Systems and methods for measuring the value of a sensed variable

Abstract

A method and system for generating a quantitative perception profile of a predetermined stimulus are described. The method can be used, for example, in colorimetry or flavor quantification. In the case of flavor quantification, flavor descriptors are selected (22, 21) from a database (7) for a specific product and presented (27) pairwise in all possible pairwise combinations to each of a plurality of tasters via a user device (2). For each pair of flavor descriptors, the user selects the descriptor that is considered to be stronger compared to the other descriptor. The results of the selection are captured (28, 24) at a host (1) and mapped by a normalization component (26) as intensity values into a flavor description space in the database (8) to create a quantitative flavor profile. The normalization component (26) performs a weighting on each descriptor selection result, which adjusts the intensity contribution of the descriptor to the flavor profile according to a probability function that is calculated depending on the likelihood that other tasters select a descriptor other than the considered descriptor from all pairs containing the considered descriptor. A filtering component (25) detects abnormal or inconsistent selection responses and issues an adaptive control instruction (32) to modify or repeat the descriptor pairs in the sequence presented to the user.

Description

Technical Field

[0001] The present invention relates to computer-implemented systems and methods for empirically measuring a human's perception of a predefined stimulus. In particular, the present invention relates to representing the perception of a stimulus in a multi-variable parameter space based on a set of predefined variables characterizing the stimulus. In a first exemplary application of the present invention, the stimulus may be a colored area of a computer display, and the perception variables may include luminance, contrast, or the perceived color, where the perceived color is expressed according to a standardized library of descriptors such as Pantone® numbers or definitions in a color space (e.g., RGB). In a second exemplary application of the present invention, the predefined stimulus may be the flavor or aroma of a food product, or other perceivable characteristics, in which case the present invention relates to representing the perceived characteristics of the food product in a normalized flavor profile space - for example, a flavor space that can be mapped to a predefined flavor annotation library. Background Art

[0002] In many industrial fields, it is important to be able to reliably measure the perception of specific characteristics of a product. For example, manufacturers of computer displays can typically use colorimetry to specify the color gamut and intensity depth of a particular display with high precision. Colorimetry can be defined as "the technology for quantifying and describing human color perception physically", and its purpose is to represent visible color characteristics as physical correlates of color perception, such as color space tristimulus values. Display manufacturers can employ color engineers experienced in colorimetry to characterize the color space and intensity characteristics of the display output when the display is undergoing calibration and quality control. Since color engineers cannot reliably measure the actual perception of color and intensity stimuli (as they are perceived by the users of the display), traditional colorimetry techniques are limited to measuring the characteristics of the display and the characteristics of the light emitted by the display (color temperature, spectral radiance, reflectance, tristimulus values, etc.). However, in practice, the user's perception of color depends not only on the physical characteristics of the emitted light. For example, it also depends on environmental variables such as the type and luminance of ambient light, viewing angle, air cleanliness / diffusion, moiré interference, or even room temperature or time of day, and it may also depend on user-specific variables such as the user's age, eye physiology, or afterimage effect. In theory, it would be possible to make detailed measurements of such environmental and user variables and thereby generate a high-resolution perception profile that correlates different values of the perception variables with different values of the color / intensity stimuli. However, this would actually be complex and laborious, and substantially impossible to achieve in real time. This also means acquiring a large amount of data that is not directly of interest to the manufacturers of computer displays (foveal scans of the user, local environmental data, etc.), and would only provide perception profile data for specific combinations of colorimetry variables and stimuli.

[0003] In other technical fields, similar problems exist with empirical measurements of colorimetry, such as in the manufacture of paints and coatings, where perception is of reflected light or transmitted light, or of both. Experienced color formulators can express colorimetric perceptions of hue, tone, and surface texture in coded terms. Surface texture can have a profound effect on the perceived surface color, especially where colorimetry is mainly determined by the reflectance characteristics of the surface. In cases where two or more color formulators need to cooperate, they can use a common library of coded colorimetry terms and calibrate their measurements against the common library.

[0004] The colorimetry techniques described above can, in principle, be used to quantify and normalize any human perception where the sensed quantity is measurable. For example, transducers and acoustic analyzers can be used to accurately measure sound waves in air, and, for example, the auditory perception of a sound spectrum can be represented as a multi-dimensional sound net or acoustic map that can be used in architectural acoustics engineering.

