Principle of Non-Specificity

[C. Coppock]

Principle of Non-Specificity [1]

In cases of recognition and in detailed formalized forensic comparisons such as with the ACE-V methodology, the exact same information is never used twice to effect individualization. Holistically derived inferential information cannot be run through a standardized formula, as it is unique in itself. This is a principle of: non-specificity within the cognitive comparative process. The use of specific (exact) information is not necessary, but rather the information must be relevant and sufficient for the purpose. The reason that the same exact information cannot be duplicated is due to the infinite (holistic) variability in where the relevant information can be found, and how those bits, portions, and combinations of information are used. It is the end result that matters. The sufficient information needed to sustain recognition (or formalized individualization) can possibly be found anywhere in the supporting information data set and cannot, due to the complex and unique cognitive process, be derived from specifically defined data. Therefore all specified data, must be generalized in nature.

Ref:
[1] Coppock, Craig: Complexity Of Recognition Inductive And Inferential Processes In Forensic Science.
2004, clpex.com
Recent update 08-24-2008 Available at Google Groups- [Forensics: Forensic Student Info]

Note: When we speak of a minutiae or Galton points, we are in fact being very general. There are many aspects, details, and relative relationships in a single Galton point we do not specify. It is also understood that each examiner will have a different understanding and experience base to apply to this single Galton point information. Therefore, we could consider a single point as a collection of information. It is the sum of its parts. Multiply this with the ACE-V methodology and we have a very complex and understandably unique cognitive process that repeatably yields the correct results with its proper application.

[C. Coppock]

It may help us understand our Forensic Comparison process with research in Measurement Theory (MT). MT deals with needed axioms and error to make imprecise measurements practical.

[Kasey Wertheim]

A very high-level onine reference exist at: http://www.capgo.com/Resources/Measurem ... heory.html

Introduction to measurement theory
Measurement is the process of associating numbers with physical quantities and phenomena. The process is accomplished through the comparison of a measured value with some known quantity (standard) of the same kind. The subject has become of vital importance in sciences, engineering and to much everyday activity.

While measurement theory began with the Greeks in the 4th century BC, the first useful work appeared in the 18th century by English mathematician Thomas Simpson on observation error - perhaps the most important single aspect of measurement theory.

Measurement error
Practically all measurement of continuums involve errors. Understanding the nature and source of these errors can help in reducing their impact and in may instances prevent the drawing of incorrect conclusions.

In earlier times it was thought that errors in measurement could be eliminated by improvements in technique and equipment, however most scientists now accept this is not the case. Today, nearly all scientific and engineering results are routinely reported with likely error bounds. The types of errors that must be understood include instrumental errors, systematic errors, random errors, sampling errors and indirect errors.

Systematic error
An error that can be predicted and hence eventually removed from data is a systematic error. Systematic errors may change with time, so it is important that sufficient reference data be collected with the data set to allow the systematic errors to be quantified and subtracted from the data set: [Instrumental errors, Sensor placement errors, Indirect errors]

Instrumental errors
Examples of measurement equipment systematic errors include calibration errors, input zero drift and gain drift. Measuring equipment can also induce nonsystematic errors. Instrument errors are considered in more detail in the Measurement Methods pages.

Sensor placement errors
An often overlooked systematic error source is associated with the location of the sensor. Errors can be caused by measured parameter gradients or the impact of other parameters on the sensor. For example, in precision air temperature measurement, it is likely that temperature gradients exist or radiant energy be heating the sensor - so just what is being measured? Also radiant heat may heat the sensor directly giving an erroneous reading.

Indirect errors
These are associated with calibration and conversions. Generally these errors are small, but can become significant with some types of measurement for example light intensity.

Nonsystematic errors
A nonsystematic error is one that cannot be predicted due to a randomness in its nature. Nonsystematic errors limit the ultimate accuracy of a measurement process by a masking effect that leads to information loss. They include: [Quantizing error, Rounding and truncation errors, Sampling errors, Random or noise errors, Sensor cross sensitivity errors]

Quantizing error
All measuring equipment has a resolution limit, input variations below which can not be detected or measured, leading to unrecoverable information loss. In systems with evenly spaced quantization boundaries, quantization errors can be reduced by adding noise to the input and averaging many samples.

Rounding and truncation errors
In processing the measuring system's readings, the precision of calculation (number of significant digits) can compromise results.

Sampling errors
The frequency and sample window time can impact accuracy, especially for changing or noisy quantity.

Random or noise errors
Random noise is always present in measurement, and sometimes is the dominant source of error. Depending on the noise spectrum, the noise error can generally be reduced by averaging many readings.

Sensor cross sensitivity errors
Few measuring systems respond only to the parameter being measured. All sensors have a degree of sensitivity to other parameters. For example a temperature sensor's output may change with pressure, humidity and/or ionizing radiation.

