Natural Language Algorithms: Henry Faulds

Henry Faulds, in his book Dactylography; or, The study of finger-prints, published in 1912 had an interesting insight that arose from a clash between his alphabetic language roots and the ideographic language roots of the Asian countries where he resided. Faulds was the editor of a monthly magazine called The Chrysanthemum, devoted to Japanese topics of literary, scientific and antiquities. As a result, he spent time near the printing press where the type blocks were classified for easy access. Obviously, these type blocks were of Japanese characters that consist of ideographs, a written character symbolizing the idea of a thing without indicating the sounds used to say it. They were organized in such a way that similar characters would be grouped together by both the block and the font. Something like this:

Image Source: https://www.ebay.com/itm/224954773269

Faulds goes on to say, “However, each Chinese ideograph, for dictionary purposes, is supposed to be built up around an element called by western lexicographers its key or radical, and of these there are two hundred and twelve. You look for the radical in an unknown character, and then look for that radical in its serial place in the two hundred and twelve. It is a question then, as in finger-prints, of counting strokes, and if the strokes are alike in number in any two instances, of looking then as to how they are arranged. Two characters with the same number of pen-strokes under the same radical or key, may bear quite a different aspect.

A Chinese character is defined and limited, but a finger-print pattern often, or usually, trails off into indeterminate lineations of little value for classification purposes. Hence we seek in the latter to isolate for study the central part of the pattern, where the intricacy of the ramifications usually rises to a maximum’ [1]

While the description is in the context of being struck with the idea of creating a classification for filing and retrieving prints based on what he saw with the printing blocks, what he unwittingly did was provide the earliest and most descriptive form of fingerprint identification.

What is fingerprint identification?

The best way to investigate this is to go back to some of the earliest literature on the topic, before the knowledge was institutionalized. Before there is institutional knowledge, there are naturalist accounts of science that iterate into a school of thought. Accounts of methods and processes can be less formal and philosophical underpinnings can be lacking, but the natural language descriptors are deeply insightful and there is no real jargon yet invented so parsing meaning can be relatively straight forward.

According to the authors of early texts , we see that an identification can be described as:

‘...definitions by means of which any finger impression may be accurately described, and explains, when two prints are brought under examination, the numerous points in each that should be compared with a view to establishing identity or dissimilarity’ [2]

This statement is pulled from Edward Henry, the man who implemented the first fingerprint classification system, in which a human algorithm [3] was applied to fingerprints in order to file and retrieve them systematically.?

Upon retrieval of the fingerprints, the points (characteristics) were compared. What does that mean?

According to Frederick Kuhne, author of The Finger Print Instructor, 1916, it meant the following:

‘After an impression has been classified, a search of the files is made by comparing the impression with those of the same classification, to determine whether the person to whom said prints belong had a previous finger print record. In order to properly make comparisons, it is necessary that the person making the search be familiar with the various points or characteristics appearing in finger impressions, numerous peculiarities of ridges being readily perceivable to the naked eye, such as the general contour of the pattern, ridge bifurcations (individual ridges branching out into two ridges), what is known as an island (this being caused by a ridge bifurcating but again joining into one), abrupt ending ridges, some peculiarity as to the formation of the delta or core, short ridge lines, etc.’ [4]

In the Question and Answer section of the book he reiterates and elaborates.

Q. What are characteristics in finger impressions?

A. They are peculiarities of the ridges, such as abrupt endings, bifurcations, the formation of what is termed an island, short ridge lines, ridge dots, some peculiarity as to the formation of the delta or core; in fact, any peculiarity out of the ordinary may be considered a characteristic point.