[0005] Techniques similar to colorimetry can also be used to quantify human perceptions of touch, taste, and smell. Like colorimetry, human perception of the flavor and aroma of substances is a perception of the interaction of taste receptors with physical or chemical quantities (such as the molecular composition of the substance), which physical or chemical quantities can be measured using, for example, chemical or mass spectrometry techniques. Like colorimetry, in theory, external environmental variables and user-related variables (e.g., genetic factors that may affect the perception of molecules such as phenylthiocarbamide) can be taken into account in a similar way to measure in detail the perception of the flavor and aroma of substances for an individual, but in practice this task would be infeasible and would generate a large amount of redundant data. Like colorimetry, a library of coded terms can be used to characterize human perception of flavor and aroma.

[0006] In the following description, examples of flavor perception will be mainly used to describe the present invention. However, the principles of the present invention can also be applied to other kinds of human perception, especially where the perception variables are objectively measurable (such as sensations of color vision, audiology, texture / touch, etc.).

[0007] Prior Art

[0008] To accurately quantify the perception of color stimuli (e.g., graphics on a computer monitor) at high resolution, it is known to compare the displayed color against a printed chart or color swatch. Alternatively, an experienced color formulator can select the closest value or combination of descriptor values with reference to a conventional set of standardized reference descriptors. However, this is a time-consuming process and is prone to inaccuracies and inconsistencies.

[0009] To accurately quantify the perception of the flavor of a substance, as mentioned above, it is theoretically possible to isolate the individual flavor-defining components of the substance and analyze (e.g., using chemical or biochemical techniques) the perceived flavor characteristics of each component and each combination of components for different types of tasters. However, this is likely to be impractical. Instead, typically a panel of experienced human tasters is employed and the tasters are trained to ensure consistent and reproducible evaluations of the relevant flavors or aromas. Establishing a panel of tasters is expensive and the tasters need ongoing training and calibration against other tasters in the same panel and / or other panels. Such a panel of tasters, which typically consists of a small number of trained tasters, is expensive and they have limited capacity in terms of the number of tastings they can perform. Tasters are typically likely to be asked to distinguish between thousands of different flavor annotations. They work together in one place in a standardized environment and their taste evaluations are monitored to ensure that individual tasters are applying flavor descriptors in the same way as other members of the panel. The tasters are trained on scales for the intensity of flavor, texture, and aroma. To ensure consistency and accuracy, each taster must submit multiple redundant evaluations, which are then averaged within a group. This creates a substantial need for data storage and processing capabilities to develop even a single flavor profile. Further, since these scales must be collected separately for each food product tested and not reused, this method generates large storage capacity requirements when measuring across multiple products or panels. It is not reasonable to extend traditional perception quantification systems to hundreds of thousands or millions of users because the data processing and storage architectures required to handle all the data needed to create a high-resolution descriptor space are impractically large. Such a system would also not be able to adapt in real time. Further still, this work is time-consuming. For example, a skilled and experienced taster may even spend an hour characterizing the flavor of a product. Traditionally, different systems use different proprietary libraries of flavor descriptors, which means that it is difficult, if not impossible, to compare the flavor evaluations of one system based on one flavor descriptor library with those of another system. This also makes it difficult for trained panel members to switch to different systems.

[0010] Similarly, with colorimetry, while color parameters related to the physical characteristics of the emitted light can be numerically mapped or transformed between color representation systems, systems that characterize color perception variables may be difficult or impossible to map in the same way.

[0011] Further, since existing quantitative evaluation methods for color and food require specialized training, this known technique makes it impossible to measure an individual's taste perception.

[0012] In perceptual quantification systems such as the colorimetric and tasting systems mentioned above, the descriptor space can include a large number (tens of thousands or even hundreds of thousands) of variables (e.g., color or flavor annotations) for characterizing stimuli (e.g., the color of the displayed graphic or the flavor of the food). A simpler and more consistent method for generating such a high-resolution multi-variable descriptor space is needed. SUMMARY OF THE INVENTION

[0013] The present invention attempts to overcome at least some of the above-mentioned and other disadvantages inherent in the prior art. In particular, the present invention aims to provide the method according to claim 1 and the system according to claim 16. Further variations of the present invention are stated in the dependent claims. By obtaining and normalizing a large number of pairwise descriptor quantifications based on a small set of descriptor variables, it is possible to generate a high-resolution descriptor space with fine granularity (i.e., a much larger number of descriptor variables). The small size of the set on which the pairwise quantification (user selection) is based means that the descriptors of the set can be given simple identifiers that are understandable to many people and to those without detailed experience in specific descriptor quantification. Although a library of fine-grained descriptor identifiers was previously required to describe the variables for defining a high-resolution descriptor space, the method of the present invention enables the creation of a high-resolution descriptor space using a much smaller set of user-friendly descriptor identifiers by presenting many simpler descriptor pairs for selection to many users. Furthermore, the system is thus scalable to an unlimited extent. Previous systems were limited by the descriptor granularity that could be distinguished by individual users (panelists). The method of the present invention allows the creation of a higher-resolution descriptor space without requiring users to have higher skills or expertise.

