Secure image generation for health and wellbeing

Abstract

A computer-implemented therapy tool receives a secure script from an end user display device, the secure script comprising information about the end user. The tool uses the secure script to generate a response from a generative machine learning model, the response comprising an image or a vector encoding an image. The tool sends the response to the end user device to trigger display of the image at the end user device in order to promote health and wellbeing of the end user.

Description

SECURE IMAGE GENERATION FOR HEALTH AND WELLBEING

[0001] The present disclosure relates to generating images to promote health and wellbeing of an end user, such as an end user with dementia, depression, low self-esteem or other mental health condition, in a secure manner. The disclosure is particularly related to, but not limited to, an image-led cognitive behavioural therapy tool and interactive platform to deliver measurable uplifts in an individual’s health and wellbeing.BACKGROUND

[0002] Many countries are experiencing an increase in the proportion of the population who are older and living with conditions such as memory loss, dementia, anxiety, bereavement, chronic pain, frailty, social isolation, loneliness, loss of identity and other conditions. In some cases residents in care homes for the elderly become withdrawn and difficult to communicate with. Mental health concerns are also prevalent among younger age groups.

[0003] Research has shown that behavioural and lifestyle changes can be stimulated by an emotional connection. Behavioural and lifestyle changes can lead to positive improvements in both physical and mental health, which in turn increase longevity and quality of life. This is highly relevant to aging populations, such as in the UK, where elderly individuals are particularly vulnerable to social withdrawal. Social withdrawal can lead to a cycle of social isolation wherein an individual, through lack of interaction with others, becomes more withdrawn and less likely to initiate interaction.

[0004] The examples described herein are not limited to examples which solve problems mentioned in this background section.SUMMARY

[0005] Examples of preferred aspects and embodiments of the invention are as set out in the accompanying independent and dependent claims.

[0006] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

[0007] A first aspect of the disclosed technology is a computer-implemented method comprising: receiving a secure script from an end user display device, the secure script comprising information about the end user; using the secure script to generate a response from a generative machine learning model the response comprising an image or a vector encoding an image; and sending the response to the end user device to trigger display of the image at the end user device in order to promote health and wellbeing of the end user.

[0008] The disclosed technology provides a method of invoking an emotive response from an individual in order to promote health and wellbeing, especially where the individual is suffering from dementia or memory impairment. The technology is also useful for promoting health and wellbeing of those suffering from low self-esteem, depression and other mental health conditions across a wide spectrum of age groups. The information about the end user can include a variety of information such as age, gender, ethnicity, place of birth, place of habitation, occupation or general interests. In some cases the information about the end user includes personal highlights comprising information about positive events or achievements in the life of the end user. In some cases the information about the end user includes results of an assessment of cognitive ability, where the assessment may be carried out immediately prior to the method using the secure script. By using a secure script the information about the end user is kept secure whilst also allowing images depicting content pertinent to the end user to be generated.

[0009] Preferably, the secure script comprises information about the end user in the form of category data. Beneficially, category data reveals less about the end user than specific values and may be used to form the secure script.

[0010] Preferably, the secure script comprises a vector encoding of the information about the end user. Representing the information as a vector is one method of keeping the information about the end user secure; since the vector is not understandable by a party without access to a decoder for decoding the vector into the information about the end user. The vector encoding is concise and reduces the amount of bandwidth needed to send the information and reduces the amount of memory needed to store the information. Preferably the vector encoding of the information about the end user is in an embedding space of the generative machine learning model. This gives efficiency since the vector is able to be input to the generative machine learning model directly without any mapping or transformation.

[0011] Preferably, the secure script is sent to the generative machine learning model over a public communications network. In such a way, the provision of the script to the generative machine learning model can be carried out in any location where such communications networks are available. For example, the script can be sent over a mobile phone network.

[0012] Preferably, the secure script comprises an encoding of an image. This is useful where an end user has a favourite image and wants to see more images like that.

[0013] Preferably, the generating of the script comprises selecting key words from a library of key words, the selecting taking into account the information about the end user. The key words can be representative of key interests regularly found in common between individuals who are similar. For example, category data corresponding to individuals whoare from a certain city or country may return key words related to landmarks in that city or country.

[0014] Preferably receiving the secure script at a therapy tool from a client device comprises receiving the secure script in plaintext. This is possible where the secure script comprises a vector since malicious parties who intercept the secure script are unable to understand it. No encryption is needed in this case which improves efficiency.

[0015] In another aspect of the technology there is an apparatus comprising: a processor; a memory storing instructions which when executed on the processor implement a method comprising: receiving a secure script from an end user display device, the secure script comprising information about the end user; using the secure script to generate a response from a generative machine learning model, the response comprising an image or a vector encoding an image; and sending the response to the end user device to trigger display of the image at the end user device in order to promote health and wellbeing of the end user. Display of the image at the end user device facilitates emotional engagement of the end user. The emotional engagement leads to a more positive state of mind, readying the individual for ensuing stimulation, and identifying goals / focus for healthy behavioural activities.

