AI Hardware and Software Can't Measure Without Learning and Re-Learning
Updated: Apr 19, 2022
To educate the Court as to the "why" of the unreliability of an approved instrument to measure across the measuring interval if it has not been recently calibrated, inspected, maintained, re-calibrated.
It seems strange to think that a modern "approved instrument" has to "learn" - that it has a kind of artificial intelligence. In reality it's true, an approved instrument has to learn the meanings of 67 mg/100mls, 94 mg/100mls, 117 mg/100mls, and 184 mg/100mls. The artificial intelligence uses extrapolation from its response during calibration to known values (from a NIST reference - un étalon with an unbroken chain of connection to SI units of the CGPM/BIPM in Paris) (maybe 0, 50, 100, 200, 300 or 0, 40, 80, 100, 300) that it is shown. This "calibration" is done by the calibrating or re-calibrating technician at the factory or Canadian Authorized Service Centre. A Certificate of Calibration is then prepared. This learning should never take place in the hands of the police. It should be done by a neutral entity that knows what it is doing and follows Standard Operating Procedures.
Here is a sample cross-examination of a Crown expert. This cross-examination needs improvement.
Q. Assuming that a new Intoxilyzer 8000 or 8000C comes off the assembly line... A. Yes.
Q. ...at C.M.I. Inc., in Kentucky. The instrument’s come off the assembly line, it’s not yet been calibrated. The instrument doesn’t – at that point in time – know how to relate various readings of 70 milligrams per 100 mils, or 100 milligrams per 100 mils or 130 milligrams per 100 mils, or 150 milligrams per 100 mils, against the measurement standard. Right? A. That’s correct. It’s actually looking at voltages as opposed to concentrations. The concentration of the standard is known and it produces a voltage change in the instrument that is then related to a blood alcohol concentration. Q. But the new instrument that comes off the line. The hardware’s been built. The initial software has been loaded into the instrument. The instrument needs to learn what 70 milligrams per 100 mils looks like; what 100 looks like; what 130 looks like; what 150 looks like. A. Correct. Q. And that’s done through a process called calibration. A. Yes. Q. And calibration is the process that we talked endlessly about earlier, having to do with its definition. Right? A. Yes.
Q. Now, without that learning process, of learning what 70 looks like, what 100 looks like, what 130 looks like and so on, the instrument can’t produce a reliable measurement result. A. Correct, until it’s been calibrated. Q. And from C.M.I.’s perspective, that measurement result needs to be possible across a measuring interval. A measurement range. A. Yes, across the range that the instrument’s capable of analysing for. Blood alcohol concentration. Q. Now, the Intoxilyzer 8000 or 8000C, the manufacturer advertises in its brochure that it is capable of a range of – and maybe you can help me with this – from – I’m reading on the second page of the brochure – is that 0-600 milligrams per 100 mils, the equivalent? A. That’s the equivalent units, yes.
Q. As I see it. It’s actually typed as 0.00 to 0.600 percent B.A.C. in grams per 210 litres, but for us that translates into 0 to 600 milligrams per 100 mils.
A. Correct. Q. So, the calibration needs to take place right across that measuring interval? A. You would have to ask C.M.I. what their procedure is for the calibration of the device across the entire range. I do not know that. Q. But if the instrument is going to be scientifically reliable across the whole range, then one is going to need to do that learning process at each of the values between 0 and 600 milligrams per 100 mils. A. I don’t do calibration so I can’t speak to what the procedure is. You would have to ask C.M.I. or somebody who’s trained to do that.
[If CFS scientists don't understand how it is that an approved instrument is scientifically reliable across its measuring interval then how can any CFS scientist be qualified/capable of giving scientific opinion about what is and is not a scientifically reliable measurement on an approved instrument?]
[Notice the reference to "somebody who's trained to do that" above. The answer implies that the only persons who can give opinion evidence about this stuff are technical experts - those who point to technical manuals as their source of knowledge - see the Science Manual for Canadian Judges. There are no scientists available to the Court from the CFS or ATC who can give scientific opinion about the reliability of the measurement result throughout the measuring interval.]
