it may be much harder to achieve compatibility with a wide window of MATLAB versions.
We’re fully engaged in building our new interface infrastructure and look forward to delivering richer, improved interfaces. We appreciate your continued support as we do our best to support you!
Cleve Moler passed away May 20, 2026 and, quite rightly, much has been said about his contributions to the fields of mathematics and scientific computing. As creator of MATLAB his work had a profound impact on millions of scientists and engineers. The MATLAB universe includes tools for users in many different disciplines, from automobile and process control to finance. I’ve been using MATLAB for 40 years now and of course all of Eigenvector’s software products are built with it, including our for-MATLAB products like PLS_Toolbox and stand-alone products like Solo. It greatly accelerated the field of chemometrics, chemical data science and machine learning.
But for me, the significance of his work can be shown in a simple example. Last week a question came up about degrees of freedom in multivariate calibration. Just to make sure I was thinking correctly on why you have to account for mean centering as consuming a degree a freedom I created this example.
MATLAB commands illustrating that mean centering data consumes a degree of freedom.
Cleve gave us the tools to do linear algebra calculations easily in a language that felt natural. It’s hard to overstate what this did for education in this field. The ability to simply try things to confirm understanding is huge. And writing MATLAB code was so much easier because you didn’t have to pre-allocate memory or multiply matrices in loops and it gave sensible error messages making it easy to debug. My first day working with MATLAB was my last day writing code in BASIC or FORTRAN. In my earliest days in chemometrics I’d hear a talk at a conference about a new idea or method and then go back to my room and program it up. MATLAB made data science so much fun! Exciting times.
I am fortunate to have had a few interactions with Cleve. One of the earliest, in 1995, involved a discussion on how multi-way operations, i.e. tensor factorizations and multiplications might be built into MATLAB. The last, in 2024, had to do with singular matrix pencils in the generalized eigenvalue problem Ax = lambda Bx. This is a problem sometimes in the Generalized Rank Annihilation Method (GRAM). In a note about this to our Eigenvector staff Cleve added:
Let me tell you: when Cleve Moler asks if you are interested, there is only one answer!
Many years ago we gave Cleve one of our shirts with the Eigenvector logo embroidered on it. He said, “I get lots of swag, but I’ll actually wear this.” It makes me happy to think that maybe he did. Given the work he did on LINPACK, EISPACK and eigenfunctions of vibrating membranes, I think he found it amusing that somebody would start a company called “Eigenvector.”
Cleve Moler’s work had an enormous influence on me; without it Eigenvector Research might not exist. That impact flows to my fellow Eigenvectorians and the thousands of users of our software. Rest in peace Cleve. Thank-you for everything, or as we say here at Eigenvector, EVRIthing.
Like most software packages, our PLS_Toolbox, Solo and other products come with a maintenance program. When you are “in maintenance” you have access to support through our helpdesk@eigenvector.com email address. And unlike the software support of many firms, our Helpdesk is monitored our software development staff team members including Scott, Donal, Lyle, Nate and Bob. This is the crew that is actually responsible for writing our software. So when you contact Helpdesk you’re actually talking to the people behind the code. That’s why our customers say our Helpdesk is actually helpful!
Developing PLS_Toolbox and Solo in MATLAB.
Of course most of our customers never need any help. So why should they purchase maintenance? Because we’re always improving our software, and when you are in maintenance you have access to all the improvements and new features. For instance, here are the highlights of what we’ve added in our major releases since October 2021:
Version 9.0 — Faster engine, modern Machine Learning methods
Updated Solo engine to MATLAB R2020b for improved speed and graphics. Added Python-integrated methods including deep learning neural networks (ANNDL, ANNDLDA) and nonlinear pattern recognition (UMAP, t-SNE).
Version 9.1 — Parallel Computing, improvements to non-linear methods
Parallel computing enabled to speed computation in many non-linear methods, improvements to ANNDL, ANNDL-DA, UMAP
Version 9.2 — Smarter classification and spectral filtering
HMAC automated hierarchical model builder. Grey CLS with GLS and EPO filters for interference correction.
Version 9.3 — Temperature correction, interpretation and expanded classification
Temperature-interpolated Classical Least Squares (CLSTI). Shapley values and sensitivity for nonlinear model interpretation (EVRISHAPLEY). Linear Discriminant Analysis (LDA).
Diviner: a high-throughput search engine that accelerates calibration model development by systematically exploring preprocessing and variable selection combinations.
Our next major release will be PLS_Toolbox/Solo version 9.6 which we plan to ship this summer. This version will include further improvements to Diviner plus the addition of lean chemometric methods and tools for working with volume fractions. We’ll also be building Solo with MATLAB R2024b which will assure continued compatibility with the latest Windows and MacOS operating systems. And for you PLS_Toolbox users stay tuned for news about compatibility with MATLAB R2026.
When developing calibration models for spectrometers the reference values are often given in weight percent. This makes sense as it is easy to make up reference samples by weighing the components. But in absorbance spectroscopy the measurement is related to the volume fraction of each component [1,2]. In cases where the density of all components is about equal there is little difference between the mass and volume fraction. But when the difference in densities are large the difference is considerable.
As an example, consider a hypothetical binary liquid mixture where component A has a density of 0.8 and component B a density of 1.2. Assuming linear mixing (no volume change on mixing) a 50/50 by weight mixture is 60/40 by volume (A/B). Assuming Beer’s law, the spectrometer would see this sample as 0.6 times the pure component spectra of A plus 0.4 times the pure component spectra of B. Based on this it’s easy to see that a calibration done on volume fractions would be linear whereas one done on mass fractions would not.
Consider the 5 component mixture data collected by Windig and Stephenson [3]. This is a mixture of butanol, dichloromethane, methanol, dichloropropane and acetone. Samples were measured in the NIR (1100-2498nm, 700 wavelengths) and the data was split into 35 calibration samples (2 replicates each) and 35 test samples (also with replicates). Classical Least Squares (CLS) models were built using both mass fraction and volume fraction. The volume fractions were computed from the original mass fractions using densities of [0.810 1.326 0.790 1.160 0.790] obtained from the literature. The calibration curves are shown for mass fractions in Figure 1 and volume fractions in Figure 2. The errors for mass and volume fractions are shown in Tables 1 and 2 respectively. Because both the mass and volume fractions are in the same range (0-1) with nearly equal variance these tables are comparable. Compensating for the slight differences in variance and aggregating over all 5 analytes the model based on volume fractions has lower RMSEP by about 24%.
Figure 1. Calibration curves for Windig mixture data based on Mass Fractions.
Figure 2. Calibration curves for Windig data based on Volume Fractions.
Table 1: Root-mean-square errors of Calibration, Cross-validation and Prediction for Windig data based on Mass Fractions.
Mass Fractions
Butanol
Dichloro-methane
Methanol
Dicholoro-propane
Acetone
RMSEC
0.0102
0.0081
0.0100
0.0157
0.0098
RMSECV
0.0118
0.0090
0.0128
0.0202
0.0124
RMSEP
0.0108
0.0091
0.0079
0.0131
0.0092
Table 2: Root-mean-square errors of Calibration, Cross-validation and Prediction for Windig data based on Volume Fractions.