Q. What is the value of ridge characteristics in finger print work??

A. They serve as a positive means of identification.

Q. To what extent are these characteristics used?

A. They are used when making a search of the files, by comparing them with the impressions having the same classification; they are also used in the prosecution of criminal cases, where impressions are left in the commission of a crime. In such cases the impressions are enlarged and the characteristic points of both impressions are diagrammed to correspond with each other.? [5]

By paraphrasing what was written, we can say an identification is a precise form of classification that results in an equivalence relation between two impressions. It’s important to note that this is done with some specific goals in mind; ‘as a means of fixing identity’ , a ‘positive means of identification’, the ‘prosecution of criminal cases’, or in the instance of Edward Henry who was in charge of distributing payments, preventing ‘false personage’.? This is to say that there is a recognition that fingerprint identification has a role in the social construction of institutional rights and responsibilities. This makes it different from a merely descriptive account of the world in the way a natural science might.?

While the picture above shows us what is being looked at in the print and its relative location, it’s worth repeating Fauld's statement, because he’s describing the phenomenology of the comparison process.?

‘You look for the radical in an unknown character, and then look for that radical in its serial place in the two hundred and twelve. It is a question then, as in finger-prints, of counting strokes, and if the strokes are alike in number in any two instances, of looking then as to how they are arranged. Two characters with the same number of pen-strokes under the same radical or key, may bear quite a different aspect’ [6]

?While his original statement above was concerned more with the recognition of the fingerprint from a mere classification/retrieval context, it’s important to anchor on his statement. Paraphrasing, he says that the center of the character is looked at first (the radical), it is then looked at sequentially among other characters (in serial). He then finishes with a statement of examining the lines and counting (counting strokes) the similarities between them (strokes are alike) including their spatial arrangement (how they are arranged).??

He goes on to provide more details, this time specifically in the context of a comparison.

‘In tracking a criminal by a single impression made by a finger, the lineations in so small a space would require to have been clearly imprinted, and to have what many finger-print patterns have not, some notable or significant characteristics about it. Then, when enlarged by photography into a picture of some thirty inches, the measurements from fixed points in the pattern should correspond with those of the person in custody, on suspicion, and the curves should be shown to concur in all their sinuosities. But, in comparing two official imprints of the ten fingers properly and clearly impressed, there should be no difficulty, the points of comparison being overwhelming.’ [7]

Here we see a few concepts emerge when it comes to the comparison of two impressions: Clarity (‘clearly imprinted’), diversity of characteristics (‘notable or significant characteristics’), spatial agreement (‘measurements from fixed points in the pattern should correspond’), and formal agreement (‘the curves should be shown to concur in all their sinuosities’). He then goes on to make the first statement of verification that I could find. He switches gears and says, once the single impressions are shown to match, a comparison of all ten fingers would show overwhelming correspondence. It would seem that a ten print comparison, post single impression identification would be somewhat of a quality assurance measure.?

This book is but one of Fauld's works. In 1905, he laid out the first statement of sufficiency in his work ‘Guide to Fingerprint Identification’.? He said, ‘Complete coincidence of ten serial lineations is strong circumstantial evidence of identity’. [8] Although, he goes on later to criticize Scotland Yard for claiming that they can identify with four points in a print. [9]

Back to the question at hand, what is an identification and more importantly how does it work? An identification is a precise form of classification in which observations about the fingerprint draw it into an ever smaller class through recursion.? What does that mean?

It means that through definitions we create boundaries and those boundaries constitute classes of objects. As it pertains to fingerprints, if we were at a crime scene and we saw what looked like a finger impression on a car, we could ask, does this impression have any ridges visible? If the answer is yes, we could then place it in a class ‘latent fingerprint’. We could then go on to ask if the impression has, as Faulds articulates, clarity and a diversity of characteristics. If the answer is yes then we have a latent fingerprint with clarity and a diversity of characteristics. This forms the basis of a new class, ‘suitable for comparison’ and now we add the classes together in sequence. We have a latent fingerprint that is suitable for comparison. As we see in these two examples we went from the general impression to the more specific latent fingerprint and we went from the more specific clarity and diversity to the more general ‘suitable for comparison’. That is to say, these properties are symmetrical, a property we will investigate more thoroughly later.