[0014] Thus, by capturing many answers to the pairwise comparison problem, it is possible to achieve a similar quality of perceptual variable (descriptor) quantification as that achieved by expert colorimetricians or tasting panels, without requiring training of colorimetricians / tasters and without requiring large-scale data processing capabilities to calculate the results. Furthermore, the achieved descriptor characterization is consistent and can be quantified in a way that provides a normalized descriptor space, thus allowing different user groups in different locations to perform different quantifications. Pairwise interrogation, as well as the automatic processing of pairwise answers, can also significantly reduce the time spent by each user - a few minutes instead of an hour or more using traditional methods. Although the time required for user participation is reduced, the accuracy of the resulting perceptual profiles can be greatly improved. Additionally, all the data points (user selections) in the descriptor space contribute more or less to improving the accuracy of the perceptual profiles.

[0015] The above advantages can in particular be achieved using an adaptive variant of the method described below, in which the selection and / or ranking of the paired descriptors is determined based on the user's past selections. This adaptive variant also reduces the amount of network traffic (less user interaction) and reduces the collection of redundant information. This increases the processing speed and reduces the storage and processing power requirements. This also allows the resolution of the descriptor space to be varied to meet different requirements. For example, certain regions of the descriptor space may contain more useful information than other regions, and these regions can be characterized with a higher resolution (greater density of descriptors per point). The invention also facilitates the creation of a standardized interface, whereby the descriptor space (e.g., color / aroma / flavor) can be matched with other descriptor libraries, such that the descriptor space and / or the acquired perceptual profiles can be easily mapped to other descriptor spaces, such as the proprietary internal descriptor spaces of, for example, color display manufacturers, food manufacturers, or perfume houses. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In the following detailed description, together with the illustrated example embodiments and implementations given in the accompanying drawings, the invention and its advantages will be further explained, wherein:

[0017] Figure 1 An example server-client system for performing the method according to the invention is shown;

[0018] Figure 2 An example of a flavor descriptor space is shown, where the stimulus is the taste of the substance being analyzed;

[0019] Figure 3 An example of a color descriptor space is shown, where the stimulus is a light red area displayed on the computer being analyzed;

[0020] Figure 4 A first example of a server-client configuration for performing the method according to the invention is shown in more detail;

[0021] Figure 5 A second example of a server-client configuration for performing the method according to the invention is shown in more detail;

[0022] Figure 6 A first variant of a user device for performing the method according to the invention is shown;

[0023] Figure 7 A second variant of a user device for performing the method according to the invention is shown;

[0024] Figure 8 An example of a flavor descriptor space with a target flavor profile is shown.

[0025] The accompanying drawings are only intended to illustrate exemplary embodiments of the present invention and are not to be construed as limiting the scope of the present invention. In cases where the same reference numerals are used in different drawings, these reference numerals are intended to refer to the same or corresponding features. However, the use of different reference numerals per se should not be regarded as an indication of any specific differences between the features being referred to.

[0026] As used herein, the term "color" includes, in addition to colorimetric parameters and spectral parameters, other parameters that can be used to characterize the light emitted, reflected, or transmitted, such as, for example, intensity, radiance, surface texture, and the like.

[0027] As used herein, the term "flavor" is to include aroma. References to "food" or "food product" should be understood to include any substance or product that is edible, drinkable, inhalable, or smellable. Detailed Description

[0028] Figure 1 A system is shown in a very simplified schematic form that can be configured to perform the method of the present invention. A backend server 1 (also referred to as a host or cloud platform) is arranged to communicate with a plurality of user devices 2 such as smart phones, laptop computers, tablet computers, or desktop computers via a communication network 3 such as the Internet. As will be described below, the host 1 is configured to send descriptors to the user devices 2 and collect and process descriptor selection responses from the user devices 2.

[0029] Figure 2 A graphical representation of a flavor descriptor space 4 for representing the flavor profile of a food product is shown. In this example, thirteen flavor descriptors are depicted on thirteen radial axes, each flavor descriptor having an intensity scale of 0 - 100%. In this example, the thirteen descriptors have been preselected as being suitable for characterizing dark chocolate products. Two different dark chocolate products are characterized by dotted lines 5 and dashed lines 6.