[Gerald Clough]

With reference to Kasey's enumeration of errors:

I seems to me that the source of many errors made in good faith is the error of misinterpreting any part of an image. This subjective interpretation must be made if that feature is to become useful in forming an objective conclusion. The image feature may represent one of three realities, a plain continuous ridge segment, an artifact, and a genuine diversion from a plain continuous ridge. If a genuine diversion, it may be (1) a diversion or discontinuity, the precise nature of which cannot reasonably be guessed, (2) a diversion or discontinuity that can be described as exhibiting a certain behavior of ridge flow and continuity that cannot, however, be described with certainty at every point and cannot therefore be depicted by a system of schematic lines, (3) a diversion or discontinuity that can be described with considerable certainty at every point and depicted by a system of continuous and discontinuous schematic lines. There is, of course, another subjective judgment, that of whether or not some area of the image is so poorly represented that contains no data and can be recognized only as a lack of data at that point. This includes the areas where recorded impression ends in a blank or black area.

Leaving aside any argument about the value or weight of some general kind of genuine feature that cannot be precisely described, and assuming that the goal at this point is only to subjectively establish each part of the image according to the breakdown above, how do we classify the types of error that can appear at this stage? Aside from plain bad eyesight, simply a poor sensor requiring frank badly founded guesses, I seem to keep thinking of errors as (1) an incorrect guess that assigns characteristics to a portion of the image so poorly represented that it contains no data, (2) an incorrect guess that inaccurately describes the precise nature of a feature when the quality of the impression can support only a general description, and (3) an incorrect guess that precisely but inaccurately describes a portion of the image that is represented sufficiently well that it could have been accurately represented. A gross example of (1) might be an interpretation of an ending ridge at the termination of a ridge impression at a blank area. An example of (2) might be assigning a description as a bifurcation diverting from the ridge to lying to the right when the image was only sufficient to determine that a ridge lying between two on either side either ended or diverted into the ridge to the left or right. An example of (3) might be assigning a bifurcation to one ridge when the image quality was sufficient to accurately describe the diverting ridge to be formed to the adjacent ridge on the other side. And I think we can most reasonably consider these errors to be often the result of unconscious biases, rather than conscious mental coin tosses.

To avoid confusion over the sort of error we are talking about, I rightly or wrongly omitted the case where we correctly guess the precise nature of a feature when the image quality cannot reasonably convey anything but the general nature.

Again, we are not concerned at this moment about whether or not some feature of general description is used in comparison, merely how it is interpreted during analysis. My question for Kasey and Craig is: Are these errors meaningful in terms of Measurement Theory, or are they best described in a less fundamental way in terms of recognition, since I think what we most often mean when we talk about biased interpretation is some untoward unconscious factor?

Or has my description become so convoluted as to confound the experts? (A worthy result in itself.)

[Kasey Wertheim]

I'm definitely no expert in Measurement Theory - I merely posted some additional relevant info for the benefit of those who might read on, but not specifically seek it out themselves. But there are several aspects of measurement theory that apply to LPE's.

For example, Dad and I have always mentioned in our training courses that the eye and subsequent human vision system is the capture mechanism through which the human brain is first engaged in the comparison process. If we consider this the "sensor" in MT, we could relate latent print examiner vision problems such as inadequately corrected eyesight or grayscale perception deficiencies to the MT model.

Likewise, even our critics say that a latent print examiner cannot be removed from the ACE-V method to assess error rate separately because we are "the instrument" that applies it. As an instrument, we have unique calibration that directly contributes to error. Our ability (training, experience, talent, motivation, daily variables) defines who we are and how we conduct our work. We accept that some instruments are more accurate than others. Some are faster than others. Hopefully the increase in speed is because of an increase in proficiency, not a tradeoff with accuracy... but we (hopefully) get better with experience.

Part of instrument calibration is testing borderline samples in an accurate fashion. MT relates the impact of "noisy" samples, and a lot of that could apply to latent prints. Think of the skin as the ground truth, and any impression of the 3-D surface as an opportunity to mask the true nature of the sample with the "noise" introduced during that particular rendering. Multiple samples can sometimes allow you to get a better understanding of the source and therefore the ground truth. When an extremely noisy sample (a latent print) is evaluated by an instrument (the examiner), the accuracy of the instrument becomes critical.

When the instrument isn't calibrated correctly under the MT model, this can lead to error. Training is what allows us to be more finely calibrated to deal with "noisy" borderline latent prints. If an examiner places incorrect weight on a feature in a "noisy" image as distortion when in fact it is different, it sets up the potential for an erroneous identification. If an examiner incorrectly classifies a distortion artifact as a difference, it sets up the potential for an erroneous exclusion. With training, an examiner is more able to accurately determine distortion from dissimilarity and becomes a better instrument for measuring similarity or dissimilarity in latent prints.

I haven't looked into enough detail on each of these MT topics to really do justice on how they could exactly help out our discipline. I read Craig's post, and having background and insight into human factors and error, I thought the relation was good enough to bring up some general principles to the board. I would encourage any reader who is interested (you included, Gerald) to pursue a deep-dive into one or two of these, use them to push the latent print discipline forward, and write it up for publication and peer review.

