| Author | Message | ||
     Tony Davies (Td) |
Hello JB In this paper: "Near infrared networking� the ultimate control" [Proc NIR95 page 479-483] Nils Bo Büchmann described the use of 2,000 samples. That was in 1995, I expect that 10 years later the number is higher! Many years ago, I had an ANN add-on program for Unscrambler (the last DOS version). I loved to watch it run. Every now and again it would produce a calibration which fitted the test set very well. The problem is that you need a truely indepenent set that you only use once. If it gives a good answer then your ANN may have discovered a good calibration but in most cases it will not, so you forget the rule about only using it once and continue until you get a good result. You are fooling yourself! The test set is no longer independent. My advice is to listen to Pierre; he has more experince than anyone in the NIR community of every type of calibration method. Best wishes Tony | ||
| Pierre Dardenne (Dardenne) |
Hi, Training an ANN with 100 samples is quite dangerous indeed. Depends also on what are the inputs: raw data, selected variables, few PC's? If the respond is non-linear, SVM could be used instead of ANN. SVM can handle non-linearities and is less sensitive than ANN for overfitting when the number of samples is small. ANN and SVM need of course 3 sets: cal, stop and test. With so few samples a procedure as LOOCV must be applied. Pierre | ||
| JB Sirven |
Hello Tony, Effectively I do not work with so many samples. I rather have 50-100 samples for training my network. But I tested PLS and neural networks on the same dataset and neural networks provided much better predictions. Yet what you state makes me wonder. Have you got any references about the effect of the number of samples on neural networks performances ? What do you exactly mean by "being fooled by ANNs" ? Jean-Baptiste | ||
| hlmark |
JB - "Applied Regression Analysis" by Draper and Smith, has recommendations on how to handle non-linearities in the data, in both the X and Y variables. The simplest approach is, if you know the functional form of the non-linearity (from theory, say), to linearize the data using the appropriate function. If you don't know the functional form there are other, empirical, methods. As I mentioned, these algorithms have been applied to virtually ALL kinds of applications. \o/ /_\ | ||
| Tony Davies (Td) |
Hello Jean-Baptiste, The majority of NIR data is non-linear and are handled successfully by PLS. How many samples do you have? The experience from NIR is that you need a lot of samples > 1000 before ANNs will give reliable results. It is VERY easy to be fooled by ANNs (but good fun). Best wishes, Tony | ||
| JB Sirven |
Hello hlmark, For a simple reason : the relation between input and output variables is non linear. MLR, PCR and PLS are useful if you have a linear relationship between variables. Even if variants of PLS exist, which somehow model non linear relationships, neural networks are more powerful in this area. Jean-Baptiste Sirven. | ||
| hlmark |
JB - why not use one of the more standard algorithms for quantitative analysis: MLR, PCR, PLS? They've worked very well for a long time on a wide variety of applications. \o/ /_\ | ||
| hlmark |
JB - why not use one of the more standard algorithms for quantitative analysis: MLR, PCR, PLS? \o/ /_\ | ||
| JB Sirven |
Hello everybody, I am currently working on analytical spectroscopy in the UV-VIS range. I recently started analyzing my spectra through neural networks. I used a classical 3-layer network. Supervised learning provided very good results in terms of accuracy, precision and limit of detection in my area of application. My question is a rather general one : I know that a lot of neural networks types exist, mostly for qualitative purposes (i.e. identification/classification). For quantitative analyzes, basically for measuring a concentration from a spectrum, what other types of networks can I use instead of the multilayer perceptron, in order to still improve the performances ? Thank you for your answer ! Best regards, Jean-Baptiste SIRVEN. |