[0016] In another aspect of the technology there is a client device comprising : a display ; a processor ; a memory, the memory storing instructions which when executed on the processor implement a method of: receiving data about the end user; generating a secure script using the received data; sending the secure script to a therapy tool; receiving a response from the therapy tool, the response comprising an image or a vector encoding an image; and displaying the image at the display. The client device is secure and securely interoperates with the therapy tool to provide a service to the end user. The client device does not need to trust the therapy tool since the script is secure.

[0001] Preferably the image or the vector encoding the image have been computed by a generative machine learning model in response to the secure script. This gives the benefit that the image is pertinent to the end user and is likely to promote health and wellbeing of the end user.

[0002] Preferably sending the secure script to the therapy tool comprises sending the secure script in plaintext form which is efficient as compared to using encryption.

[0003] Preferably the client device comprises an encoder and generates the secure script using the received data by encoding the received data using the encoder. This gives an efficient practical way to make the secure script.

[0004] Preferably the client device comprises a decoder configured to decode the vector encoding the image into an image. This gives efficiency since the image does not need to be sent and is much larger than the vector.

[0005] Preferably the client device is arranged so that receiving the response from the therapy tool is contingent on receiving information at the client device about outcome of all or part of a game. The information is received via user input, via a game executing on the client device, by wireless transmission from another computer or in other ways. Thus the images are able to act as rewards as part of a gamification of behaviours that promote health and wellbeing. In some examples, an application on the client device includes games which may be tailored to the physical and cognitive abilities of the end user. The games can be based on the profile of the user. In non-limiting examples, the games can be a spot the difference game or a memory based game with the themes identified as passion points of the end user in order to promote memory and cognitive ability. For example, the gamification could be a game of pairs where the user has to select matching images of boats. In some examples, a reward is provided to the end user after successful completion of the game. The reward may be a new, previously unseen image, or an amination related to the passion point behind the game. For some users, the game may include physical challenges such as walk 200 steps today, or challenge another users to a fitness or strength challenge. In such a way, the application on the client device can encourage positive changes in activity levels of end users both in terms of physical and mental activity. Further, particularly where the game involves additional users, this has the additional benefit of social interaction which has been shown to reduce withdrawal and symptoms of depression. Another way in which the application can provide gamification is physical games within a residential setting. For example, a fishing game could be introduced within a care home. Further, where individuals have expressed an interest in baking via the therapy tool, baking activities can be introduced.

[0006] Preferably the client device comprises: a camera and emotion recognition functionality; or a graphical user interface arranged to receive the data about the end user by receiving any one or more of: a rating of a temperament of the end user, a rating of a level of cognitive ability of the end user, an age category, a hobby, a topic of interest.

[0007] It will also be apparent to anyone of ordinary skill in the art, that some of the preferred features indicated above as preferable in the context of one of the aspects of the disclosed technology indicated may replace one or more preferred features of other ones of the preferred aspects of the disclosed technology. Such apparent combinationsare not explicitly listed above under each such possible additional aspect for the sake of conciseness.

[0008] Other examples will become apparent from the following detailed description, which, when taken in conjunction with the drawings, illustrate by way of example the principles of the disclosed technology.BRIEF DESCRIPTION OF THE DRAWINGS

[0009] FIG. 1 illustrates schematically interaction between an end user and the imagebased cognitive behavioural tool;

[0010] FIG. 2 is a flow chart of a session of using the image-based cognitive behavioural tool

[0011] FIG. 3 is a flow chart of data management within the image-based cognitive behavioural tool;

[0012] FIG. 4 is a schematic diagram showing a library of the image-based cognitive behavioural tool;

[0013] FIG. 5 is a schematic drawing of an interaction of an example end user with the image-based cognitive behavioural tool;

[0014] FIG. 6 is a schematic drawing of a further interaction of the example end user with the image-based cognitive behavioural tool

[0015] FIG. 7 is a schematic drawing of the architecture of an example image-based cognitive behavioural therapy tool.

[0016] The accompanying drawings illustrate various examples. The skilled person will appreciate that the illustrated element boundaries (e.g., boxes, groups of boxes, or other shapes) in the drawings represent one example of the boundaries. It may be that in some examples, one element may be designed as multiple elements or that multiple elements may be designed as one element. Common reference numerals are used throughout the figures, where appropriate, to indicate similar features.DETAILED DESCRIPTION

[0017] The following description is made for the purpose of illustrating the general principles of the present technology and is not meant to limit the inventive concepts claimed herein. As will be apparent to anyone of ordinary skill in the art, one or more or all of the particular features described herein in the context of one embodiment are also present in some other embodiment(s) and / or can be used in combination with other described features in various possible combinations and permutations in some other embodiment(s).