Volume Fractions
Butanol
Dichloro-methane
Methanol
Dicholoro-propane
Acetone
RMSEC
0.0098
0.0071
0.0062
0.0118
0.0070
RMSECV
0.0115
0.0091
0.0083
0.0160
0.0088
RMSEP
0.0096
0.0066
0.0058
0.0087
0.0063
24% is a nice reduction, but why isn’t it more? Or maybe a better way to ask this is why isn’t the calibration based on mass fractions worse than it is? It is worth noting that the estimated pure component spectra for the two cases are not the same. They are compared in Figure 3. Estimates for the pure component spectra of dichloromethane show the largest deviations, but differences exist in all components. The pure components are, of course, estimated to obtain the best fit to the observed mixture spectra so they will differ depending on whether mass or volume fraction is used. In other words, the pure spectra are derived to best fit the data whether it is in mass or volume fraction.
Figure 3. Comparison of pure component spectra estimates based on mass fractions and volume fractions.
What if you have data in mass percent and don’t know the density of the constituents? Consider the casein-glucose-lacate (CGL) data fromNæs and Isaksson[4]. This data includes 153 calibration and 78 test samples from a designed experiment, 117 wavelengths, 1104-2496 nm. These are powders, and while you can look up their densities, they are highly dependent upon how the samples have been handled. A CLS calibration based on mass fraction is shown in Figure 4. Errors are collected in Table 3. The calibration curves suggest that there is some non-linearity as might be expected if the densities of the constituents were markedly different.
Figure 4. Calibration curves for CGL data based on mass fractions.
Table 3. Root-mean-square errors of Calibration, Cross-validation and Prediction for CGL data based on Mass Fractions.
Mass Fractions
Casein
Glucose
Lactate
Moisture
RMSEC
0.0369
0.0687
0.0450
0.0096
RMSECV
0.0398
0.0717
0.0456
0.0097
RMSEP
0.0301
0.0469
0.0313
0.0068
It is possible to estimate a set of densities for the CGL system using optimization. The objective function for this is the sum of squared difference between the calibration spectra and the reconstructed spectra using the CLS estimates and the estimated volume fraction. The MATLAB function fminsearch was used to find a set of densities to minimize this error. Note that because only the ratio of densities matter, the first density for casein was set to 1. The minimum for this problem is fairly shallow and solutions get far from 1 with little improvement in the objective function so a small penalty was added for the distance of the solution from unit densities. The result is [1.00 2.03 0.66 1.00], which reduces the reconstruction error of the spectra from 2.46 to 0.405, a factor of ~6. The reality of these densities is debatable so we will refer to them as apparent densities and apparent volume fractions.
A CLS calibration using the apparent volume fractions is shown in Figure 5 and the errors are collected in Table 4. When the errors are adjusted for differences in variance upon conversion from mass to apparent volume the average error is reduced by 43%. Note that the non-linearities obvious in the mass fraction calibration curves have been mitigated. (It is also possible and perhaps even preferable to do this calibration without the moisture. When this is done very similar results including improvements when using apparent volume fractions are obtained.) If final estimates in mass fraction are desired it is of course a simple matter to convert the volume fraction estimates from the model back to mass fractions.
Figure 5. Calibration curves for CGL data based on apparent volume fractions.
Table 4. Root-mean-square errors of Calibration, Cross-validation and Prediction for CGL data based on Apparent Volume Fractions.
Volume
Casein
Glucose
Lactate
Moisture
RMSEC
0.0240
0.0197
0.0173
0.0033
RMSECV
0.0253
0.0211
0.0181
0.0034
RMSEP
0.0209
0.0162
0.0154
0.0029
In conclusion, the use of volume fractions can clearly improve calibrations. Here is it shown for CLS models but it is also true that inverse least squares models can be improved (more on that later). The amount of improvement is dependent upon the variation in the (apparent) ratio of densities of the pure components in the mixture. Provided sufficient data is available, apparent volume fractions can be estimated which in turn improves calibrations.
[1] H. Mark, R. Rubinovitz, D. Heaps, P. Gemperline, D. Dahm, and K. Dahm, “Comparison of the Use of Volume Fractions with Other Measures of Concentration for Quantitative Spectroscopic Calibration Using the Classical Least Squares Method,” Appl. Spectrosc. 64, 995-1006 (2010).
A few weeks back Rasmus Bro sent us an email about a little experiment he performed. His colleagues were discussing the utility of having students write reports, wondering if the reports could be replaced with output from ChatGPT. So he submitted the two figures below and prompted it with:
I am writing a report for a university course on advanced chemometrics. I have built a PLS prediction model predicting Real Extract in beer (40 samples) from the attached VISNIR spectra (spectra.jpg) and yielding the regression vector in regressionvector.jpg. I am asked to critically describe what I learn from the regression vector
ChatGPT produced an excellent report, the full text of which you can find here. The only problem? The very first thing out of its virtual mouth was wrong. The report stated that “Positive [regression] coefficients mean absorbance at that wavelength is positively correlated with Real Extract” and that “Negative coefficients mean absorbance at that wavelength is negatively correlated with Real Extract.”
This not correct in general, and not correct in this instance either, although it is closer in this case than usual. The regression vector for this example is shown below, with coefficients colored blue for negative correlation between Real Extract and absorbance at that wavelength and red for positive correlation. You can see that the negative coefficients tend to be blue, and the positive ones tend to be red, but there are plenty of instances of the reverse of this.
In general, the correlation between absorbance at a wavelength and the chemical analyte concentration or property of interest can be positive, negative or zero. It all depends on the degree of overlap of the components and whether the interfering species are positively or negatively correlated with that concentration. And all of that is actually independent of the regression vector. The regression vector is the part of the analyte signal that is orthogonal to the interferents. This doesn’t change regardless of how these interferents are correlated with analyte of interest (except in the extreme case of them being perfectly correlated or anti-correlated in which case orthogonality can’t be determined).
We teach this in our Variable Selection course, and it demonstrates one of the reasons that variable selection is so difficult. You can’t just pick variables that are correlated with the analyte of interest. The figure below shows on the left the spectra of a system where the analyte of interest is entirely overlapped by an interferent, (which is not at all uncommon in NIR!). On the right it is shown that, depending on the degree of correlation between the interferent and the analyte, the correlation between wavelengths and analyte can be positive, negative or zero. More complex systems can produce even more unintuitive behavior, such as instances where the variables most correlated to the analyte of interest have no signal from it all.
All this demonstrates the problem with ChatGPT: it is no smarter than all the text it has scraped up. (Maybe it will get smarter after reading this.) Common misconceptions, like this one, will likely be regurgitated. That means that EVERYTHING it produces is suspect and therefore requires that it be verified before being used for anything important. This is even scarier when you realize that it doesn’t know when it is training on misinformation that it produced, thus reinforcing and perpetuating misconceptions. Getting back to original human-produced sources is getting increasingly difficult.
Bottom line: if you are going to use ChatGPT or its brethren allow yourself some time to verify its output. When you do, you might find it isn’t saving you as much time as you think.
At Eigenvector we’ve been teaching classes in chemometrics and machine learning since our founding more than 30 years ago, and I’ve been doing them slightly longer than that. Most of our classes are taught with equations that describe the various models we cover and how they are computed. (The exception of course is our Chemometrics without Equations series.) But of course in order for the equations to be helpful you actually need to understand them. It has been said that linear algebra is the language of chemometrics, and I think that applies to most data science fields. Certainly some calculus is required, but the most fundamental methods are succinctly expressed in linear algebra notation.