[0030] For the colorimetric applications described above, Figure 3 a similar descriptor space is shown. If the stimulus is a light red graphical element displayed on a computer screen, the descriptor variables can include hues such as "lilac", "rose pink", "coral", "light orange", etc. The visual perception of Caucasians tends to be sensitive to such light pink hues, and manufacturers of video displays may need to perform a higher resolution quantization of the descriptors in this color subspace.

[0031] In the remainder of the description below, flavor / aroma applications are mainly cited as examples, but it should be understood that the same principles can be used in other applications such as colorimetry, audiology, and the like.

[0032] Figure 4 shows in more detail Figure 1 An example of the functional elements of a first example implementation of host 1. Note that host 1 does not need to be a single device, but may include multiple devices or virtual devices running on a cloud platform or virtual server. A set of flavor descriptors (also known as flavor annotations) applicable to the specific product or product type being analyzed is stored in database 7. For example, in the case of a chocolate product, the flavor descriptors may include Figure 2 the thirteen descriptors shown in

[0033] The flavor descriptor pairs are presented at each user device 2 (e.g., on a display screen) for selection by the user of device 2. The user indicates which of the two presented flavor descriptors more closely reflects his or her perception of the product flavor, for example, by swiping left or right on a touch screen or touching a graphical button or other form of button. Each user selection is transmitted to the receiving interface 24 at host 1 for filtering 25 and normalization 26. Preferably, the paired interrogation 30 is repeated for each of the flavor descriptor pairs, and this process 31 is carried out for each repetition in user device 2. In a colorimetric application, colors corresponding to the descriptors may optionally be displayed in addition to or instead of the text descriptor terms.

[0034] As described above, these pairs can be transmitted from host 1 to user device 2 sequentially. Alternatively, a set of pairs for one product for one taster can be transmitted to user device 2 all together or in batches of multiple pairs. Similarly, user device 2 can transmit the response selections to the receiving component 24 in real time (i.e., as the selections are made), or user device 2 can accumulate some or all of the responses before transmitting all the responses together or in batches of multiple responses to the receiving component 24.

[0035] A filtering component 25 can be provided to filter out abnormal user selections, or to exclude a particular user's selections from the selection data to be mapped to the flavor descriptor space in the database 8, or to weight user selections according to some other criterion. Optionally, the output 32 from the filtering component 25 can be used as an adaptive control input to the selection component 21 such that the selection of flavor descriptor pairs can be performed adaptively, for example to repeat the interrogation of descriptor pairs for which the filtering component 25 has detected abnormal user selections. As will be described below, the filtering component 25 can be implemented as a machine learning function and / or a knowledge base of previous user selection patterns.

[0036] The user selections corresponding to the flavor descriptor pairs are adjusted by a normalization processor component 26 to map the selections to a multi-dimensional flavor descriptor intensity space, which can be implemented as an appropriate data structure in the database 8. The normalization component 26 weights each flavor descriptor so as to normalize the intensity values relative to other descriptor intensity scores. The normalization process will be described in more detail below.

[0037] Figure 6 An example of a graphical representation of one of the descriptor pairs on the touch screen of a smart phone is shown. In response to the displayed prompt, the device user (taster) indicates which of the two displayed descriptors, "citrusy" and "nutty", more closely describes the flavor of the substance being tasted. The user then selects the "winning" descriptor, and this selection is recorded and transmitted to the host 1, and then the next descriptor pair is displayed on the screen.

[0038] Figure 5 An alternative implementation of the server-client system for carrying out the present invention is shown. Although only one user device 2 is shown, it should be understood that a similar transmission / reception arrangement exists between the host 1 and each user device 2. In this example, alternatively, Figure 4 the functions of the host 1 are partially performed at the user device 2. Thus, the transmission component 22 transmits flavor descriptors (e.g., a set of 13 descriptors in the case of the dark chocolate product shown in Figure 2 to the selection component 21 located at the user device 2. The selection component 21 can be, for example, a mobile application or a part of the functionality of a mobile application. The selection component 21 optionally generates flavor descriptors under the control of a randomization component 23 (in Figure 2In the chocolate example, there are 78 pairs), and the pairs (optionally randomized) are preferably presented one at a time by the user interface output component 27 (such as a display, audio output, etc.) at the user device 2. For each pair of descriptors, the user's selection is recorded by the user interface input component 28 (such as a touch screen, keyboard, mouse, audio input) and transmitted by the transmission component 29 of the user device 2 to the receiving component 24 at the host 1. As in the implementation illustrated in Figure 4 Once all the pairs of descriptors have been presented and the selections recorded, the responses can be transmitted in real time, individually, in batches, or all together.