Next we train our eyes on the data in the print. Even the properties of the print have been given their own classification. These are known as level 1, level 2 and level 3 data respectively and can be thought of as diversity of data along the ridge in aggregate ( ridge flow/pattern type - level 1), in path continuity (minutiae - level 2) and within or along the edge (pores and ridge edges -level 3) . Something Faulds himself articulated when he said ‘A single ridge is not even always of uniform thickness throughout its course. [10]

Given this information, as Faulds himself articulates, you look for the radical, or in the case of a latent impression an anchor point (usually a focal point like the core or delta if they exist). Once you find an anchor, you use it as a basis to venture out into making a map of minutiae relative to the anchor point or other previously mapped minutiae. Or as Faulds described, ‘counting strokes’.? The result is something that can be transformed into sentences that create relative class distinctions. In the case of analyzing a fingerprint, you locate the core and you count two ridges over to the right and see a bifurcation that opens down. The bifurcation is three ridges away from a short ridge that has one thin end and one thick end. Based on the description you’ve just created the following relative classes: has core, has bifurcation, opens down, three ridges away, short ridge, thin end. If you do that for all the data in the print you are considering, it adds up into complex patterns.? Faulds himself recognizes this when he states ‘the ridges are not only distinct, but gather up into complex patterns of much character and variety’ [11]

The way this complexity builds is through recursion. Simply stated recursion is just a way of defining a problem in terms of itself. It’s a function that calls itself. The most simple way to think of it is like a Russian Nesting Doll in which the function ‘doll’ adds to itself to create a bigger, nested doll which adds to itself to create an even bigger nested doll, etc. The function works both through addition (small fits into bigger doll) and through subtraction (big reveals smaller doll).

More formally, it looks like this if we were writing computer code:

Image Source: https://www.geeksforgeeks.org/recursive-functions/

So in our case, classification is the function and the output of the function feeds into the function to create a more complex output. This more complex output could then be the basis for recursion at a new level.

Let’s take a look at an example. We could create some relative classes based upon position, sequence, shape and direction. The classes ‘has bulb shape’ and ‘points down’ and ‘continues in a circular direction’ and ‘points up’ and ‘ends next to an edge that sticks out’ in addition to ‘is center’ all exist in this photo.?

Image Source: Author’s fingerprints

If I just said it in plain english, I would say in the center of the print there’s a ridge that points down, has a bulb shape and if you trace it around to where it ends you’ll trace a circular shape until you get to the end, which points up. The end that points up is next to an area that sticks out further than the uniform thickness of the ridge in a round shape and if you trace that ridge around, there are no breaks but at the bottom of the ridge it gets pointy on the outside and flat on the inside.

I could take input like this and feed it into artificial intelligence and have it define a recursive function. I did, and here’s what it produced:

minutiae = [

  Minutia("bulb", "points down", None, "center"),

  Minutia("fork", "continues in a circular direction", "points down", "right"),

  Minutia("end", "points up", "round", "above"),

  Minutia("ridge edge", None, "sticks out", "left")

]

for i in range(len(minutiae) - 1):

  minutiae[i].next = minutiae[i + 1]

fingerprint_description = describe_fingerprint(minutiae[0])        

Then we'd need to call the function:

print(fingerprint_description)        

The output would be:

In the center of the print, there's a bulb ridge that points down. Right from that, there's a fork ridge that continues in a circular direction from the bulb ridge and points down. Above that, there's an end ridge that points up and has a round shape. Left from that, there's a ridge edge that sticks out.

An explanation of the process:

1. The code creates a list of Minutia objects with properties like shape, direction, ending shape, and spatial position.
2. It links the minutiae together using the next attribute to create a sequential structure.
3. The describe_fingerprint function is called with the first minutia, and it recursively builds the description based on the minutiae's properties and relationships.
4. The output is a clear and natural-sounding description of the fingerprint's minutiae, incorporating their spatial, shape, direction, and sequential qualities.