We’d really like our students to understand what is going on “behind the scenes” with the methods they are using, so we start most of our classes with the course we now call “Linear Algebra for Machine Learning and Chemometrics.” It covers the basics, including
Vector and matrix operations (addition, subtraction, multiplication)
Inner and outer products
Vector spaces, subspaces and null spaces
Projections onto vectors and subspaces
Basis sets, orthogonal and orthonormal matrices, linear independence
Gaussian elimination and solving systems of equations
Matrix inverses and least squares
Matrix rank, rank deficiency and ill-conditioned matrices
Singular Value Decomposition (SVD)
Pseudoinverses
Most college graduates from the hard sciences took a course in linear algebra that covered these topics, but if they were like the one I had, never related these concepts to actual practice. Given that we employ these methods to describe and manipulate real chemical data we have many examples of how the concepts are used! An example of that, where we illustrate an outer product as a concentration profile times a pure component spectrum, is shown below.
If you understand the concepts outlined above you’re in pretty good shape to start exploring multivariate methods and you’ll have an appreciation of how they work and, perhaps more importantly, why they sometimes fail. Rank, in particular, is such an important concept in modeling and data science that I don’t see how you could do a credible job of modeling without a solid grasp of it. Likewise for ill-conditioning, shown below.
We assign our Introduction to Linear Algebra as pre-course homework (reference below). It covers most of the material in our Linear Algebra short course, and includes exercises which can be followed in MATLAB. Please feel free to download and use it. If you want to have a good foundation for your data science education, begin at the beginning!
As part of our effort to make a future version of PLS_Toolbox compatible with the new MATLAB2025a interface tools (and deprecation of Java) we surveyed our users to find out more about what they do and how they work. We got a nice response and some of the results were quite interesting, so I thought we’d share them here. We used Survey Monkey to conduct the survey so for the most part I just copied the figures from it.
When asked what version of MATLAB was being used with PLS_Toolbox nearly 50% of our users responded with R2024b. This surprised me as I figured there would be a smaller percentage of users with the most current version of MATLAB. I’m reticent to update my MATLAB (and other software) versions unless I have a compelling reason to do so. My experience with R2024b is that it is an excellent release, (some are better than others), so that might be part of the reason. We also found that ~75% of users were on the current version of PLS_Toolbox, 9.5.
When asked how they primarily interacted with PLS_Toolbox about 75% answered with a mixture of use of the interfaces and use of the command line. The most frequent responses were an equal mix of both and mostly interfaces with occasional command line. I’m in this latter category myself. But this response really highlights one of the major strengths of PLS_Toolbox: you can use it either way. You can make models from the command line that are completely compatible with models generated by the interfaces, and vice-versa. How you work largely depends on the situation and personal preferences.
When asked about what types of data users were working with, it was no surprise that over 80% were working with NIR followed by 50% using Raman. Raman use has grown a lot in the last several years and we expect that trend to continue. UV-VIS was a larger group than I expected at a little over 35%. I was pleased that over 30% use it with chemical process data as that is how I got into this business to begin with. We were surprised to find that the “other” option produced quite a few instances of NMR use.
We asked our customers how frequently they used various analysis tools and methods. It was no surprise to us that over 80% did exploratory analysis, which includes PCA, frequently. Over 80% also use linear regression, which includes PLS, frequently. PCA and PLS are certainly the workhorses of chemical data science. About 70% of users accessed the classification methods in PLS_Toolbox at least occasionally while over 50% of them used the non-linear regression methods at least occasionally. Diviner, our new semi-automated machine learning tool for regression model development was used at least occasionally by about 30%. Given that Diviner is the newest major tool in PLS_Toolbox this seems like a good start and I expect that number to grow.
Over 70% of customers responding to the survey use PLS_Toolbox weekly or daily. Of course I expect that the heavier users were more likely to answer the survey, but still, that’s a lot. I only get a chance to use it weekly myself!
Many of our users have been with us for a long time. Over 70% responded that they’ve been using PLS_Toolbox for more than 5 years. When you combine that with the amount they use it above that accounts for a lot of use. This ultimately contributes to the reliability of PLS_Toolbox: issues get found and our development staff fixes them.
Finally, we asked users to rate their satisfaction with PLS_Toolbox on a scale from 1 to 100. A histogram of the responses is shown below. 55% of our customers gave PLS_Toolbox a rating of 90 or higher. The average was 86. Still room for improvement and of course we’re working on how to bring the bottom end of the scale up!
We of course asked users about what they find difficult and what they’d like us to add, etc., and got some good suggestions. So plenty to work on, as always! But I’ll leave my favorite comments here.
Love the product, creativity and quick support responsiveness!
Very complementary to my own MATLAB coding – often I jump back and forth. Nice PCA interface. I appreciate the missing data capabilities in a lot of the code.
I love Diviner. Please add preprocessing sets for the spectrosopic/chromatographic methods.
I have stuck with this product through the years as I trust the software, I find it very powerful (certainly better than anything my colleagues can do in free or shareware), and it is constantly changing and improving.
Thanks to all our users who completed the survey and provided this feedback!
The 19th Eigenvector University wrapped up on Friday, May 16. A little smaller this year, we had 32 students from locations as far away as Milan, Italy and Santiago, Chile. EigenU 2025 started with online classes the week before including the new “Introduction to PLS_Toolbox and Solo” developed by our Lyle Lawrence. In person classes started Monday, May 12 at the Washington Athletic Club in Seattle with “Chemometrics I: Principal Components Analysis.” This year Neal Gallagher and I covered the teaching duties for this class on PCA, arguably the most important method in modern data science. And in spite of the fact that I’ve taught this 100+ times over 30+ years I’m still enthusiastic to do so. PCA is just so useful and the insights gained from it are often significant, and, well, it’s just fun to see what’s in data sets!
The rest of the EVRI staff caught up with us on Monday evening. As always, EigenU is an opportunity for us to get together and we enjoyed a nice dinner at Tulio on Tuesday (below). This is a rare photo in that it includes Jill, my spouse and the person that runs the Eigenvector “back office.” It is always great to have the whole crew together!
Eigenvector staff including (from left) Lyle Lawrence, Manny Palacios, Barry Wise, Jill Wise, Scott Koch, Nate Watson, Donal O’Sullivan and Bob Roginski.
Wednesday night we were at the Top of the WAC with the “PowerUser Tips & Tricks” session where the EigenGuys take turns presenting their favorite underutilized or new software feature. Also the “PLS_Toolbox/Solo User Poster Session” was held, where users get to share analyses they’ve been working on. This year’s best poster winners were Shirmir Branch with “Using PLS_Toolbox for the Analysis of Complex Molten Salt Chemistry” and Benjamin Panitz with “Development of a Chemometrics Model for Rapid Fish Egg Quality Assessment in Aquaculture.” Both of these posters addressed really interesting issues. Shirmir’s involved developing methodology to monitor metals including radioactive elements in the molten salt cooling system in nuclear reactors. The problem of selecting optimal fish eggs for production in hatcheries. Shirmir and Ben took home Bose SoundLink Bluetooth Speakers for their efforts!
Eigenvector Research Vice-President Neal B. Gallagher presenting Shirmir Branch with the prize for her winning poster.
Eigenvector Research President Barry M. Wise presenting Benjamin Panitz with a prize for his winning poster.
Professor Rasmus Bro joined us again this year, teaching several classes including the new “PLS_Toolbox from the Command Line.” If you want to learn how to automate analyses or make them run sans interfaces this is the place to start. We’ve found that most of our experienced PLS_Toolbox users use a mix of interfaces and command line, as do we in our roles as consultants. It is a very powerful combination!
EigenU 2025 concluded with Thursday evening’s workshop dinner, this year in Haggerty’s Sports Bar, and Friday’s “Non-linear Machine Learning for Calibration and Classification.” In all sixteen hands-on classes in chemometrics and machine learning were presented along with the two evening events. A very busy week for sure! We wish all our students the best of luck as they employ their new data science skills back at home.