[0018] The inventors have recognized there is a growing need for technical tools to facilitate communication with elderly individuals such as in residential care settings. The inventors have developed a technical tool which generates images that invoke a response in an individual and may facilitate interaction between a carer and the individual.Qualitative results of using the tool indicate that it is extremely successful in facilitating interaction, health and wellbeing, even for individuals who have otherwise been withdrawn and unresponsive for many weeks. The technical tool may be used by professional carers working in care and nursing homes as well as in domestic settings. The technical tool may be used by unpaid carers looking after relatives in their later years and helping them to live longer, healthier and happier lives, ideally by remaining in their own homes thereby reducing pressure on public health services. The technical tool is also useful to facilitate communication or to promote health and wellbeing of end users in other age groups, including but not limited to those suffering from mental health conditions such as depression and low self-esteem. In some examples the technical tool is used without a carer being present.

[0019] The inventors have developed a technical tool that uses synthetically generated images generated using a generative machine learning (ML) model such as DALL-E, Stable Diffusion, Midjourney or others. Using generated images (as opposed to other modalities of content) is found to be particularly effective. Research has shown that the mechanisms used by the brain to access memory decline at different rates and to different extents. In particular, the mechanisms used to retrieve memories when they are linked to images decline at a slower rate compared with mechanisms used to retrieve memories linked to other stimuli. This results in the ability to retrieve memories linked to images being similar between young adults and older adults. Images are also a universal language, unlike text and speech which are language dependent, and have been shown to be more effective at triggering an emotional response. However, in order to invoke such a response, a working knowledge of an individual’s interests, passions and history is required such that relevant images can be provided. Thus various technical challenges are encountered when developing a technical tool for presenting generated images. The technical challenges include but are not limited to: security, scalability, latency reduction, usability.

[0020] Identifying points of interest or passion in an individual who may be, on occasion, non-verbal for a variety of reasons can be challenging. This is particularly relevant in a care environment where carers and other medical professionals, at least initially, do not have a long standing or personal relationship with the individuals. Without this valuable insight in to an individual’s preferences and history, medical professionals can struggle to engage with an individual and miss opportunities to provide them with vitally important mental stimulation and interaction. The technical tool described herein makes it easier for different carers to maintain a more consistent mode of positive communications.

[0021] Figure 1 shows an image-based cognitive behavioral therapy tool 100 (referred to as therapy tool 100) deployed as a cloud service. In Figure 1, an end user 102 is viewing an image on a hand held client device 110, accompanied by a carer 104. It will beunderstood that the carer 104 can also be a medical professional, friend, family member, social worker or other connection of the end user 102. Where the end user is a child, the carer 104 can be a teacher or support worker. Although the example of figure 1 shows a hand held client device 110 it is also possible to us a large display screen such as to facilitate communication with more than one end user at the same time. Although the example of figure 1 shows a carer 104 it is not essential for a carer 104 to be present.

[0022] The therapy tool 100 is computer implemented and in some examples is deployed on one or more web servers or other computing entities in communication with the client device 110. The client device 110 is preferably portable or hand held. In some cases the client device 110 is a wearable computer such as a smartwatch. In some cases the client device 110 is a smart phone or tablet computer. The client device 110 is able to communicate with the therapy tool 100 over a wired or wireless connection such as via the Internet or other communications network. In some cases a secure connection such as using HTTPS other secure communications network technology is established between the client device 110 and the therapy tool 100. But this is not essential as in some examples the client device 110 communicates with the therapy tool using a public connection. The client device 112 optionally has an encoder 116 and / or a decoder 118 as explained in more detail below.

[0023] The therapy tool 100 is connected to or has access to a generative ML model 114 such as via an application programming interface API. The generative ML model is functionality to generate images which may be photorealistic images or images of other styles. The generative ML model 114 is able to receive a text input and generate an image depicting content semantically related to the text input. Any suitable generative ML model 114 is used and a non-exhaustive list of example generative ML models for generating images is: DALL-E, Midjourney, Stable Diffusion, Fotor Image Editor, Pareto, Adobe Firefly, Microsoft Bing Image Creator, Microsoft Designer.

[0024] The generative ML model 114 comprises an optional encoder 122 for encoding a text input into a vector in a multi-dimensional space. The generative ML model 114 comprises a decoder 120 for decoding a vector in the multi-dimensional space into an image. The generative ML model may be a neural network that has been trained using images with known associated text describing content depicted in the images.

[0025] The therapy tool 100 receives a secure script 106 from an end user display device such as client device 110. The secure script comprises information about the end user 102. The therapy tool 100 uses the secure script 106 to generate a response 108 from a generative machine learning model 114, the response comprising an image or a vector encoding an image.

[0026] The therapy tool 100 sends the response 108 to the client device 110 to trigger display of the image at the client device 110 in order to promote health and wellbeing ofthe end user 102 such as through emotional engagement or connection with the end user. In some cases the generative ML model 114 generates an image which is in the response 108. In some cases the generative ML model 114 generates a vector encoding of an image. The vector is sent to the client device in response 108 and a decoder 118 at the client device 110 decodes the vector into an image.