The interesting thing is that we can then assign a truth value to each of the minutiae because these are propositions that can be disputed. So, if minutia 1 and all of the properties ascribed to it exist as in the above example, we could assign it a truth value, which is a classification in its own right. We could say minutiae [1] = true as an example.

The idea here is not to advocate for explicitly listing the properties of minutiae out extensively, that would be tedious. Rather it shows the structure behind the precise classification of features in a fingerprint from a mere psychological aspect. Regardless, identification is more of a visual computation than it is a linguistic one, so practically it doesn’t make sense. The mark up of a fingerprint usually happens visually and a more common convention is to just put a colored dot where you’re looking. If there are varying levels of clarity, then a color scheme that corresponds to a good, bad, ugly scale can be used. These modifiers can act as signifiers of a truth value of less than 1.

And unsurprisingly, while algorithms vary, if we survey the AFIS algorithm literature there is an overlap in a type of AFIS matching known that employs a structural method which first filters the fingerprint, much in the same way an Examiner analyzes a fingerprint. It uses a central minutiae then looks at the nearest neighbor minutiae and determines the type of neighboring minutiae, distance to the central minutiae, the relative angle, and ridge count. In this way, much as the way an Examiner views minutiae, the algorithm aggregates the data in meaningful clusters of properties. The difference being, humans have access to continuous data all along the ridge whereas an algorithm only finds relevance in discrete terms which it then builds in some tolerance rotationally and dimensionally.

Image source: https://www.researchgate.net/publication/220756612_Effectiveness_of_Assigning_Confidence_Levels_to_Classifiers_and_a_Novel_Feature_in_Fingerprint_Matching

This brings us back to Fauld's statement. Compare his statement to the? AFIS algorithm picture.

?‘You look for the radical in an unknown character, and then look for that radical in its serial place in the two hundred and twelve. It is a question then, as in finger-prints, of counting strokes, and if the strokes are alike in number in any two instances, of looking then as to how they are arranged. Two characters with the same number of pen-strokes under the same radical or key, may bear quite a different aspect’ [12]

Image Sources: (left) https://www.researchgate.net/publication/220756612_Effectiveness_of_Assigning_Confidence_Levels_to_Classifiers_and_a_Novel_Feature_in_Fingerprint_Matching

(right):? https://www.nature.com/articles/s41598-017-15536-w

Look at the similarities. A center point in the algorithm and a radical in an ideograph are ascertained. From there, the arrangement of a number of items with relative angles are counted which allows further subclassification. Fauld's last statement goes on to note that mere counting is not sufficient to differentiate two ideographs.? This makes Faulds the first to articulate that a point standard was not a valid method for identification.?

If we survey the eye tracking literature on fingerprint examination, we see that these concepts map nicely into the strategy of Examiners. They tend to use the core and delta as center points and then branch out around these areas. Indicating that like the radical of the ideograph and the centerpoint of the algorithm, focal points act as an anchor and clusters of data are examined around the anchors.

Image source: https://iab-rubric.org/images/pdf/papers/2020_TBIOM_LatentEyeGaze.pdf

Per the study, ‘The above results show that examiners use core and delta points as reference while searching for minutiae. The relative positioning of minutiae points around the core or delta act as a distinct trait during the comparison...’ [13]

I’m going to add some conjecture at this point. I think that there is another missing correlation between the concepts outlined above.? If we incorporate the recursive elements from earlier and compare that to the image I included of the ideographs, we can see that ideographs have strokes that recursively iterate onto radicals, radicals recursively iterate into characters, characters recursively iterate into words, and words recursively aggregate to sentences, which are propositions that have truth values. Similarly, it would seem there is, in effect, a grammar of feature selection that builds into what amounts to propositions about the existence of minutiae that have truth values.? Similarly, in the AFIS Algorithm, the additive properties of minutiae aggregate into fuzzy truth values.?

If we bring the notion of propositions with truth values into the comparison phase, we can see how a comparison works in a formal way. In the AFIS algorithm example above, a corresponding truth table is formed based upon truth values, much in the same way we articulated the truth value of a minutiae in the human example.?

