Shelia Guberman (born 25 February 1930 in Felsztyn, Kamianets-Podilskyi Oblast, Ukraine, USSR) is a scientist associated with the USSR and the United States who works in nuclear physics, computer science, geology, geophysics, medicine, artificial intelligence, and the psychology of perception. He proposed the D-waves theory of Earth seismicity, algorithms of Gestalt-perception (1980) and image segmentation, and programs for the technology of oil and gas fields exploration (1985).^[1]

Guberman is the son of writer and poet Aizik Guberman and his wife Etya, a teacher. From 1947 to 1952, he studied at the Institute of Electrical Communications in Odessa, USSR, graduating in radio engineering. From 1952 to 1958, he worked as a field geophysicist in the Soviet oil industry, and subsequently studied as a postgraduate at the Oil and Gas Institute in Moscow from 1958 to 1961. In 1962 he received a PhD in nuclear physics, followed by a PhD in applied mathematics in 1971, the year he was appointed to a full professorship in computer science.

After authoring the first applied pattern recognition program in 1962, Guberman specialized in artificial intelligence, implementing principles of Gestalt perception in computer programs for geological data analysis. In 1966, he was invited by mathematician Prof. Israel Gelfand to lead the Artificial Intelligence team at the Keldysh Institute of Applied Mathematics of the Russian Academy of Sciences. He applied pattern recognition technology to earthquake prediction, oil and gas exploration, handwriting recognition, speech compression, and medical imaging. From 1989 to 1992, Guberman held the chair professorship at the Department of Geography at Moscow Open University.

He has lived in the United States since 1992. Guberman is the inventor of the handwriting recognition technology implemented in the commercial product by Paragraph International, a company founded by Stepan Pachikov, which is used today by Microsoft in Windows CE.^[2] He is also the author of core technologies for five US companies and owns a patent on speech compression.^[3]

The common approach to computer handwriting recognition was computer learning on a set of examples (characters or words) presented as visual objects. Guberman proposed that it is more adequate for the psycho-physiology of human perception to present the script as a kinematic object, a gesture, i.e. synergy of movements of the stylus producing the script.^[4] The handwriting consists of 7 primitives. The variations which characters undergo during writing are restricted by the rule that each element can be transformed only into its neighbor in the ordered sequence of primitives. During the evolution of Latin-like writing, it acquired resistance to natural variations in character shape: when one of the primitives is substituted by its neighbor, the interpretation of the character does not change to another one.

Based on this approach, two USA companies, ParaGraph and Parascript, developed the first commercial products for on-line and off-line free handwriting recognition, which were licensed by Apple, Microsoft, Boeing, Siemens, and others.^[6,7] Most commercially available natural handwriting software is based on ParaGraph or Parascript technology.^[8]

The hypothesis that humans perceive handwriting as well as other linear drawings (and communication signals in general) not in visual modality but in the motor modality was later confirmed by the discovery of mirror neurons.^[9] The difference is that in the classical mirroring phenomena the motor response appears in parallel with the observed movement ("immediate action perception"), whereas during handwriting recognition the static stimulus is transformed into a time process by tracing the path of the pen on the paper. In both cases, the observer is trying to understand the intention of the correspondent: "the understanding of what the person is doing and why he is doing it, is acquired through a mechanism that directly transforms visual information into a motor format".^[10]

Speech is traditionally presented as a time sequence of phonemes, comprising vowels and consonants.^[11] Each vowel is mainly determined by the relationship between the volume sizes of the front and back of the vocal tract, a ratio defined by the horizontal back-and-forth position of the tongue, the back-and-forth position of the lips, and the size of the pharynx, which can extend the vocal tract cavity far back. Most consonants can be described using three parameters: the place of articulation (such as lips or teeth), the time pattern of interaction with the vocal tract (explosive or not), and whether the sound is voiced or unvoiced. Due to the inertia of articulatory organs such as the tongue, lips, and jaw, phonemes interfere with adjacent sounds to change their acoustic quality through co-articulation, resulting in each phoneme sounding different depending on its context.

Guberman presents a parallel model of speech production, proposing that vowels and consonants are generated in parallel rather than in sequence.^[12] This separation is possible because vowel and consonant production involves two distinct channels managing different groups of muscles that jointly define the vocal tract geometry and the resulting voice signal. For the vowels [o] and [u], the lips are controlled by the mentalis and orbicularis oris muscles for protrusion and rounding, whereas for [i] and [e], the buccinator and risorius muscles retract the lips. The tongue participates in creating vowels through the innervation of the superior longitudinal and vertical muscles for lifting and moving the whole tongue back and forth, while the genioglossus muscle is used for all consonants articulated in the front of the mouth when the jaw is fixed.^[13] For the lip consonants [p], [b], [v], and [f], the lips and jaw are moved up and down by the labii inferioris and orbicularis oris muscles, with the zygomaticus minor moving the lower lip back specifically for [v] and [f].