[0027] Because the image is generated using the information about the end user the image depicts content which is pertinent to the end user and is found to provoke an emotional response in the end user 102. The end user 102 and the carer 104 are then able to have a conversation triggered by the image such as to recall events in the life of the end user 102 related to the image. In situations where the carer 104 is not present the end user 102 is able to reflect on their memories. The end user 102 (either when the carer 104 is present or not) is facilitated to perform health and wellbeing promoting behaviours such as via gamification as described in more detail below.

[0028] Because the script 106 is secure the information about the end user 102 is kept secure despite being sent over a communications network to the therapy tool. In some examples, the script 106 may be sent over a public connection to the therapy tool whilst retaining security. In some cases the script 106 is secure with respect to the therapy tool 100 provider; that is, a provider of the therapy tool 100 is unable to access the information about the end user 102 used to form the secure script 106.

[0029] The script is secure by using one or more of the following methods:• Mapping the information about the end user 102 to category data so as to not contain specific values of variables such as age and rather contain an age category (such as over 80 years).• Use of encryption.• Encoding the information about the end user into a vector using an encoder, where the vector is in a multi-dimensional space of the generative ML model 114.

[0030] As mentioned above, the client device 112 optionally has an encoder 116 and / or a decoder 118. The encoder 116 is a neural network encoder for encoding information about the end user 102 into a vector for input to the generative ML model 114. A non- exhaustive list of examples of information about the end user 102 is: age, gender, ethnicity, place of birth, citizenship, level of cognitive function, medical conditions, hobbies, significant geographical places in the end user’s life history, significant occupations in the end user’s life history, skills. The information about the end user may comprise topics of interest to the end user. In an example the encoder is a neural network language model such as BERT, RoBERTa or any other neural network language model that computes contextualised embeddings. The decoder is a neural network configured to receive a vector and compute an image. In an example the decoder is a decoder froma diffusion model such as the open source Stable Diffusion model or other diffusion model for generating images.

[0031] The vector produced by the encoder 116 may be sent by the client device 112 to the therapy tool 100 over a communications link of any type. Because the vector is a concise representation the bandwidth needed to send the vector is reduced as compared with sending the full information about the end user 102. Thus efficiencies are gained in terms of bandwidth of communication link between the client device 110 and the therapy tool.

[0032] Because the vector produced by the encoder 116 is meaningless to the therapy tool, the end user 102 is able to make use of the therapy tool 100 service without revealing their full information such as age, medical conditions, memory function level etc. The vector produced by the encoder 116 may be sent to the therapy tool 100 without encrypting the vector in some cases, since the vector alone is meaningless to a malicious party who intercepts the vector (the vector is a list of numbers).

[0033] Where the response 108 comprises a vector encoding of an image bandwidth efficiencies are gained as compared with sending the image itself in the response 108. This is particularly useful where the client device 110 is used in a domestic setting with a low speed internet connection.

[0034] In some cases the information about the end user 102 is mapped to categories or clusters rather than being encoded using encoder 116. In these cases encoder 116 at client device 112 is omitted and an encoder 122 at the generative ML model 114 encodes the information about the end user 102 from the therapy tool 100.

[0035] More details about how the information about the end user 102 is obtained is now given. In some cases the information about the end user 102 is input to the client device 112 and stored in memory in the client device 112. Preferably, the computing device comprises a specific application for receiving the information about the end user, which may be category data. The information about the end user is entered in category format in some cases and can include end user specific data such as age, gender, occupation, nationality or place of residence. It will be understood that this list is non-exhaustive. The category data, or other information about the end user 102, can also be captured in responses to questions. For example, the carer (or the computing device where there is no carer present) can ask the end user what their interests are and they could indicate that they like bird watching or baking. In figure 1 , the carer 104 is shown inputting category data to the computing device 110. In various examples the questions are pre-prepared by a neurologist and / or psychologist. Where there is no carer 104 present the answers to the questions are input by the end user.

[0036] In some examples, a profile is created for the end user within an application on the client device 112. At initial set up, information about the end user 102 comprising datasuch as age, gender, ethnicity or place of residence can be added to the profile. This data can be provided by the end user themselves, or by a carer, friend, relative or other connection. The initial input can create the category data described above. As more is known about the end user, the profile can be updated to include more details such as general interests. In particular, this additional data can be gleaned from the end user’s interaction with therapy tool 100.

[0037] In some cases there is a plurality of clusters of end users 102 where the clusters have been determined using a clustering algorithm such as k means. Each cluster has a plurality of characteristics. The characteristics of the clusters may be known to the client device 112 so that rules at the client device 110 may be used to assign the end user 102 to one or more of the clusters. Identifiers of the clusters to which the end user 102 is assigned can then be included in the secure script. For example, men in a certain age bracket with a particular nationality could be assigned to one cluster and women of the same age bracket and the same nationality could be assigned to a second cluster. A third cluster could exist which includes both the first and the second group based on age and nationality. In such a way a single profile may be assigned to a plurality of clusters.