[0039] The randomization component 23 can be implemented using any suitable random or pseudo-random number generation algorithm. Further, based on the responses given by individuals and groups, the questions are adapted to avoid collecting redundant information that is "known" based on previous answers. Additionally or alternatively, each answer can be dynamically filtered for accuracy and consistency when compared to the previous answers of an individual user. In this way, the questions can be adapted to the tasting skill level of the respondent in order to capture accurate and relevant data at the level that is likely to be achievable for each individual user. This increases the depth of information that can be captured while eliminating redundant or misleading data that would increase storage and processing requirements.

[0040] As mentioned above, it is advantageous to generate pairs of descriptors dynamically in real time. This so-called "intelligent" generation of pairs of descriptors can be adapted to eliminate or reduce the occurrence of pairs of important additional characterization data that are judged to be unlikely to provide a stimulus (e.g., pairs that are uniformly treated in the same way by all users, or pairs related to a region of the descriptor space where individual users have been judged to have a poor consistency record). Some pairs may be located in regions of the descriptor space where there is already sufficient data density, so these pairs can be partially or completely suppressed. Or the opposite may be the case - regions of particular interest in the descriptor space may be sparsely populated, in which case more pairs can be generated for that region. In cases where users are randomly located around the world, it is possible that some regions are underrepresented or overrepresented, in which case the number and type of pairs can be dynamically adjusted in real-time response. The age of the data points in the descriptor space can also be considered. For example, if the data in a particular region is unusually old, the number of new pairs of descriptors can be increased to replace or supplement the existing data. In some applications, the user's perception can change over time, so it may be important to update the descriptor space.

[0041] The filtering component 25 can be configured to use hierarchical descriptor data, where descriptors are provided at different levels of granularity, inheriting descriptor features from higher descriptors in the hierarchy. For example, in a flavor application, high-level descriptors (creamy, fruity, sweet, vanilla, etc.) can be used first to quantify the flavor of a yogurt product, and the filtering component 25 can employ a machine learning algorithm that automatically detects when a particular type of user response has reached a predetermined convergence level (i.e., when presenting further paired descriptors, a diminishing return level is achieved). At this stage, the filtering component 25 can be configured to automatically move to a higher granularity of descriptors (lower in the hierarchy). This change in granularity occurs in either direction and can be repeated as often as needed. For example, when the stimulus is the taste of strawberry yogurt, the filtering component 25 can relatively quickly move to a high granularity using strawberry flavor descriptors (of which there are many).

[0042] If convergence of user responses is detected for a particular region in the descriptor space, the filtering component can, for example, in a flavor application.

[0043] The filtering component 25 is provided for, e.g., detecting illogical, inconsistent, or incomplete selection results and optionally commanding the selection component 21 to repeat a particular pair or otherwise adapt the sequence of pairs presented to the user(s). To address situations in which a taster provides an incomplete set of selection results, the filtering component 25 can include a Bayesian or other statistical analysis engine for, e.g., inferring the value of a missing selection parameter from the taster's other selection results.

[0044] The normalization component 26 is configured to normalize the selection results and map them into the flavor profile space in the database 8, such as Figure 2 the example depicted in. In a traditional manual flavor evaluation environment, where flavor annotations are evaluated by individual trained tasters, the measurement of each flavor annotation can be done independently. This means that additional flavor annotations can be added to a product, and if all other flavor annotations remain the same, that flavor annotation can be measured and "added" to the flavor profile. In contrast, in the method of the present invention, all descriptor selections are made relative to other descriptors, so it is not possible to evaluate the selection result of an individual descriptor alone. Each descriptor selection score must be normalized relative to other descriptor scores.

[0045] The "winner" in each pair (i.e., the stronger flavor descriptor selected by the user from each descriptor pair) is processed by the normalization component 26. The normalization component can implement the following normalization function, for example:

[0046] intensity(X) = (wins(X) +1 ) / (wins(X) + 1 + losses(X) + 1).