Several conclusions follow from the hypothesis of Parallel Phonetic Coding. First, because vowels are defined by the ratio between front and back vocal tract volumes, vowels are present at every moment of speech, including silence, where the neutral vowel [ə] occurs when no vocal tract muscle is innervated. Second, every consonant in speech appears against the background of a vowel: the final consonant in a word is pronounced against the background of [ə], and in consonant clusters, all consonants except the last are produced in parallel with [ə]. Historically, Russian orthography required writing a special character denoting the neutral vowel—Ъ—after a word-final consonant, a rule that was abolished in 1918. Third, in written representations of words such as "soda" and "word", the number of vowels in a syllable reflects the relative duration of the vowel. Such parallel coding is reflected in Hebrew, where two dots beneath characters indicate the vowel [e] in the word יצֵירֵ ("peace"), and in Arabic, where the consonant stream carries the semantic root while the vowel stream modifies that meaning or marks grammatical categories (e.g., kitab "book", katib "writer", ia-ktub-u "he is writing", and ma-ktab "school").

In the 1970s and 1980s, Guberman developed artificial intelligence software and the appropriate technology for geological applications, using it to predict the locations of giant oil and gas deposits.^{[14,15,16,17]} The technology employs morphostructural zoning maps (a method proposed and developed by Prof. E. Rantsman) to outline morphostructural nodes at fault intersections, combined with a pattern recognition program that identifies nodes containing giant oil or gas fields.^[18] This model, proposed by Prof. Yury Pikovsky of Moscow State University, assumes that petroleum migrates from the mantle to the surface through permeable channels created at the intersection of deep faults.^[19] In 1986, the team published a prognostic map for discovering giant oil and gas fields in the Andes in South America based on the abiogenic petroleum origin theory. The map forecasted that eleven nodes, which had not been developed at the time and covered only 8% of the total area of all the Andes basins, contained giant oil or gas fields.

Thirty years later, in 2018, the results of comparing the prognosis with reality were published.^[20] Since the publication of the 1986 prognostic map, only six giant oil and gas fields were discovered in the Andes region: Cano–Limon, Cusiana, Capiagua, and Volcanera (Llanos basin, Colombia), Camisea (Ukayali basin, Peru), and Incahuasi (Chaco basin, Bolivia). All of these discoveries were made in locations identified on the 1986 prognostic map as promising areas. The result is convincingly positive and represents a strong contribution in support of the abiogenic theory of oil origin.

In the middle of the 20th century, seismologists observed the phenomenon of chains of earthquakes consistently arising along major faults, which was later interpreted as waves of tectonic strain.^{[21,22,24,23]} In 1975, Guberman proposed the D-waves theory to separate local processes of stress accumulation from the triggering of earthquakes.^[24,23] The basic postulates of the theory hold that a strong earthquake alters the mass distribution in Earth's core and its rotation rate (ω), and when ω reaches a local minimum, disturbances known as D-waves propagate along meridians from both poles at a constant speed of 0.15° per year. According to the theory, a strong earthquake is triggered where tectonic stresses have accumulated at the moment two D-waves originating from the North and South poles intersect.

The source of irregularity in Earth's rotation can be a strong earthquake that displaces massive rock volumes, requiring the angular speed of rotation ω to change in order to conserve rotational momentum; due to the slow speed of D-waves (0.15°/year), it takes more than 200 years to reach regions where earthquakes of magnitude M > 8 occur.^[26] This hypothesis was supported by seismological data showing D-waves triggering strong earthquakes (magnitude M ≥ 7.0) in Alaska and the Aleutian Islands, California, Southeastern Europe, Asia Minor, Southern Chile, the South Sandwich Islands, New Zealand, France, and Italy, with the probability of this occurring by chance being less than 0.025 in each case.^[25] Testing the postulate of polar wave origin required long-term seismic records, which are documented in China dating back to 180 A.D.^[28] Among the six strongest documented earthquakes in China, an initial earthquake generated two polar D-waves: the wave from the North Pole triggered a second earthquake in 332 years, while the wave from the South Pole triggered a third earthquake in 858 years, with an overall average deviation of only 0.4° between the D-wave position and the triggered earthquake epicenter—less than historical epicenter determination error.^[28]

The D-waves hypothesis further implies that epicenters of the strongest earthquakes occur predominantly at discrete D-latitudes defined by (90/2^n)·i (for i = 0, 1, 2, ... and n ≤ 5).^[28] To test this distribution, high-seismicity areas of Earth were divided into stripes 5.625° wide parallel to D-latitudes of order n ≤ 4. Across 43 regions where earthquakes with M ≥ 8.0 occurred, the strongest earthquake in 31 of those regions had its epicenter located within a 1°-wide stripe centered on a D-latitude. Because a 1°-wide stripe occupies 0.36 of each 5.625°-wide region, a random distribution would yield an expected 15 epicenters (43 × 0.36), making the probability of 31 epicenters falling within the stripes less than 0.005.