[0038] In some examples images are generated by the generative ML model 114 in advance and stored in a library of images. This helps reduce latency when responding to a secure script 106. The images within the library comprise metadata describing the images. The metadata is either manually completed labelling or captioning computed using an automated image captioning service. The labelling or captioning may for example comprise key words which are assigned to the images based on their content. For example, the key words can be objects, places, themes or dates.

[0039] In some examples, the therapy tool 100 selects a batch of images from the library of images to supply to the client device 112. At least some of the images can be selected based on the category data. For example, the category data used to generate the secure script includes that the end user likes bird watching. The therapy tool 100, on receiving the secure script 106 and consulting the metadata of the images within the library, identifies images of birds.

[0040] Preferably, each image is carefully selected to minimise risk of triggering a negative response.

[0041] In some examples, the images selected based on matching with the category data are a portion of the batch of images supplied to the end user. For example, 50 images can be supplied to the end user wherein 15 of those images are selected from the matching with end user indicated preferences, 15 are selected based on other category data such as age, ethnicity or city of habitation. The final 20 images can be selected at random. It will be understood than any ratio of images is possible, and tailored depending on the initial data available. For example, wherein the end user has not been able to input theirgeneral interests, a higher portion of the images can be selected on a random basis. Similarly, wherein a profile has been developed over several interactions with the therapy tool a higher portion of the images displayed can be based on the known interests, or passion points, of the user.

[0042] In some examples, a selected initial batch of images is shown to the end user. The end user provides feedback on the images to the client device 112. The client device 112 can record this feedback, optionally against the end user’s profile. A variety of methods can be used to receive feedback from the end user.

[0043] One such method comprises the end user assigning each image a score out of, for example, 5. In such cases, the end user will assign a 5 to a favourite image, or 1 to an image they are least interested in. The carer can input this selection on the end user’s behalf. In order to assign the score, a number could be selected on a screen of the computing device or spoken verbally by the end user or carer and picked up by a microphone of the computing device. Another method comprises the end user being prompted to select a subset of preferred images form the provided batch of images. In some instances, the carer can select the favourite images or assign scores based on the perceived reaction of the end user, such as via hand signals or facial expressions. In other examples, the client device 112 comprises facial recognition software which can recognise a positive or negative response from the end user’s facial movements and identify preferred images accordingly. Yet another method of receiving feedback from the end user includes monitors which measure physiological changes in the end user. For example, a heart rate monitor can indicate an increased heart rate upon viewing a preferred image. Similarly, an accelerometer on a end user’s wrist can supply feedback by measuring hand gestures. For example, the end user may wave when they see a preferred image.

[0044] The therapy tool receives feedback from the client device 112, indicating the end user’s response to the supplied images. The feedback may be detected by sensors in the client device. The feedback may be from the end user and / or the carer. The therapy tool matches the feedback with the meta data of the preferred images. In some examples, the therapy tool provides, as an input to the client device, the meta data of the preferred images. The meta data of the images can be supplied directly to the client device from the therapy tool. Alternatively, key themes can be identified from the meta data, and said themes supplied to the client device for inclusion on the profile of the end user. The updated profile can then be used to provide an updated secure script 106 to the therapy tool 100 for a subsequent session of the end user using the client device. A session begins when the end user logs in to the client device 112, and ends when the end user logs out of the client device 112. An interaction refers to a complete loop between the end user and therapy tool 100 and comprising the therapy tool receiving a secure script 106, thetherapy tool supplying a response 108 to the client device 112 and the therapy tool 100 receiving feedback from the end user, or carer, via the client device. A single session can comprise multiple interactions between the client device 112 and the therapy tool 100. In some examples, the profile of the end user, also referred to as a digital identity of the end user (which may include their skills) is updated between interactions within a single session.

[0045] During a subsequent session, the therapy tool 100 receives an updated secure script 106. The updated secure script 106 is used to inform selection of a further batch of images to be shown to the end user. As set out above in relation to the initial batch of images, the profile of the end user is used to select a subset of the further batch of images supplied. The end user’s feedback is then provided to the client device, as outlined above in relation to the initial batch of images. In such a way, an iterative process is formed wherein the profile of the end user becomes more informed with each interaction or session. A subset of the images within the batch of images is selected based on the feedback provided during previous sessions. As more information is known about the end user, and the profile updated accordingly, a higher portion of the supplied images can be based on the end user’s profile, via an updated secure script 106. In such a way, the end user is likely to receive a set of images containing more images which incite a positive response from the end user.

[0046] The subsequent sessions can be days, weeks or months apart. Each session could comprise several interactions wherein the end user receives a new batch of images in each round and provides their feedback as set out above. Th secure script is then updated between rounds, in the same way as between sessions.