[0047] In other words, the intensity value of descriptor X is calculated as a function of wins(X) above, where wins(X) is the number of times descriptor X is chosen over another descriptor, and loss(X) is the number of times another descriptor is chosen over descriptor X. This calculation is performed for each of the descriptor dimensions (e.g., 13) in the flavor descriptor space. Thus, the base value (i.e., the initial value before any wins or losses are recorded) is 50%. Note that this is merely an example of a function for normalizing the selection results. Other normalization formulas can be used.

[0048] The filtering component 25 can advantageously include components for detecting illogical or inconsistent selections by the user. For example, it can detect logical inconsistencies in the responses, such as (fruit flavor * / nut flavor) (nut flavor * / spicy flavor) (fruit flavor / spicy flavor *), where the asterisk indicates the user's selection for these three descriptor pairs. This can indicate that a particular user has difficulty discerning flavor annotations or certain types of flavor annotations. When such an inconsistency is detected by the filtering component 25, the relevant pair or set of pairs can be automatically re-prompted by sending an instruction via the adaptive control communication link 32 until consistency is found. The filtering component 25 can use established machine learning techniques to determine rules for a particular user and / or a particular food or food type, and the rules can be fed back to the 32 selection component 21 so as to adapt the selection of descriptor pairs in such a way as to refine the responses for a particular user and / or a particular food or food type for higher accuracy.

[0049] Tasters can be filtered based on their ability to consistently and reliably distinguish flavor annotations. In this case, the filter component 25 and the selection component 21 can be configured to present multiple sets of flavor descriptor pairs for the same product and evaluate their consistency according to predetermined logical and numerical rules. For example, each successive ranking of flavor annotations can be required to fall within a threshold of error in order for the selection results of the taster to be accepted for mapping into the product profile in the flavor descriptor space in the result database 8. The filter component 25 can be configured to detect logical inconsistencies between the responses of an individual user (e.g., if the user selects descriptors A, B, and C such that A > B, B > C, and C > A, then intervention can be made). In this case, the filter component 25 can automatically adapt the sequence of descriptor pairs to avoid or eliminate inconsistent responses, or reduce the weight of the responses of that user.

[0050] The various system elements referred to herein as "components" described in functional terms can be implemented as dedicated circuits or hardware elements, or as instructions stored on a data carrier or in an operating memory and executed by a processing unit, or as a combination of these.

[0051] The methods and systems described above are designed to provide an accurate and consistent quantification of product flavor. The normalized quantification of flavor descriptor information allows for the generation of scaled flavor profile information, as opposed to the declarative information of traditional tasting panel methods. This enables tastings performed by different groups of tasters to be numerically compared to one another, such that flavor descriptor profiles from different groups can be represented in the same descriptor space. Machine learning implemented in the filtering component 25 can, for example, include learning the different perceptual profiles of different tasters or groups of tasters, and can use the learned rules to automatically adjust the flavor intensity profile. If, for example, the same product is tasted by two different groups using the same or similar descriptors, the filtering component 25 can be configured to automatically learn the differences in the calculated intensity scores for each descriptor from the two different flavor quantifications. These learned differences can then be used by the filter component 25 to increase or decrease the weight of the descriptor intensity values for a particular group accordingly. This facilitates a meaningful efficiency gain in preference testing by being able to remove "irrelevant" data points based on taste perception variations. This results in a lower capacity utilization for market research activities.

[0052] Furthermore, the method of the present invention can be adapted to enable the creation of an ideal or "target" profile for a product. Compared to traditional regression models that define preference drivers, this method requires less storage and processing power due to the adaptive nature of the problem interface and the improved accuracy of preference data when combined with individual perceptual measurements. In addition to pairwise selection (i.e., binary selection) based on each taster's perception of the intensity of relevant flavor annotations, cues can also be utilized to present the selections to determine the taster's preference within each pair. Thus, in addition to "which flavor is stronger" ", questions related to the same flavor descriptor pair can be asked such as: "which flavor would you prefer to increase" " and / or "which flavor would you prefer to decrease" The selected responses to these preference pairs can be mapped and normalized in a similar manner to the flavor descriptor perceptual selection described above into the flavor descriptor space of the product being analyzed and can be used to create a target flavor profile that can be compared to the perceptual profile already described and thereby indicate what flavor changes can be advantageously made to the product to appeal to the public within the broader group represented by the participating tasters. Thus, by combining the original profiles created by the users, using the scaled measurements of each user's feedback on which annotation to increase / decrease, a target profile can be generated for each user. This is expressed as an equation where the variable is the flavor tone. The coefficients are determined via the users' feedback. The users' feedback can be filtered and aggregated to create a "group" equation with aggregated coefficients. This can be used to generate the target profile. For example, if the original profile is defined by the function F(x, y, z) and the aggregated group function is 0.3x, 0.4x, 1.2z, then the target flavor intensity for x becomes 0.3x, where x is the original intensity generated by the control group described above.