Earthquakes represent an essential component of tectonic movement, and strong earthquakes occur at morphostructural nodes (the intersections of faults), indicating that major morphostructural knots are also concentrated near D-latitudes.^[29] The occurrence of strong earthquakes at discrete D-latitudes in turn influences the broader tectonic configuration of fault networks.^[30] Combined with Professor Pikovsky's hypothesis that morphostructural knots act as conduits delivering oil from the mantle to the Earth's crust, giant oil and gas fields are also predominantly located at discrete D-latitudes, a parameter utilized in petroleum exploration. It was additionally found that most infrastructure failures on oil, gas, and water pipelines, as well as railroad rails, occur within these morphostructural knots.^[31]

Two types of treatment exist for patients with hemorrhagic strokes: passive (medicamental) and active (surgical). Prof. E. Kandel, one of the pioneers in the surgical treatment of hemorrhagic strokes, turned to mathematician Prof. I. Gelfand for help in comparing the effectiveness of these two treatments. Guberman was chosen as the main architect of the project. First, it was decided to change the goal from choosing the best treatment in general to finding the best treatment—conservative or operational—for a particular patient, under the philosophy of treating the patient rather than the disease. For this purpose, it was decided to use pattern recognition technology previously developed for geology. Two decision rules had to be developed: one for predicting the outcome (life or death) of conservative treatment for the particular patient, and another for predicting the outcome (life or death) of surgery for the same patient. The decisions were based on neurological and general symptoms collected within the first 12 hours after the patient arrived at the hospital.

The obtained decision rules underwent preliminary testing for two years, during which collected data were sent to the computer and the two prognoses forecasting outcomes of operation and conservative treatment were placed in the patient's file. A month later, the computer predictions were compared with the actual outcomes, yielding an overall result of 90% correct predictions. Clinical implementation followed, where the computer decisions were immediately sent to the surgeon on duty, who made the final decision. In five years, 90 patients received computer forecasts.^[33,34] In 16 cases where the computer strongly recommended operation, 11 patients were operated on and survived, while for 5 patients the computer warning was neglected for various reasons and all 5 died. In 5 cases where it was strongly recommended to avoid operation, 3 patients were treated accordingly and survived, whereas 2 were operated on contrary to the computer advice and died.

From 1966 to 1991, he served as Chief Scientist at the Keldysh Institute of Applied Mathematics in Moscow, Russia. Between 1989 and 1992, he held a Chaired Professorship in the Department of Geography at the Russian Open University in Moscow. From 1989 to 1997, he was Chief Scientist at ParaGraph International in Campbell, California, United States. He also served as a Visiting Scientist at Lawrence Berkeley National Laboratory in California, United States, from 1995 to 1996. From 1998 to 2007, he was the Founder and CEO of Digital Oil Technologies in Cupertino, California, United States.

More than 180 papers have been published in scientific journals in Russia, the US, France, Germany, Italy, and Austria. Selected recent papers on computer science and psychology include "Reflections on M. Wertheimer`s "Productive Thinking": Lesson to Artificial Intelligence" (Gestalt Theory, 2001), "Clustering Analysis as a Gestalt Problem" (Gestalt Theory, 2002), "Reflections on Ludwig Bertalanffy's "General System Theory: Foundations, Development, Applications"" (Gestalt Theory, 2004), and "What is «self-organization»? A journey of a small child" presented at the 7th Congress of the UES Systems Science European Union in Lisbon (2008). Further publications in computer science and psychology include "Gestalt and Image Understanding" with Vadim V. Maximov and Alex Pashintsev (Gestalt Theory, 2012), a 2013 critical review of Desolneux, Moisan & Morel (2008): From Gestalt Theory to Image Analysis (Gestalt Theory), "On Gestalt Theory Principles" (Gestalt Theory, 2015), and "Gestalt Theory Rearranged: Back to Wertheimer" (Frontiers in Psychology, 2017). A 1972 publication is also noted under selected papers on tectonophysics.

Sources discussing his work include E. Zueva's 2009 paper, "The history of computer vision in the Keldysh Institute of Applied Mathematics of the Russian Academy of Sciences," published in Mathematical Machines and Systems. Another source is the 2009 article by E. M. Kudryavtsev, E. F. Maklyaev, S. D. Zotov, and A. A. Lebedev, titled "Comparison of Hypothetical D-Waves of Planetary Scale, Causing Earthquakes in Sh.A.Guberman's Model, with Slow Solitary Elastic Waves (SSEWs) Detected Experimentally," appearing in the Bulletin of the Lebedev Physics Institute. Additionally, Joel N. Shurkin authored "Thoroughly Electronic Russian" in 1994.