Image source: https://www.researchgate.net/publication/220756612_Effectiveness_of_Assigning_Confidence_Levels_to_Classifiers_and_a_Novel_Feature_in_Fingerprint_Matching

If this is done to both a latent fingerprint and a known fingerprint, then a comparison really just becomes the comparison of truth values of the minutiae found in each. When an examiner does this it takes the form of a chart. This chart is from 1920, showing how early this equivalence relation was understood.? ‘the characteristic points of both impressions are diagrammed to correspond with each other’ (footnote 5 above)

Figure 8 : A fingerprint comparison chart

Image Source:

https://archive.org/details/fingerprintssimp00holtuoft/page/116/mode/2up

This corresponds nicely to an equivalence relation in set theory where there is correspondence between sets.

.?

Image Source: https://www.youtube.com/watch?app=desktop&v=nuewUugSqbk

And we know this type of thinking exists within the Examiner community because the language of equivalence relations is used frequently. Take the following example. A suspect is identified through fingerprints. When the suspect is ready to go to trial a motion to compel fingerprints is filed in many jurisdictions to ensure the fingerprints used in the identification match the defendant in custody. An Examiner shows up and takes a rolled set of fingerprints.? The whole case is not reworked in its entirety however, only the two sets of rolled prints are compared. When asked by the attorneys in the case, the Examiner gives the following analogy. If A = B and B = C then A = C. The idea being, if the latent prints (A)? in the case are identified to rolled prints (B) and the rolled prints (B) match the newly obtained rolled prints (C), then the latent prints (A) match the newly obtained rolled prints.(C). This is known as a transitive property of equivalence relations.

Image source: https://slideplayer.com/slide/5292138/

In the example we see the equivalence relations play out at two levels. First between latent and the rolled prints and secondly between the two sets of rolled prints. It stands to reason this relationship extends down to the level of the minutiae. It is unsurprising then given the recursive nature of the analysis that minutiae in a print demonstrate a reflexive property (e.g. this ridge ending is this ridge ending) as well as a symmetric relationship between the values in a truth table (e.g. the ridge ending in the latent two ridges from the core exists in latent (A) and in the rolled prints (B) finger #2 two ridges from the core).

Looking at the literature on how to implement quality assurance measures, we see the testing of these relationships. For instance in prints with lower quality Examiners tend to mark minutiae with a good, bad, ugly color codes, which directly speaks to the confidence in the fidelity of the reflexive property of individual minutiae in a print. In cases where conclusions are being challenged using consensus as a quality assurance measure, the fidelity of the reflexive property of minutiae is exactly what is in debate.?

When comparing the minutiae between two prints (the truth table values show congruence) working from the rolled print to the latent print in a predictive way can be used to test the fidelity of the conclusion. Conversely, literature on bias attempts to blind the Examiner to the fidelity of symmetric properties of minutiae between two prints. Studies show that changes in markup after exposure to a rolled print does not change the error rate for false identifications and necessarily cuts against this claim. [14]

In summary, change tends to move at the speed of ambition or error, not necessarily thoughtfulness. A cursory reading of the fingerprint literature tends to relegate Faulds to a pioneer who converged on the idea of fingerprints as a means of identification but who was beaten to the finish line of the race at the time, devising a fingerprint classification system. Reading him in his own words we see that to be a gross misunderstanding due in part to the ambition and bravado of his contemporaries.? Faulds laid the groundwork for a thoughtfulness that has been sorely missing in the contemporary statistical revolutions of forensic science.

G. K. Chesterton said:

There exists in such a case a certain institution or law; let us say, for the sake of simplicity, a fence or gate erected across a road. The more modern type of reformer goes gaily up to it and says, “I don’t see the use of this; let us clear it away.” To which the more intelligent type of reformer will do well to answer: “If you don’t see the use of it, I certainly won’t let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it.” [15]

Over 100 years ago, Faulds was laying the seeds of this very thought when he lambasted the institutional aristocracy of his day, Scotland Yard. In a discipline that necessarily engages the metaphysical aspects of the mind, they chose to portray subjective probabilities as certainty.