It may be that the only way to improve your grit is to do things that are hard. — Clare E. Wise, M.D.
The end of the ski season is upon us and with it the end of the ski racing season. It has been a great year and I’ve been fortunate to officiate at alpine ski racing events at (almost) all levels this year, with athletes from 6 to 86 (yes, really).
We’re big fans of ski racing and I’m pleased we’re able to sponsor two ski teams under the Eigenvector banner, the Schweitzer Alpine Racing School (SARS) and the Mission Ridge Ski Team (MRST). Given that ski racing has nothing to do with chemometrics, machine learning or, say, spectroscopy, you may be wondering why we choose to sponsor ski racing. The answer? Because it’s hard. And learning to do things that are extremely difficult under highly variable and largely non-ideal circumstances improves grit and resilience. This is true for young people and adults.
Grit made its way into the popular lexicon upon the publication of Angela Duckworth’s eponymous book in 2016. Grit is defined as “passion and perseverance in pursuit of long term goals,” or alternately, as “courage and resolve, strength of character.” (For a quick introduction to grit see Duckworth’s 2014 TED talk.) Grit is closely related to resilience, “the capacity to withstand or to recover quickly from difficulties and setbacks.” Though hard to measured accurately, (grit is assessed using self reported questionaires), grit is the single biggest factor in predicting the success of people in everything from math scores to recovery from surgery.
So how do you build grit? Duckworth’s TED talk from 11 years ago ends with a “we don’t know.” Since then additional research has been done. Developing long term habits, such as doing the New York Times crossword puzzle every day, (or in my case, studying French on Duolingo), appears to contribute. But the most effective way, and maybe even the only way to build grit and resilience is to repeatedly do things that are hard. Things that you often fail at but learn to recover from and progress.
Ski racing provides many opportunities to fail and recover. Is totally objective: time is the only thing that counts and it is measured to the hundredth of a second. But course sets, surface and weather conditions vary widely and races are seldom run under ideal conditions. It is inherently unfair because no two athletes ski exactly the same course: the surface degrades and the weather can change in an instant. Add to that frequent schedule changes and the resulting equipment and mental preparation gymnastics. There are many, many variables that are out of the control of the athletes.
Athlete at Downhill Start during Eigenvector Western Region Speed Series at Schweitzer Mountain Resort, Sandpoint, Idaho.
Beyond grit and resilience, ski racing (and other sports and activities too, of course) improve the mental health and social development of young people. Research discussed in Greg Lukianoff and Jonathan Haidt’s “The Coddling of the American Mind: How Good Intentions and Bad Ideas are Setting a Generation up for Failure” and Haidt’s recent “The Anxious Generation: How the Great Rewiring of Childhood is Causing an Epidemic of Mental Illness” indicates the rise of “safetyism” contributes to arrested development and “failure to launch.” Participation in skiing in general is the opposite of safetyism, skiers can and do get injured. Ski racers even more so. Having to pick yourself up from “face-plants” is a pretty regular thing. And while we wouldn’t wish serious injury on anyone and do our utmost to prevent it, ski racers and other athletes learn a lot about what they are capable of when they are recovering. Haidt’s (and other’s) work also indicates that time spent outdoors with friends away from cell phones is clearly good for the mental health of young people.
For a nice summary of recent research on this topic, please see “Too Grit to Quit: Building Grit & Resilience in Orthopedic Patients & Providers,” a Grand Rounds talk by Clare Wise. Though focused on the world of orthopedic surgery, the research covered includes many generally applicable results, along with some interesting information about what the US Ski Team is doing to improve the grit of their athletes (a pretty gritty bunch to begin with).
Our daughters Clare and Mattie, who have both been ski racers and coaches, have many fellow ski racer friends and I am constantly amazed at what a fun, outgoing, and very successful group they are. Their ski racing experience surely contributed to this. Eigenvector is proud to support that experience for future and current generations.
I’ve learned a number of things using Diviner, our new semi-automated Machine Learning tool for creating regression models. When you make models in large numbers on each data set you are presented, you begin to see some trends. For me, one of these was the discovery of how some data preprocessing methods work well surprisingly often. (I’m still flabbergasted that SNV followed by autoscaling works well for many spectroscopic data sets.) Another thing that happens is that you begin to see “anomalies” much more often when you make models by the hundreds.
The anomaly of interest here is the occasion when the error of calibration (root-mean-square error of calibration, or RMSEC) exceeds the error of cross-validation (root-mean-square error of cross-validation, or RMSECV). This is generally not expected as what this is saying is rather counter-intuitive: a model’s ability to predict data when it is left out is better than its fit to the data in the model. (!) Cross-validation curves for a Partial Least Squares (PLS) model generally looks like Figure 1 below.
Figure 1: Typical Calibration Error (RMSEC) and Cross-validation Error (RMSECV) Curves for a PLS Model for Glucose from NIR Spectroscopy.
If you look closely in Figure 1 you’ll see that the RMSECV line (blue) is above the RMSEC line (red) in all cases. This is perhaps more obvious if you plot the ratio of RMSECV to RMSEC, as shown in Figure 2. In all instances this ratio is greater than 1. We refer to this as the “Overfit Ratio,” and it has become one of our favorite diagnostics for selecting model complexity. Ideally, a model’s fit error on the calibration data should not be much lower than its prediction error on the same data. If it is then the model is fitting the data much better than its ability to predict it which suggests it is fitting the noise in the data and is, therefore, overfit. It is typical for the overfit ratio to increase as the model complexity increases.
Figure 2. Overfit Ratio RMSECV/RMSEC for Model in Figure 1.
But every once in a while we get cross-validation curves that look like Figure 3, where the ratio RMSECV/RMSEC is great than 1 except at 6 PLS components. Note that this is for a 5-fold cross-validation split of the Casein/Glucose/Lactate (CGL) NIR data (compliments of Tormod Næs), where we are making a model for Casein and have pre-processed the data with a second derivative (21 point window, 2nd order polynomial).
Figure 3. Calibration and Cross-Validation Error Curves for Casein (5-fold, 2nd derivative, 21-pt window).
When presented this way the anomaly is easy to ignore. In Diviner, however, we like to plot the model overfit RMSECV/RMSEC ratio versus the predictive ability, RMSECV. The idea being that the best models are the ones that have good predictive ability without being overfit, i.e. the models in the lower left corner. But when RMSECV < RMSEC, then RMSECV/RMSEC < 1, and these points stick out like a sore thumb on the plots, as shown in Figure 4, lower left corner. This is rather hard to ignore.
Figure 4. Overfit Ratio RMSECV/RMSEC versus Predictive Ability RMSECV for 80 Models with Different Combinations of Preprocessing and Variable Selection from Diviner. Number of PLS Latent Variables Shown.
My first question when I saw this phenomena was, “is it reproducible, or is it some quirk in the software?” The answer to that is that it is definitely reproducible. I should note here that we are doing conventional/full cross validation: the entire model is rebuilt from scratch using the full data set minus each left out split, no shortcuts, such as not recalculating the model mean, or cross-validating only the current factor, are used.
The next question is “what conditions lead to this phenomena?” On that I’m substantially less certain. First it seems likely that this is NOT possible when using Leave-one-out (LOO) cross-validation. When leaving only one sample out of a model, the model certainly must swing towards that sample when it is included rather than excluded. I don’t think it has anything to do with some quirk of preprocessing during cross-validation. We see this sometimes using very simple preprocessing methods including row-wise methods (such as the derivative used here or sample normalization) that don’t change with the data split.