[0047] In some examples, a subsequent batch of images is provided to the end user which is not informed by end user feedback and is instead selected in the same way as the first batch of images. For example, wherein the initial profile did not indicate personal interests, and there were none, or only a few, preferences indicated from an initial round the subsequent batch of images can be supplied to offer a broader range of subject matter. In some examples a subsequent batch of images will contain a higher portion of random images, wherein random refers to images that are not selected based on a cluster to which the end user belongs. In such a way, this prevents ‘pigeon-holing’ of end users and ensures personalised care is provided.

[0048] In some examples, the therapy tool collates the feedback from a plurality of end users using the same or different client device 112. As described above, the end users can be assigned to clusters. Wherein a plurality of end users belonging to the same cluster indicate a preferences for a particular theme or passion point, identified via the key words, future end users allocated to this cluster will be more likely to receive images related to this passion point. The therapy tool 100 can comprise a threshold value wherein only oncea certain number, or percentage, of profiles within a cluster indicate a preference for a particular theme or passion point, represented by the key words, the particular theme or passion point becomes more likely to be shown to individuals of that cluster. Similarly, once this threshold is exceeded additional weighting can be applied to this preference for that cluster as more individuals of that cluster indicate the same preference. In such a way the therapy tool 100 identifies common interests within clusters and uses this insight to improve the selection of images presented to end users. For example, the therapy tool may identify a specific age group of women from a given city all share an interest in sewing. The therapy tool can therefore include at least one picture related to sewing in a batch of images, and in particular the first batch of images, provided to an end user who is a women of the same age group from that particular city. Note that sewing is one example only and is not intended to be limiting. Other topics such as classes, tapestry, visiting museums, holidays, presents, celebrations, grandchildren and others are used in some examples.

[0049] The therapy tool 100 can identify the themes or passion points referred to above using the meta data of selected images, and in particular the key words assigned to the images. For example, images within the library may be tagged related to hobbies or skills such as fishing, sewing, baking, music, instruments, singing or football. The therapy tool 100 is therefore able to identify that certain groups of individuals share an interest in one of these hobbies. The groups can be represented as clusters by the therapy tool. As the therapy tool 100 receives more data, weighting can be applied to the identified preference. For example, a higher number of individuals of the same group who share a preference will increase the weighting. Preferences with a higher weighting are more likely to be used in the selection of images to be shown to the end user. Particularly strong preferences, indicated by a high percentage of a high number of individuals within the group, could also be used to inform peripheral groups. Peripheral groups refer to groups that share some but not all characteristics of a specific group. For example, individuals of the same gender and nationality but of an adjacent age group. In such a way, the therapy tool is continuously improving the images offered to individuals, allowing the iterative process of identifying passion points to be streamlined.

[0050] In some examples, the profile of an end user is used to generate a digital identification for the end user. The digital identification can include passion points, identified through the positive feedback received during interaction between the end user and the therapy tool. The digital identification provides an insight into what is important and interesting to the end user. One benefit of the digital identification is that it can be provided to family, friends or other visitors to use as conversation starters when spending time with the end user. This is particularly valuable wherein the end user is withdrawn or otherwise not engaging with others. Another advantage of the digital identification is thatit can be provided to medical professionals upon first meeting an end user. This can encourage positive interactions between medical professionals and the end user, which can lead to better care and understanding. A further advantage is that the digital identification can follow the end user to a new location. For example, if the end user is in care and is moving between residences, the new residence can immediately start to engage with the end user upon arrival. This is particularly relevant wherein an end user may suffer from dementia or mental health challenges and resultantly can find the process of moving more unsettling than other individuals.

[0051] The examples discussed herewith predominantly refer to the use of images. However, in each example, displaying the images could be augmented by interaction with other senses, such as sound. As with the images supplied above, the sounds and music can also be labelled or captioned. For example, the sound of a train could share meta data with pictures of trains. This is particularly relevant wherein a end user is partially sighted. Songs can be particularly useful in stirring up positive memories and associations for end users. For example, a first record purchased or a first dance song from a end user’s wedding.

[0052] FIG. 2 is a flow chart of a method performed by a client device 110 such as that of FIG. 1 and by a therapy tool 100 such as that of FIG. 1. The operations performed by the client device are in a left hand column of FIG. 2. The operation performed by the therapy tool 100 are in a right hand column of FIG. 2. The relative vertical position of the operations on the page represents chronological order.