[0053] An example of the target flavor profile 9 is shown on the flavor profile space 4' illustrated in Figure 8 . In this example, the target profile 9 indicates that increasing the intensity values of the "sweet", "milky" and "caramel" flavors and / or decreasing the intensity values of the "spicy", "barbecue", "nutty" and "aromatic" flavors in the product being tasted is likely to improve the flavor.

[0054] Figure 7 A further variant of the invention is shown. In addition to presenting pairs of flavor descriptors (in this case "spicy" and "milky") for selection, additional input fields 35 can be provided for the user to enter other information that can be quantified using the flavor descriptors. It is possible to present, for example, a list of alternative flavor descriptors that are not in the set of descriptors from which the pairs are selected and allow the user to provide additional selection or preference information by selecting a descriptor from said list. Alternatively, the additional field 35 can be a free-form field for collecting additional tasting feedback from the user.

[0055] Although the present invention has been described using pairs of flavor descriptors, it is also possible to use triples or even higher order combinations of flavor descriptors for making selections. However, pairs of descriptors have been found to be preferred because they make it easier for tasters to distinguish flavor notes and the number of possible pairs is significantly fewer. However, the adaptive nature of the algorithm allows the method to determine when it is possible to present more than two flavor notes while maintaining accuracy. This way of including more than two selections is an additional efficiency gain because it reduces the returns to the server for problem determination, thereby reducing the processing load time.

Claims

1. A computer-implemented method for representing a descriptor profile of a stimulus in a multi-dimensional descriptor space and using descriptor variables capable of characterizing a predetermined stimulus, the method comprising: - A first step of providing, to a user device of each user among a plurality of users, a plurality of different descriptor pairs from a host device, wherein each pair includes descriptors representing two different dimensions of the descriptor space; - A second step of, for each user, presenting to the user, using a graphical user interface on an output device of the user device, a text prompt and one of the plurality of different descriptor pairs provided in the first step, wherein each member of the descriptor pair is presented adjacent to the other member on the graphical user interface, and wherein each member is presented in a separate selectable region of the graphical user interface; - A third step of, for each user, the user interacting with the selectable region associated with a member of the presented pair by using the graphical user interface of an input device of the user device to indicate which member of the presented pair more closely represents the user's perception of the stimulus, such that the graphical user interface of the user device receives the user's selection of a member of the presented descriptor pair that more closely represents the user's perception of the stimulus, and upon receiving the user's indication, the descriptor pair is removed from the graphical user interface; - Repeating the second step and the third step successively for each of the plurality of different descriptor pairs provided in the first step, such that the second step and the third step are performed for each of the plurality of different descriptor pairs provided in the first step; - A fourth step of transmitting the user selection to the host device using a network interface of the user device; - A fifth step of creating a training set using a previous user selection pattern, wherein the training set establishes a predetermined set of filtering rules for classifying user selections; - A sixth step of training a machine learning algorithm using the training set, wherein the machine learning algorithm is iteratively trained until it can detect when the user selection reaches a predetermined convergence level; - A seventh step of filtering out one or more of the received selections that fall outside a predetermined range by referring to the predetermined set of filtering rules and deleting the one or more identified selections; - An eighth step of, using a normalization processor of the host device, calculating, for each of the descriptors received from the user device, a descriptor value in the perception space and normalizing the value of each descriptor against other descriptor values in order to generate a normalized perception profile of the predetermined stimulus in the descriptor space.

2. The method according to claim 1, wherein The first step includes generating all possible pairwise combinations of descriptors.

3. The method according to claim 2, wherein, Generating all possible pairwise combinations is performed by a processing unit of the user device.

4. The method according to claim 1 or 2, wherein The first step includes: generating descriptor pairs at the host device and transmitting the generated descriptor pairs from the host device to the user device via a network.

5. The method according to claim 1, wherein, The first step includes: using a randomizer of the host device or the user device to provide descriptor pairs in a random or pseudo-random sequence.

6. The method according to claim 1, wherein The first step includes an optimization step that adaptively provides descriptor pairs depending on one or more of the following: - historical information of previous user selections, - previously filtered historical information of user selections, or - the current distribution of descriptor data points in the descriptor space.