I suspect it has to do with the stability of the model at the given number of components. If there are two components in the model (with all the data) that have nearly the same co-variance values (variance captured in x times variance captured in y) then a small change in the data, such as leaving a small amount of it out, can cause the factors to rotate or come out in different order. Thus the model is quite different in the last factor or factors. This instability can lead to a “lucky guess” as to the predicted values on one or more of the left out sample sets.
As a test of this model instability theory the agreement between the calibration and cross-validation y-residuals is considered. Figure 5 shows the cross-validation residuals versus the calibration residuals for the 6 component model. The samples are colored by their T2 values. For comparison the residuals from the 7 component model are shown in Figure 6. It is clear that there is much better agreement between residuals in the 7 component model, which shows the usual overfit ratio greater than 1 behavior.
Figure 5. Cross-validation versus Calibration y-residuals for the 6 Component Casein Model Showing Substantial Differences Between Residuals, colored by T2 values.
Figure 6. Cross-validation versus Calibration y-residuals for the 7 Component Casein Model Showing Good Agreement Between Residuals, colored by T2 values.
It is interesting to note that samples with the largest difference in residuals for the 6 LV model tend to have high T2 values, as shown in Figure 5. Many of these samples also have relatively high Q-residual values as well, as shown in Figure 7. The fact that the samples that have the largest disagreement tend to be samples that could have a high impact on the model when included versus excluded suggests (to me at least) that model instability at the given number of factors is key to the overfit < 1 problem.
Figure 7. Cross-validation versus Calibration y-residuals for the 6 Component Casein Model Showing Substantial Differences Between Residuals, colored by Q-residual values.
There is much still to be done to fully investigate the cause of the RMSECV<RMSEC phenomenon. The practical question, of course, is what to do about it when it is observed. My suggestion is to simply not trust the model for the given number of factors. Therefore do not consider it as a candidate for final model selection. It’s fun to think you could get a model that predicts better than it fits but it likely the results of model instability that produces some good lucky guesses.
Eigenvector Research has always worked to make PLS_Toolbox compatible with the most current version of MATLAB, plus an approximate five year window of older versions. Compatibility with the upcoming MATLAB 2025a, however, presents unique challenges. In their transition to an entirely HTML-based interface, The MathWorks (TMW) has removed support for Java in 2025a. As a result, many elements in our graphical user interfaces will not function in 2025a. To make matters even more complicated, the Java functionality that we have relied on has not been completely replaced by HTML equivalents. Thus, PLS_Toolbox will not be compatible with MATLAB 2025a.
TMW is aware of these issues and we are working with them (and other MATLAB experts) to find solutions that will maintain the high level of user-friendliness in our current interfaces in future versions of MATLAB. We want to make sure that the solutions we come up with are user-friendly, stable, and have a long lifetime. This will be challenging and will take time and careful consideration. We expect to have a version of PLS_Toolbox with undated interfaces available in late 2026.
If you use PLS_Toolbox then we recommend that your organization not adopt MATLAB 2025a or later. If you choose to do so, we expect that our command line functions will still work as always, but most of our interfaces will not. In its current state, we cannot offer support for MATLAB 2025a. We will, of course, happily support PLS_Toolbox, MIA_Toolbox and Model_Exporter on MATLAB 2019b through 2024b. And of course we will continue to expand and refine the capabilities of our software as we work to become compatible with future MATLAB versions.
Note that users of our stand-alone software Solo and its variants (Solo+MIA, Solo+Model_Exporter) will not be affected by this change, nor will users of our prediction engine Solo_Predictor.
Please write to bmw@eigenvector.com if you have any questions about this. We’re happy to give more detailed explanations of the technical issues. We would also be interested to know if your organization is affected by the removal of Java from MATLAB and how you are dealing with it.
Eigenvector Research was founded on January 1, 1995, which means that we just turned 30. When I mentioned writing a piece for the occasion of our 30th anniversary, our Donal O’Sullivan replied “I don’t know if you want people to know we’re that old!” And I understand where he’s coming from. In the software business, especially in data science, maybe you don’t want to advertise that you’ve been around for 3+ decades. Perhaps better to look like the bright shiny new thing.
But I think that 30 years of experience counts for something. We’ve seen quite a few shiny new things get misused and abused, mostly by people that don’t appreciate the basics that you just can’t get around. Things like if your model is purely based on data (rather than physics and chemistry) you can’t expect it to work outside the range and subspace of the calibration data. And the more effects you have contributing to the variance in your data, the smaller the unique part of the signal you are looking for will be, until it disappears all together. And if you are fitting your data better than your ability to predict it you’re fitting the noise or clutter. And, and….
The downside of turning 30 in the software business is what you might call technical debt, except most definitions of technical debt focus on the cost of implementing short term work-arounds instead of long term solutions. But if you’ve been in the business as long as we have, you know that, in spite of the advantages they initially offered, most software frameworks are eventually abandoned. (Remember ActiveX?) Learning to utilize new technologies to add new methods and features to software is fun, but replacing old architecture is hard work. Over the last couple years we’ve put a lot of effort into updating the infrastructure that supports our users behind the scenes. In the coming year we’ll be focused on updating some of the technologies behind our end user software PLS_Toolbox, Solo and their variants while still improving the existing methods. It is going to be challenging!
Looking back, 2024 was a great year. It was definitely the “year of training.” Between our in person classes (Eigenvector University in Seattle and Rome) our online classes (both open and for specific companies) and our recorded courses we reached more students than ever. And we also had our biggest year of software sales. Associations with our instrument company partners and software resellers was a big part of that. We were also happy to help many new users transition from other software packages. And of course our students, users and consulting clients all benefited from that 30 years experience.
Our biggest software development of 2024 was the release of Diviner, our semi-automated Machine Learning tool for accelerating the development of calibration models. Diviner automates the construction of regression models and keeps the analyst in the loop so that they can learn from the process. We have lots of plans for improving and expanding the use of Diviner: even as useful as it is now, there is still much to do!
Example output from Diviner. Each point represents a model with different preprocessing, variable selections and meta-parameters. The best models are the ones that have the best predictive ability but are not overfit.
When young people turn 30 they are often a bit depressed to be leaving their 20’s behind. (A search on “turning 30” reveals lots of this, along with a lot of stuff that is really, really not useful!) But generally that feeling is soon replaced with the realization that they’ve entered a very productive period of their life, where they can make real progress on careers and relationships, both business and personal. As a company we feel the same way. Here’s to life beyond 30!
I believe that days go slow and years go fast And every breath’s a gift, the first one to the last – Luke Bryan, ‘Most People are Good’
Eigenvector Research starts its 30th year this January. Reflecting on that made me think of the Luke Bryan quote above. Some days have been long, for sure. For me that’s particularly true when I’m doing administrative stuff. There are also some days that go incredibly fast. When you are “in the zone” writing code to investigate a new method you get to five o’clock and wonder where the day went. But all of the years go fast. This really hits me when I think of the things I started long ago but still haven’t finished (like that journal article on Gray Classical Least Squares). And of course when you realize your children are adults with job titles like “Senior Financial Analyst” and “Orthopedic Surgical Resident” you wonder where the years went.
Every breath is a gift, especially when you are doing something you love. In the field of chemometrics (chemical data science, machine learning) there’s just no end of things to think about and explore. And of course part of the fun is working with such a great crew. Here’s our technical staff ( I guess I can still include myself in that) at Shuckers during the 17th Eigenvector University last May. We group up differently depending on the project (software, consulting, training, webinars, helpdesk, etc.) so I get to work with EVRIbody from time to time, as do we all. That really keeps it interesting, in part because we span a wide range of ages (almost 40 years) and hobbies (ski racing, rowing, trail running, biking, baseball, photography, playing the bagpipes) and other interests (cosmology, cooking, movie and music trivia, French). (Et bien sûr j’adore encore partager le bureau derriere avec ma femme Jill, la Directrice Générale!)
