When you write the application for foundations, you write your research as promising as you can, to maximize the probability to get the funding. Now when I have got it and spent it. It is time to reflect whether I have done planned tings.

When designing machine noses the heat mapping is a helpful tool

Technical challenges

Since last spring, we have greatly progressed on technical issues. Back then, we struggled with contaminating the measurement instrument. It was contaminated after few samples, and we had to clean it repeatedly to get the thing work. For success, we needed hundreds of proper measurements in a row, so the needed improvement was quite significant.

To manage the contamination, first we tried the common medicine in artificial olfactomics: heating. It helped, but not as much as we needed. Then after studying the subject deeper, we got to the field of aerosol technology. There we got a good idea with help of aerosol researcher Dr. Sampo Saari. Actually, it turned to be a little too good idea. I mean economically promising. Economically interesting ideas always slows the research, since it is not possible to ignore the patenting options et cetera.

During these studies, we got one good study on smoke particles, and issue related to people’s health. (These results will be published soon.) After the episode with contamination and smoke, our system worked a far better, but some time had gone as well. Now we could measure samples pretty much, as we planned in the research plan.

Machine learning

After the machine learning has been applicable for sensor technology, the paradigm of many automated measurement has been changed quite significantly. Earlier, engineers needed to know every function inside the machine. Now with machine learning it is enough that you design proper hardware, so that, the machine gets repeatable and rational input, containing the critical information. Then, with a proper teaching methods machine really learns how to interpret result correctly. This seems to be especially effective in fields where you have complex measurement results, but you need only yes, or no, or a small amount of possible final answers. In long run, it is amazing to think how machines might work in the future, could they reach understanding? Could they choose between own and common good?

The next thig in our research was to tackle the sample consumption. Since the machine learning and the principle of our measurements, we could not exactly say how the most important information is spread to the measurement spectrum. That is why it was beneficial to measure with higher details and resolution than needed, and later we could optimize and decrease the spectrum resolution. We tested this improved technology, with test samples, and it seemed to work fine. The sample consumption was decreased at least hundred times, but experiments with valuable samples and meaningful for our society are still on the way.

So did I reach the plans in the research plan? Yes and no. We could greatly improve the system, we got good meaningful results and the research is going on fine on significant subject. Also no, there should be more public results available. Combining an economically driven and common good driven research is challenging. I’m just thinking that if on some day, machines would struggle on moral issues, should we give them a warm and sarcastic welcome party, or encourage them not to give up if they fall.

Markus Karjalainen is a doctoral student in Tampere University of Technology. He is doing research in gas spectroscopy and artificial olfactomics. He got annual TES grant for making the doctoral thesis.