MR. BISS: Q. So, the instrument has to learn each of these values. It has to learn them so that the instrument has an accuracy of plus or minus 3 percent or 3 milligrams per 100 mils right across its measuring range, or its measuring interval, so at 50, at 70, at 100, at 130, 150 and so on. A. Correct.
[This question as asked, does not reflect a current scientific concept of "accuracy" (see the VIM vocabulary) but it does reflect the manufacturer's concept of "accuracy" in its manufacturer's specifications. It also reflects what the CFS/ATC scientist will admit. Take a look at Hodgson's definition of "accuracy" and the VIM definition of accuracy. "Accuracy" is really something we aim towards but in reality we can only achieve an average of values (a technique to promote accuracy) that falls within uncertainty of measurement.] Q. Right? Now, in order to complete that
learning process the people at C.M.I. in their lab would run a program on the instrument which permits the creation of something called the calibration curve. I think you used the word calibration curve much earlier today. A. Yes. Q. And the calibration curve is, whether it’s done on paper in some scientific experimental processes or whether it’s done electronically, the point is, there’s a relationship between the measured result and the electrical signals that are coming off the instrument so that the instrument is reliable at each of those values of 70, 100, 130, 150 and so on.
A. Again, across the entire range that it’s measuring. Q. Yes. And so for purposes of the creation of that calibration curve, what the lab people at C.M.I. would do would be to show the instrument values at various – they would use measurement standards of various values. So the measurement standards might be 50, might be 100, might be 300, might be 0, that the person in the lab would show to the instrument so that the instrument can learn what those values look like and create its calibration curve. A. Again, in general terms, that’s how a calibration curve works when you’re calibrating an instrument of some kind. Again, I don’t know the specific procedures that C.M.I. follows, though. Q. Right. Well, let’s look at the last exhibit which was Exhibit Number 25 in terms of procedure. And after specifications, it has calibration procedure for the Intoxilyzer 8000... A. Yes.
Q. ...in that document. And it indicates in
that procedure that the equipment that you would need to do a calibration would be first of all, a wet bath solution. I think that just means a – a wet bath simulator solution. A. Well, simulator using an alcohol – a liquid alcohol standard as opposed to a gas cylinder... Q. Yes. A. ...standard. Q. So they would use a wet bath solution of zero, which essentially is water. A. Yes. Q. They would use a wet bath solution at 40 milligrams per 100 mils. MR. BISS: Your Honour, do you have the same page? THE COURT: Yeah, I’ve got it. MR. BISS: Okay. Q. They would use a wet bath solution at 80 milligrams per 100 mils. A. Yes. Q. And 100 milligrams per 100 mils. A. Yes. Q. And at 300 milligrams per 100 mils. A. Correct. Q. And the procedure would involve using that variety of solutions and the reason why they need to use that variety of solutions is because the instrument has to learn what each of those values means. A. Correct. Q. And that creates a calibration curve. A. Well, it creates the data that the instrument uses to generate a calibration curve.
It is extremely important that defence lawyers learn these basic concepts. CFS/ATC scientists seem to gloss over them without recognizing how important they are to an understanding of what makes the measurement results on an approved instrument "reliable". The next step for a defence lawyer after this, is to understand how the instrument's ability to reliably measure, across the measuring interval, changes over time through a process called drift. See the Hodgson article referred to by the SCC in St. Onge. Unless inspection/maintenance (and perhaps re-calibration) takes place on a regular basis, the calibration curve originally learned by the instrument becomes inadequate (the curve is wrong) to fit the instrument's changed (drifted) response to the concentration in the sample chamber, at at least some values across the measuring interval.
See also "How to Calibrate - Teach an AI - Build the Inverse Logarithmic Calibration Curve"
#artificialintelligence #calibration #inspection #measuringinterval #annualmaintenance #NIST #BIPM #SIUnits #SOPs