The Eigenvector Technical Staff: Lyle W. Lawrence, Sean Roginski, R. Scott Koch, Shamus Driver, Barry M. Wise, Neal B. Gallagher, Robert T. “Bob” Roginski, Manuel A. “Manny” Palacios.
It piles up. When I checked this morning there have been 24,532 commits to our software version control system. Version 9.3 is the 38th release of PLS_Toolbox. We’ll have our 18th EigenU this year and our 13th Eigenvector University Europe. I lost count long ago of how many short courses we’ve done but it’s way over 200. We’ve done over 30 webinars, and have had consulting projects with 50+ companies and national laboratories. And no end in sight!
Neal and I joke that we got into this because life is too short to drink bad coffee, bad beer, do boring work or live in a crappy place. That sounds flippant but it is actually true. So a big THANKS! to all our customers who have enabled us in this software odyssey and intellectual pursuit. We have a lot planned for 2024 that we think you will find useful! We hope to continue to serve you.
“The age-old concept of data modeling with spectroscopy has been revitalized through the integration of Machine Learning and AI initiatives. The pharmaceutical industry, in particular, is embracing data science, exploring the potential of deep learning and AI tools in spectroscopy applications. The emergence of open-source tools adds transparency to the ‘black box’ of these advanced technologies, sparking discussions around regulatory concerns.”
I’ve been doing data modeling and spectroscopy for 35+ years now and I’ve never felt particularly un-vital (vital–adjective; full of energy, lively). However, there is certainly more vibrancy in chemical data science right now largely due to the hype surrounding Artificial Intelligence/Deep Learning (AI/DL). Rasmus Bro wrote me with “I feel like AI/DL has sparked a new energy and suddenly we are forced to think more about what we are and how we are.” But the part of this bullet that really got my attention is the part I’ve underlined. I wasn’t at the meeting so I don’t know exactly what was said, but I must take issue with the idea that open-source has anything to do with the transparency of ‘black-box’ models.
First off, there is a pervasive confusion between the software that generates models and the models themselves–these are two separate things. The software determines the path that is followed to arrive at a model. But in the end, how you get to a model doesn’t matter (except perhaps in terms of efficiency). What’s important is where you wound up. That’s why I’ve said many times: VALIDATE THE MODEL, NOT THE SOFTWARE THAT PRODUCED IT! I first heard Jim Tung of The MathWorks say this about 25 years ago and it is still just as true. The model is where the rubber meets the road. And so it is also true that, ultimately, transparency hinges on the model.
Why do we care?
Backing up a bit, why do we care if data models are transparent, i.e. explainable or interpretable? In some low-impact low-risk cases (e.g. a recommender system for movies) it really doesn’t matter. But in a growing number of applications machine learning models are used to control systems that affect people’s heath and safety. In order to trust these systems, we need to understand how they work.
So what must one do to produce a transparent model? In data modeling ‘black-box’ is defined as ‘having no access to or understanding of the logic which the model uses to produce results.’ Open-source has nothing to do with the transparency or lack thereof in ‘black-box’ models. It is a requirement of course that for transparency you need to have access to the numerical recipe that constitues the model itself, i.e. the procedure by which a model takes an input and creates an output. This is a necessary condition of transparency. But it doesn’t matter if you have access to the source code that generated or implements e.g. a deep Artificial Neural Network (ANN) model if you don’t actually understand how it is making its predictions. The model is still black as ink.
This is the crux.
Getting to Transparency
The first step to creating transparent models is to be clear about what data went into them (and what data didn’t). Data models are first and foremost a reflection of the data upon which they are based. Calibration data sets should be made available and the logic which was used to include or exclude data should be documented. Sometimes we’re fortunate enough to be able use experimental design to create or augment calibration data sets. In these cases, what factors were considered? This gives a good indication of the domain where the model would be expected to work, what special cases or variations it may or may not work with, and what biases may have been built in.
The most obvious road to transparency includes the use interpretable models to begin with. We’ve been fans of linear factor-based methods like Partial Least Squares (PLS) since day one because of this. It is relatively easy to see how these models make their predictions. There is also a good understanding of the data preprocessing steps commonly used with them. Linear models can be successfully extended to non-linear problems by breaking up the domain. Locally Weighted Regression (LWR) is one example for quantitative problems while Automated Hierarchical Models (AHIMBU) is an example for qualitative (classification) models. In both cases interpretable PLS models are used locally and the logic by which they are constructed and applied is clear.
With complex non-linear models, e.g. multi-layer ANNs, Support Vector Machines (SVMs), and Boosted Regression Trees (XGBoost), it is much more difficult to create transparency as it has to be done post-facto. For instance, the Local Interpretable Model-Agnostic Explanations (LIME) method perturbs samples around the data point of interest to create a locally weighted linear model. This local model is then interpreted (which begs the question ‘why not use LWR to begin with?’). In a somewhat similar vein, Shapley values indicate the effect of inclusion of a variable on the prediction of a given sample. The sum of these estimated effects (plus the model offset) equal the prediction for the sample. Both LIME and SHAP explain the local behavior of the model around specific data points.
It is also possible to explore the global behavior of models using perturbation and sensitivity tests as a function of input values. Likewise, visualizations such as the one below can be created that give insight into the behavior of complex models, in this case an XGBoost model for classification.
XGBoost Decision Surface Visualization
To summarize, transparency, aka “Explainable AI” is all about understanding how models, that are the outputs of Machine Learning software, behave. Transparency can either be built-in or achieved through interrogation of the models.
Transparency at Eigenvector
Users of our PLS_Toolbox have always had access to its source code, including both the code used to identify models and the code used to apply models to new data, along with the model parameters themselves. It is not “open-source” in the sense that you can “freely copy and redistribute it,” (you can’t) but it is in the sense that you can see how it works, which in the context of transparent models is the critical aspect. Our Solo users don’t have the source code, but we’re happy to tell you exactly how something is computed. In fact we have people on staff who’s job it is to do this. And our Model_Exporter software can create numerical recipes of our models that make them fully open and transportable. So with regards to being able to look inside the computations involved with developing or applying models we have you covered.
ANN Model for Fat in Meat showing Shapley Values and Model Sensitivity Test from PLS_Toolbox 9.3.
A Final Word
There are a lot of reasons why one might care about model transparency. We like transparency because it increases the level of trust we have in our models. In our world of chemical data science/chemometrics we generally want to assure that models are making their predictions based on the chemistry, not some spurious correlation. We might also want to know what happens outside the boundary of the calibration data. To that end we recommend in our courses and consulting projects that modeling always begin with linear models as they are much more informative, you (the human on the other side of the screen) stand a good chance of actually learning something about the problem at hand. Our sense is that black-box models are currently way over-used. That’s the part of the AI/DL hype cycle we are in. I agree with the sentiments expressed in Why are we using Black-Box models in AI When we Don’t Need to? and it includes a very interesting example. Clearly, we are going to see continued work on explainable/interpretable machine learning because it will be demanded by those that are impacted by the model responses. And rightly so!
I’ll address this by starting with what stays the same. We use the same notes, and have the same instructors. For the most part we use the same data sets and go over the same examples hands-on with our PLS_Toolbox/Solo software. And we spend approximately the same amount of time on each topic.