7. The method according to claim 1, wherein The eighth step includes: if the probability of a descriptor being selected is higher compared to the probabilities of other descriptors being selected, weighting the descriptor with a higher descriptor value, and if the probability of a descriptor being selected is lower compared to the probabilities of other descriptors being selected, weighting the descriptor with a lower descriptor value.

8. The method according to claim 7, wherein If the received set of selections includes fewer selections compared to all of the multiple presented descriptor pairs, the method includes using the Bayesian engine of the host device to infer selection parameters for those descriptor pairs that are missing.

9. The method according to claim 1, comprising: Use the filter and / or normalization processor of the host device to determine weighting or correction parameters for at least one of the descriptors for at least one of the users, and store the weighting or correction parameters as rules of the machine learning engine.

10. The method according to claim 1, including the steps of: recording, for each user, change preference parameters for each of a first plurality of descriptors, and using a normalization processor to map the change preference parameters onto the descriptor space in such a way as to indicate the cumulative change preferences for each descriptor.

11. The method according to claim 1, comprising: Receive selections of one or more additional descriptors from each user that are not in the first plurality of descriptors, and add the additional descriptors to the descriptor space and the first plurality of descriptors.

12. The method according to claim 1, including: generating a first descriptor profile in the descriptor space; generating a second descriptor profile in the descriptor space using a first stimulus and a second plurality of users; and using a normalization processor to generate normalization mapping parameters for each of the plurality of descriptors for mapping descriptor values from the first descriptor profile to the second descriptor profile or from the second descriptor profile to the first descriptor profile.

13. The method according to claim 1, wherein: - the predetermined stimulus includes the taste or smell of food, - the user is a taster of the food, - the descriptors include flavor or aroma descriptors of the food, - the descriptor values include intensity values of flavor or aroma, and - the descriptor space is a flavor or aroma descriptor space.

14. The method according to claim 1, wherein: - the predetermined stimulus includes the color of a radiating or reflecting object, - the user is an observer of the object, - the descriptors include color descriptors of the object, - the descriptor values include color values of flavor or aroma, and - the descriptor space is a color space.

15. A system for determining the quantified intensity values of multiple predetermined flavor descriptors of an analyzed substance, the system including a host device and a plurality of user devices, wherein the host device includes: a first database containing multiple predetermined flavor descriptors of the substance; a transmission device that transmits the flavor descriptors to each of the user devices; Each in the user equipment is configured to: (i) receive a plurality of different descriptor pairs from the host device, where each pair includes descriptors representing two different dimensions of the descriptor space; (ii) present a text prompt and one of the plurality of different descriptor pairs received in step (i) to the user through a graphical user interface, where each member in the descriptor pair is presented adjacent to the other member on the graphical user interface, and each member is presented in a separate selectable area of the graphical user interface; (iii) receive the user's selection of a member in the presented descriptor pair by interacting with the selectable area associated with the member in the presented descriptor pair, the member more closely representing the user's perception of a stimulus at the graphical user interface, and when the user's selection is received, the descriptor pair is removed from the graphical user interface; (iv) sequentially repeat steps (ii) and (iii) for each of the plurality of different descriptor pairs provided in step (i), such that steps (ii) and (iii) are performed for each of the plurality of different descriptor pairs provided in step (i); A processor, configured to: Receive the flavor descriptor selection from each in the user equipment; Create a training set using a previous user selection pattern, where the training set establishes a set of predetermined filtering rules for classifying user selections; And Use the training set to train a machine learning algorithm, where the machine learning algorithm is iteratively trained until it can detect when the user selection reaches a predetermined convergence level; Filter using the trained machine learning algorithm by referring to the set of predetermined filtering rules to detect inconsistent or illogical or incomplete selection responses among the flavor descriptor selections received from one or more user equipment; For each of the first plurality of predetermined flavor descriptors, calculate a flavor descriptor intensity value in the flavor descriptor intensity space according to the selection response of each flavor descriptor, and normalize the flavor descriptor intensity value against the flavor descriptor intensity values of substantially all other flavor descriptors in the first plurality of flavor descriptors in the flavor descriptor intensity space.

16. The system according to claim 15, wherein, The machine learning algorithm is configured to determine rules for weighting or correcting or deleting selection responses in the selection responses received from the user equipment.

17. The system according to claim 15, wherein, The filter includes an adaptive control output signal, and the adaptive control output signal is used to transmit control instructions to the selector to modify the sequence of descriptor pairs selected by the selector depending on the anomalies or inconsistent conditions detected by the filter in the selection responses.

Images