[0053] At the client device 100 information about the end user is received. In an example, the carer 104 records 202 a temperament of the end user by tapping or clicking icons showing emotive faces. This can be done on a graphical user interface of the client device 110, or recorded separately by the carer. The emotive face icons can show a wide range of emotions such as happy, sad, excited or angry. The emotive faces can also offer variation within these emotions such as “very happy” or “slightly angry”. Additionally or alternatively, the carer can record the temperament of the end user using a number or words. For example, 1 can be entered to indicate a very low mood and 10 can be entered to indicate a very high mood. Similarly, words such as “low”, “agitated”, “happy” or “content” could be entered to indicate the temperament of the end user 102. Said words could be selected from options on the screen. In such a way, consistent language is encouraged which can more easily compare the temperament of the end user over time. The temperament of the end user can be recorded during a particular session, or recorded and stored over a longer period to track changes in overall health and wellbeing. For example, the temperament of the end user can also be recorded outside of a session with the therapy tool to give a picture of how the end user’s general wellbeing is impacted on a day to day basis. In some examples, a sleep tracker is used to monitor sleep quality ofthe end user. Improved sleep quality is linked to improved mental wellbeing. Similarly, an activity tracker can be used to monitor an increase in the end user’s activity which can be associated with improved mental and physical health. In such a way, an application on the client device 110 can be used to generate a timeline which tracks the end user’s wellbeing. Beneficially, as the profile stays with the end user, the timeline is not broken by a change in the end user’s location. For example, if the end user moves residence, the timeline can continue.

[0054] Operations 204 to 212 are optional and so are shown with dotted lines in FIG. 2. Operations 204 to 212 enable information about the end user to be obtained.

[0055] A plurality of images are displayed 204 on a screen of the client device 110. The images may be selected at random from a library of images. The client device 110 receives 206 a selection of a subset of the images by user input from the end user. The end user 102 can select the images by a touchpad of the client device, or pressing the screen of the client device. The end user 104 can also indicate preference orally by stating a number corresponding to the image they wish to select. Alternatively, the end user 102 can indicate their preferences to the carer 104 who can input the selection on their behalf.

[0056] The selection of images is provided to the therapy tool 100. The therapy tool 100, via the client device 110, then queries 208 the end user with questions regarding the selected images. The questions can for example ask the end user why they chose a particular image or, if the picture was of a particular place, ask if they have visited the location. Questions carefully pre-prepared by professionals may lead to specific activities and clinical outcomes including improvements in health. Alternatively or additionally, the end user 102 can be provided with facts about the image. For example if the image shows a boat the end user could be provided with facts about the boat such as the make, model, top speed or years in service. The end user 102 responds to the question and their response is recorded 210 by the client device 110. The question could be a multiple choice question, in which case the end user 102 can select the answer by pressing the screen or stating the selection verbally. The carer 104 can also supply this input on the end user’s behalf. Wherein the question is open ended, the response can be typed in by the end user 102 or carer 104, optionally using speech to text software installed on the device, or can be a voice recording of the end user 102 or carer 104 speaking. In some instances, the question or fact provides a talking point for the end user 102 and carer 104. Once the end user has answered the question, they can request additional questions via the computing device or choose to end the session. The end user’s temperament is then rerecorded 212 through a selection of emotive faces by the carer 104. Any of the examples described in relation to recording 202 the initial temperament can also be applied to rerecord 212 the end user’s temperament. The temperament of the end user 102 can also be recorded throughout the session. The temperament of the end user 102 can also be recorded usingfacial recognition software which allows the computing device to record the temperament of the end user 102, without input from the carer 104.

[0057] In some examples, the images provide prompts for conversation between the end user and others such as their family, friends, carers or other end users. At the end of a session, the client device can display the end user’s favourite images and the end user can select at least one of the images for conversation. For example, the carer 104 could ask the end user 102 why they selected the particular image, or if it made them remember a particular time. In such a way, the end user is provided a prompt to engage in social interaction with another individual on a topic of interest to them. As discussed above, social interactions have been shown to provide a positive impact on the mental health of individuals.

[0058] The client device 110 computes 214 a script. In an example the client device 110 computes the script by forming categories of data received at operation 200 into a message. Where operations 204 to 212 are performed the client device 110 computes 214 the script so the script comprises metadata of images selected by the end user at operation 206 or key words identified in responses at operation 210.

[0059] In another example the client device 110 computes the script by computing a vector encoding any one or more of: the information received at operation 200, metadata of images selected by the end user at operation 206, keywords identified in responses at operation 210. In this case the script is a message containing the vector.

[0060] The client device 110 sends 216 the script to the therapy tool 100 as indicated by the horizontal arrow from operation 216 to operation 218 in FIG. 2. The script is sent over a wired or wireless communication link to the therapy tool 100. In the case the script comprises only a vector and no confidential information, the script is sent in plaintext to the therapy tool over any communications link (secure or public). This gives efficiency since there is no latency or processing overhead introduced by encryption. In the case the script comprises category data the script may be encrypted and / or sent via a secure communication link to the therapy tool 100.

[0061] The therapy tool receives 218 the script. The therapy tool 100 inputs 220 the script to a generative machine learning model by using an API to the generative machine learning model or by inputting the script to an input field of a graphical user interface of a web service providing the generative ML model 220. In the case the script comprises a vector this is input to the generative ML model using an API command indicating the script is a vector so that the generative ML model does not need to encode the script. In the case the script comprises text being category data, keywords or other information about the end user, the script is input to the generative ML model using an API command indicating the script is to be encoded using an encoder 122 of the generative ML model 114.