-Kasey

[Gerald Clough]

Thanks. I wonder, though, if it's just a mistake to chase a model of the human process as a technological system. It's attractive, because we take shots on account of "science" issues and because we feel a symbiosis with automation. I'll have to think on that. I sometimes sits and thinks - when I don't just sits.

[Pat A. Wertheim]

to chase a model of the human process as a technological system

I sits and I thinks maybe it's an analogy. There ain't no perfect analogy, but if it helps us understand a thing, it ain't all bad, either.

[Les Bush]

Firstly I think Craig can step into the mantle of being authoritative about the process of individualisation since the level of material he consistently produces is well researched/relevant and his interpretations well balanced.
[quote="C. Coppock"]Principle of Non-Specificity [1]

In cases of recognition and in detailed formalized forensic comparisons such as with the ACE-V methodology, the exact same information is never used twice to effect individualization. Holistically derived inferential information cannot be run through a standardized formula, as it is unique in itself. This is a principle of: non-specificity within the cognitive comparative process. The use of specific (exact) information is not necessary, but rather the information must be relevant and sufficient for the purpose. The reason that the same exact information cannot be duplicated is due to the infinite (holistic) variability in where the relevant information can be found, and how those bits, portions, and combinations of information are used. It is the end result that matters. The sufficient information needed to sustain recognition (or formalized individualization) can possibly be found anywhere in the supporting information data set and cannot, due to the complex and unique cognitive process, be derived from specifically defined data. Therefore all specified data, must generalized in nature.[/quote]
This information is very important to me as a fingerprint expert. We establish individualisation be a combination of similarity, sequence, spatial arrangement and sufficiency. By doing so we identify the souce of both the exemplar and latent prints. And our conclusions are definitive in that we form the belief that no other person/s have a duplication of the same pattern of fingerprint detail. We have arrived at this point by following a very pragmatic path based on personal experience and collective knowledge that no other expert has found two prints to be identical from two different persons. This thread appears to be about how we can associate the variables used in individualisation with some more meaningful scientific outcome of measurement.

[quote="Kasey Wertheim"]Introduction to measurement theory
Measurement is the process of associating numbers with physical quantities and phenomena. The process is accomplished through the comparison of a measured value with some known quantity (standard) of the same kind. The subject has become of vital importance in sciences, engineering and to much everyday activity.[/quote]

So where does that leave us with fingerprints? The scale of fingerprint detail is macroscopic but no-one in science has ever said measurements have to be on the large size, just that they are relevant and reproducible. So how can our individualisations produce quantitative data that can be validated with a known association to the source skin? If we look closer at the detail in much the same way that AFIS does there is an abundance of measurements to be gleaned by any pattern of ridge detail. The opportunity exists for every latent and exemplar prints to be surveyed/analysed for their geometric relationships that define the uniqueness of their pattern of detail. These same patterns will have correspondence to the source skin arrangement while variations in their (exemplar/latent) appearance/results can be interpreted/explained by the transfer conditions. To grasp this concept will require acceptance that the permanence and uniqueness of the source skin is exclusive to one person. To the courts our tabled results of measurements are an extension and a presentation of the detail used to conclude individualisation. To the science community the measurements of latent/exemplar patterns are our way of proving the analysis to a degree of relevance with a prediction that if the source skin was available it would validate the results. Without access to the source skin the results could be strongly inferred by using other exemplars (repeat recidivists) as verifications. Regards from oz and thanks Craig for an interesting thread. Les

[C. Coppock]

There are several points of our science that I would could focus some additional attention. Regarding proof of a hypothesis, we must realized that most scientific theories, hypothesis and even some mathematical statements are not provable beyond a practice nature. This ties us in with uncertainty, measurement theory, and its attention to error [and] axioms. Much needed axioms! We also have axioms, in that we don't need to compare all friction skin in the world each time we want to offer a hypothesis of Individualization or exclusion. I attempted to cover some of this information in the paper: Complexity of Recognition. I think we are on the right track, we just need to have a stronger tie-in with general science. After a brief discussion on probability theory and its implications, a CLPE stated that there is no need to understand probability theory within our Science of Fingerprints. Perhaps our our science gap is bigger than we would like.

Sincerely,

[C. Coppock]

Yes Gerald you're correct,

Our biggest challenge is to understand our methodology, it's limits (errors), and its needed axioms, utilizing such things as Measurment Theory as a relevant guide rather than a hard rule. Most of the issues are indeed relevant, just clouded within the complex human cognitive process. Non-Specificity is a core, yet we need to refine our axioms, and organize the errors within sets. Of course, many of our process errors are simply different versions of each other.

Interestingly, I have been doing some research in Artificial Intelligence. These AI folks have already been breaking down human cogative tasks. I see some areas that are relevant to our issues and our attempts to break the process down into understandable bits. ...some of which has been rolled up into a short paper on Uncertainty and the Development of a Inference Process Theory. I will post an draft version on my Google Group [Forensics: Forensic Student Info] as soon as I can get file transfer ability.