So what’s the difference? At the risk of sounding condescending, the main difference is that you attend our in-person classes in person and with the online options you don’t. But the in-person aspect brings many unique possibilities with it. The most important one is the student-teacher interaction. In a live class the instructors can answer questions in real time, and also observe the students and sense if they are “getting it” or not. For the hands-on parts, we also usually have one or more additional instructors walking around behind people to see what’s on their screen to check that they are following and assist if necessary.
With in-person classes you also have the time at breaks and after class to talk face to face, like our Manny Palacios at left below discussing aspects of regression with an EigenU attendee. Plus there are also opportunities to interact, network and socialize with fellow attendees. Beyond this, in-person classes force you to set aside time, focus on learning and take advantage of the immersive environment.
The downside of in-person classes? Time and expense. You have to block the time off for the class plus the travel time. Unless the courses are nearby there are travel and lodging expenses and the courses themselves cost more.
For remote learning, we offer classes both live online and recorded online. The live online classes are at scheduled times, generally early morning in North America and late afternoon in Europe. Students can ask questions that the instructor can answer as part of the lecture or his assistants can answer through online chat. Plus we record these so students can review them later, which is especially helpful with the hands-on exercises. But of course the student-teacher interaction doesn’t match what possible in person, and there is not much interaction between students. And like in-person classes, you might not find the class you want in the time frame you need it. Online classes are, however, much less expensive and can be done from the comfort of your home or office, like the guy on the right above (whose desk hasn’t been this clean since the picture was taken 3 years ago).
Finally, there are recorded online classes. The main advantage of these is that they can be done completely on your own schedule, are available on demand, and like live online classes, are less expensive and don’t require travel. They do, however, put more distance between instructors and students as they become separated in both space and time! Questions are answered via email but not in real time.
In-person vs. Live Online vs. Recorded Online Pros and Cons
Pros
Cons
In-person Classes
Best student-teacher interaction
Most expensive
Real-time answers to questions
Added travel time
Interactions with fellow students
Time away from office
Forces focus on learning
After class & social opportunities
Live Online Classes
Real time answers to questions
Less student-teacher interaction
Access to recordings for review
No interaction with fellow students
Forces setting time aside
No after class & social
Less expensive
No travel required
Recorded Online Classes
Learn at your own pace
Least amount of student-teacher interaction
Available on demand
No interaction with fellow students
Less expensive
No after class & social
No travel required
Easy to put off
So what to choose? I’ve listed the pros and cons as I see them in the table above. Different people learn differently, so what’s a pro to one may be a con to another. For some, like me, attending something live forces me to set aside time then pay attention, e.g. turn off my phone and email, until it’s over. Some people don’t consider this an advantage!
That said, IMHO, if you can afford the time and expense and can work them into your schedule, in-person short courses are the best way to get started with a new subject in the shortest amount of time. They are the Cadillac (or in my case the Lincoln) way to learn. Between live online and recorded online I’d choose live if you can find the right course at the right time. But if you need it and have to have it right now, you can’t beat recorded online for on demand convenience.
I was asked the other day to provide a list of advantages that our Solo software has over its competitors for chemometrics and machine learning. Well, I don’t spend much time keeping track of what’s in other companies’ software. But I can tell you what is in Solo and why we think it’s a great value. (Apologies in advance for all the acronyms but I’ve included a guide below.)
Solo supports a very wide array of methods for data exploration, regression and classification. The standard PCA, PCR, PLS, MLR and CLS are included of course, but also MCR, SIMPLISMA/Purity, Robust PCA and PLS, O-PLS, PLS-DA, Gray CLS, PARAFAC, PARAFAC2, MPCA, batch data digester, batch maturity, LWR, ANN, Deep Learning ANNs, SVM, N-PLS, XGBoost, KNN, Logistic Regression, user specified Hierarchical Models, Automated Hierarchical Models, SIMCA, UMAP and t-SNE.
Solo has a sophisticated point and click user interface (see below) for graphical data editing and model building which users find very intuitive. Plots are easily customizable with class sets, color by properties, etc.
Solo supports the most extensive set of data preprocessing methods: centering, scaling, smoothing, derivatives, automatic WLS baseline, selected points baseline, Whitaker filter, EEM filtering, MSC, EMSC, detrend, EPO, GLSW, despiking, Gap-Segment derivatives, OSC, normalization, PQN, SNV, block scaling, class centering, Pareto and Poisson plus build your own math functions. User selectable preprocessing order and looping for more flexibility.
Solo does full cross-validation including all preprocessing steps, and includes a variety of cross validation methods as well as customizable specified data splits.
Solo supports a host of methods for variable selection: i-PLS, VIP, SR, Genetic Algorithms, r-PLS and stepwise.
Solo includes a full suite of methods for calibration transfer: DS, PDS, DW-PDS, SST, OSC, plus they can be inserted before or between preprocessing steps as required with the Model Centric Calibration Transfer Tool (MCCT).
Solo includes the Model Optimizer which can be used to create, calculate, compare and rank linear and non-linear models with a variety of preprocessing and selected variables
Solo is available on all three major platforms: Windows, MacOS and Linux. Models and data files are completely compatible between platforms.
Solo includes the Report Writer which makes models easy to document by creating PowerPoint or web pages from models. Solo maintains data history and includes model caching for preserving traceability.
Solo optional add-ons include Model_Exporter which allows users to export models as numerical recipes as well as Python and MATLAB code so they can be applied online and in handheld devices. Solo also works with Solo_Predictor, a stand-alone, configurable, prediction engine for online use. The MIA_Toolbox add-on allows users to seamlessly apply all the methods above to hyperspectral images.
Solo is completely compatible with PLS_Toolbox for use with MATLAB. PLS_Toolbox has all the features of Solo, including the point and click interfaces and graphical data editing, but also allows users to access all the functionality from the command line and incorporate these methods in user specified scripts and functions for ultimate flexibility. This allows users to work the way they want (command line or point and click) and still work together.
Solo gives users access to the Eigenvector HELPDESK, user support that is prompt and actually helpful. HELPDESK is manned by our developer staff, the people that actually write the software. Need more help on specific applications? Eigenvector offers consulting services.
Despite all of its advantages, Solo costs less than other major chemometric packages. We publish our price list so you can compare. We offer single-user and floating licenses that work great for small or large groups.
The Lake Chelan Valley Scholarship Fund (LCVSF) received 48 applications for 2023 and has awarded 22 scholarships, each in the amount of $2500. Awards went to 7 previous award winners now in their second, third or fourth year of study, as well as 10 graduates from Chelan High School (CHS) and 5 graduates from Manson High School (MHS) class of 2023.
CHS 2023 graduate awardees are Ryan Allen, Macie Cowan, Melina Cruz-Magallon, Irene Hernandez, Arden Paglia, Kaylee Patino, Kimberly Pineda, Mariana Sanchez-Mendoza, Tate Sandoval and Lauren Ware. MHS 2023 graduate awardees are Lissett Hernandez, Mackenzie Marble, Briseida Mendez-Resendez, Jude Petersen and Alondra Serrato Bailon. Renewals include Quinn Stamps and Casey Simpson (4th award for each), Cody Fitzpatrick (3rd award), Savannah Gresham, Odaliz Ordaz, Titus Peterson and Zoee Stamps (2nd award for each).
LCVSF board president Betsy Kronschnabel observed “It’s great to see renewals from students that are succeeding at their chosen school and are receiving awards for multiple years. Unlike many other scholarships, the LCVSF awards help students throughout their undergraduate education.”