He said, "It has been laid down with every superfluity of emphasis that — not one finger-print, but — four points of agreement in a possible forty or so which an average single finger-print contains, is enough to secure an infallible identification by the "experts" of Scotland Yard. Three points, it is said, may be found to agree in two finger-prints taken from different persons, but never can there be found a case with four. This, however, is absolute and utter nonsense. Either those who use such language are unacquainted with the elements of dactylography or have a strong and unwholesome bias towards the reconviction of old criminals whatever the nature of the evidence may be which is alleged in the case." [16]

Names change, but the nature of people and the institutions they form don’t. Wisdom is a commodity without a currency in an industry full of bureaucracy, ambition and ego. Science as a pursuit of certainty, ambition or a self appointed clerical decree defaces the beauty of its curiosity and negates the people who practice it, just like it did to Henry Faulds.

1. Faulds, H. (1912). Dactylography; or, The study of finger-prints. (p. 86).? London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/dactylographyors00faulrich/page/86/mode/2up?
2. Henry, E. R. (1905). Classification and uses of finger prints. (p. 61). London: George Routledge & Sons. Retrieved from https://archive.org/details/b20444102/page/n7/mode/2up
3. Cegal. (n.d.). Human algorithms. Retrieved April 6, 2023, from https://www.cegal.com/en/dictionary/human-algorithms
4. Kuhne, F. (1916). The finger print instructor. (p. 61). New York: The Macmillan Company. Retrieved from https://archive.org/details/fingerprintinstr00kuhn/page/60/mode/2up
5. Kuhne, F. (1916). The finger print instructor. (p. 121) New York: The Macmillan Company. Retrieved from https://archive.org/details/fingerprintinstr00kuhn/page/120/mode/2up
6. Faulds, H. (1912, p 86). Dactylography; or, The study of finger-prints. (p. 86). London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/dactylographyors00faulrich/page/86/mode/2up?
7. Faulds, H. (1912). Dactylography; or, The study of finger-prints . (p. 87). London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/dactylographyors00faulrich/page/86/mode/2up?
8. Faulds, H. (1905). Guide to finger-print identification. (p. 47) ?London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/b20443493/page/n69/mode/2up
9. Faulds, H. (1905). Guide to finger-print identification. (p. 71) ?London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/b20443493/page/70/mode/2up
10. Faulds, H. (1905). Guide to finger-print identification. (p. 49) ?London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/b20443493/page/n73/mode/2up
11. Faulds, H. (1905). Guide to finger-print identification. (p. 14) ?London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/b20443493/page/14/mode/2up
12. Faulds, H. (1912). Dactylography; or, The study of finger-prints. (p. 86).? London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/dactylographyors00faulrich/page/86/mode/2up?
13. Malhotra, A., Sankaran, A., Vatsa, M., Singh, R., Morris, K. B., & Noore, A. (2020). Understanding ACE-V latent fingerprint examination process via eye-gaze analysis. IEEE Transactions on Biometrics, Behavior, and Identity Science. https://iab-rubric.org/images/pdf/papers/2020_TBIOM_LatentEyeGaze.pdf
14. Ulery, B. T., Hicklin, R. A., Roberts, M. A., & Buscaglia, J. (2015). Changes in latent fingerprint examiners’ markup between analysis and comparison. Forensic Science International, 247, 54–61. https://pubmed.ncbi.nlm.nih.gov/25553355/
15. Society of Gilbert Keith Chesterton. (2012, April 30). Taking a fence down. https://www.chesterton.org/taking-a-fence-down
16. Faulds, H. (1905). Guide to finger-print identification. (p. 47) ?London: Simpkin, Marshall, Hamilton, Kent & Co. Retrieved from https://archive.org/details/b20443493/page/70/mode/2up