The LCVSF was made possible by Dr. Doug and Eva Dewar (shown below), who wished that their estates be used to help the children of the Chelan Valley. Though they had no children of their own, they loved kids and helped many young people throughout their lives. The Dewars wished to enable motivated, well rounded students to further their education, and hoped that these students would return to the Chelan Valley. LCVSF was founded in 1991, and in that year five scholarships in the amount of $1000 each were awarded. The fund has grown substantially over the years from contributions from many people, but especially significant contributions from John Gladney, Ray Bumgardner, Don & Betty Schmitten, Marion McFadden, Virginia Husted, the Dick Slaugenhaupt Memorial and Irma Keeney. Now in its 33rd year, the $55,000 awarded this year brings the total to nearly $900,000 to Chelan Valley students since its inception.
LCVSF accepts applications from residents of the Chelan valley for undergraduate education. The awards are renewable for up to four years. LCVSF welcomes applications from graduating high school seniors as well as current college students and adults returning to school.
The LCVSF board includes Betsy Kronschnabel (President), Arthur Campbell, III, Linda Mayer (Secretary), Sue Clouse, Barry M. Wise, Ph.D. and John Pleyte, M.D. (Treasurer). For further information, please contact Barry Wise at bmw@eigenvector.com.
Eigenvector Research produces a variety of software products for chemometrics and machine learning and we often get asked how they work together. Here’s the roadmap!
We have two main packages for modeling, our MATLAB® based PLS_Toolbox, and our stand alone Solo (with versions for Windows, macOS and Linux). Solo is the compiled version of PLS_Toolbox, so in practice they are nearly identical, the difference being that if you are using PLS_Toolbox under MATLAB you also have access to the command line versions of all the functionality. PLS_Toolbox and Solo are highly interfaced point and click programs and include PCA, PLS, PCR, MCR, ANNs, SVMs, PARAFAC, MPCA, PLS-DA, SIMCA, kNN, etc., a very wide array of preprocessing methods, and also tools for specific tasks such as calibration transfer/instrument standardization.
Eigenvector Software for Chemometrics and Machine Learning
MATLAB®-based
Stand-alone
General Modeling & Analysis
PLS_Toolbox Our flagship product 30 years in the making. Point and click or command line access to the widest array of chemical data science tools and methods.
Solo The stand-alone version of PLS_Toolbox, available for Windows, macOS and Linux. Point and click chemometrics and machine learning.
Hyperspectral & Multivariate Image Analysis
MIA_Toolbox Add-on to PLS_Toolbox, allows seamless use of modeling methods on hyperspectral data plus additional image analysis tools.
Solo+MIA The stand-alone version of PLS_Toolbox plus MIA_Toolbox. Point and click modeling of hyperspectral images.
Model Export for Online Predictions
Model_Exporter Add-on to PLS_Toolbox, turns models into numerical recipes or code for application in other software or platforms.
Solo_Predictor Full featured online prediction engine applies any PLS_Toolbox or Solo model to new data, MIA_Toolbox compatible.
If you are doing hyperspectral imaging then you can add our MIA_Toolbox to PLS_Toolbox, or choose Solo+MIA. This allows use of all the above methods directly on hyperspectral images plus adds a few more image specific tools.
If you want to automate model application (say you want to get a PLS model online and have it make predictions as new data comes in) there are two main routes. Solo_Predictor is a full featured stand alone prediction engine that can apply any model you make in PLS_Toolbox/Solo and there are a variety of ways to communicate with it, the most common being socket connections. Solo_Predictor is compatible with hardware from many of our Technology Partner spectrometer companies.
Model_Exporter, on the other hand, creates numerical recipes and code required to apply models to new data streams in a variety of languages including MATLAB and Python. These recipes can then be compiled into other programs or run on hand held devices (such as ThermoFisher’s TruScan or Si-Ware’s NeoSpectra).
So what software should you buy? PLS_Toolbox or Solo?
Buy PLS_Toolbox if you …
— already have access to MATLAB, it’s less expensive than Solo
— know you want to automate pieces of your modeling process
— want to customize plots using MATLAB
— want to access additional functionality from other MATLAB toolboxes
Buy Solo if you …
— want to work only within visual interfaces
— don’t need to script or program
— prefer the lower cost of Solo compared to MATLAB + PLS_Toolbox
In 1988 Donald Dahlberg, Professor of Chemistry at Lebanon Valley College (LVC), decided to take a sabbatical leave at the University of Washington (UW) Center for Process Analytical Chemistry (CPAC). At the time, his former student Mary Beth Seasholtz was a second year graduate student in Bruce Kowalski’s Laboratory for Chemometrics. Mary Beth asked Don if he’d be interested in seeing what she was doing. Before Don knew it, he was attending Kowalski’s chemometrics courses and group meetings. I met Don during this period as I was also at CPAC.
When he returned to LVC he started teaching chemometrics to undergraduate students, and involving them in research. This included collaborative research with a local confectionary company.
Meanwhile, at Eigenvector we were interested in developing chemometrics courses for a wider audience. So sometime in 2001 our Neal B. Gallagher contacted Don about the possibility of creating a chemometrics workshop that did not involve the parallel presentation of matrix algebra. They struggled over a title, but eventually settled on “Chemometics without Equations (or hardly any).” We call it CWE for short. Don, having recently retired from teaching at LVC, wrote the workshop with Neal reviewing the content.
A slide from Chemometrics without Equations explaining PCA in everyday terms.
Don and Neal first presented CWE at the 16th International Forum on Process Analytical Chemistry (IFPAC) in San Diego on January 21-22, 2002. The course was taught hands-on using PLS_Toolbox. CWE was repeated at CPAC’s Summer Institute that July and again at the Federation of Analytical Chemistry and Spectroscopy Societies (FACSS, now SCIX) conference in October 2022 in Ft. Lauderdale, FL. This marked the beginning of CWE’s 20 year run at fall conferences. It was repeated in 2003 at FACSS and in 2004 moved to the Eastern Analytical Symposium (EAS), its home through this year. The workshop has been offered every year, except in 2020 when COVID-19 prevented a physical conference.
EAS 2022 marks Don’s final presentation of the course at EAS, making a total of 20 fall conference appearances. Each time Don has been assisted by either Neal or myself. Knowing that Don was an avid bourbon connoisseur we commemorated the occasion with a bottle of Blanton’s as he completed his final class.
Neal, Don and Barry celebrating Don’s final Chemometrics without Equations course at Eastern Analytical Symposium.
Looking back on 20 years of teaching CWE Don observed:
EAS has allowed me to meet many scientist who wish to learn and use chemometrics. They have included not only scientists in chemistry, but also those in related fields such as forensic science and cultural heritage. I have had the privilege to offer special versions of the course, tailored to the latter two fields. I have been able to present the course at John Jay College of Criminal Justice, the Forensic Science Department at the University of New Haven, the Museum of Modern Art, the Getty Museum and the Library of Congress. My goal has been to introduce the power of chemometrics to those inside and outside of analytical chemistry. Even though it is time to end my presentations at EAS, I intend to continue to help those who wish to explore the field of chemometrics.
Over 20+ years Professor Dahlberg has gently introduced hundreds to the field of chemometrics with CWE taught at conferences, at in-house classes and online. Thanks Don for your service to field! Cheers and bottoms up!
If you’d like to have Chemometrics without Equations presented at your site, please write bmw@eigenvector.com and we’ll help you arrange it.
We're transitioning to a new login system. You will be required to reset your password the first time you log into the new system. Please contact us at helpdesk@eigenvector.com if you have any trouble accessing your account.
SEARCH