[0062] The therapy tool 100 receives 224 a response from the generative Al model. The response comprises either an image or a vector encoding an image. In the case the response comprises an image, the decoder 120 of the generative ML model 114 has been used to decode a vector predicted by the generative ML model 114. In the case the response comprises a vector, the decoder 120 of the generative ML model 114 has not been used to decode the vector. The therapy tool 100 sends the response to the client device 110 as indicated in FIG. 2.

[0063] The client device 110 receives 226 the response from the generative ML model. There is an optional decoding operation 228 performed by the client device 110 in the case the response is a vector. The client device 110 then displays the image 230 which has been received or which has been decoded from the received vector. In some cases the client device 110 comprises a large display screen so that a plurality of end users are able to view the display at the same time.

[0064] Figure 3 is a flowchart illustrating a method of data processing according to an example of the technology. A end user profile is created 302 by the client device. In particular, the end user profile is created in an application in the client device. The end user profile can include characteristics such as name, gender, date of birth, current residence and nationality. Areas of interest or hobbies of the end user can also be included in the profile. In some examples, the profile also includes an indication of cognitive ability. The cognitive ability can also be assessed by standard tests, carried out within the application. The data can be input to the client device by the end user themselves or a carer, relative, friend or other connection. The personal data is stored locally and securely on the client device. This gives the benefit that the end user knows their data is held securely on the local client device which is physically secure within the home of the end user.

[0065] As shown in Figure 3, a secure script is generated 304 by the application on the client device. The generation 304 of the secure script comprises de-personalising the data of the end user. In some examples, the application comprises rules which replace personal identifiers within the data with masks. Additionally, or alternatively, the application generalises the data. For example, wherein an end user has provided a date of birth this can be converted to an age category. Similarly, wherein the end user has provided a specific address, this can be generalised to a city or council authority. In such a way, the key elements of the data which allow the therapy tool to select images is provided to the therapy tool, without sharing personal identifiers.

[0066] Upon receipt 306 of the secure script, the therapy tool allocates 308 the end user to at least one cluster. The cluster can vary in specificity. For example, the end user could be added to one cluster for gender, one for age range and one for city. Additionally or alternatively, the end user can be added to a cluster for any combination, including all, ofthe characteristics provided in the profile. Based on the cluster allocation, the therapy tool 100 is able to select 310 a batch of images for display to the end user. The therapy tool selects 310 at least one image, preferably a plurality of images, from a library of images. The library is explained in more detail below, in relation to Figure 4. Each of the images within the library comprises meta data. The therapy tool can match the assigned clusters with the meta data of the images for selection. For example, if one of the characteristics is that the end user is interested in fishing, images which are tagged or categorised as related to fishing can be selected for displaying to the end user. The therapy tool can also select images from adjacent categories such as marine life or nautical themes. For example, the key words could be assigned to larger groups which contain multiple categories from which the therapy tool can identify images.

[0067] The therapy tool can match images based on characteristics which the end user has in common with historical end users. For example, wherein the end user is allocated to a cluster corresponding to their age group, the therapy tool can select images which historical end users of the same cluster have provided positive feedback on. Similarly, the therapy tool can identify that certain categorisation of images, identified through the meta data of the images, often receives positive feedback from specific clusters. The therapy tool can employ threshold values for use in identifying categories of interest for specific clusters. For example, the therapy tool can identify that a category of interest has received over 60% positive feedback from a sample size of greater than 100 end users within a certain cluster and select images from this category of interest to include in the batch of images. A weighting can also be applied to this selection. For example, wherein the sample size is smaller, or the percentage of positive feedback is smaller the likelihood, or number, of images from that category being included in the batch supplied is less than it would be for a larger sample size or higher percentage of positive feedback. In such a way, as more end users from specific clusters provide feedback and data, the therapy tool can more accurately predict the images which will receive positive feedback from a end user. This has the advantage of requiring less computing resource, due to the fewer number of interactions required to identify passion points as well as less time and energy of the end user. This is particularly relevant where the end user may have a reduced attention span due to a medical condition, or difficulties in staying focussed on a single task for a prolonged period of time. Additionally, wherein the end user is completing the task with an assistance of a carer, the time available for completing a session with the therapy tool can be limited. The increased efficiency offered by the therapy tool in identifying images using the described method ensures the time of the carer is used efficiently and increases the likelihood of a positive outcome for the end user within the available time for interaction.

[0068] The images selected through matching meta data to the at least one cluster can be a portion of the batch of images selected by the tool for presenting to the end user. The batch of images is provided 312 to the end user via the application on the computing device and feedback is received 314 from the end user. Feedback can be received from the end user through any of the means described within this application, such as assigning the images a score or selecting a sub-set of preferred images. The feedback from the end users is used to update 316 the profile of the end user. For example, the end user can indicate a particular interest in bird watching through selecting related images, or through discussion with the carer during the session. Accordingly, the profile can be updated through selections made directly through the application, or from additional information gleaned from conversations had with the end user following selection of the images. For example, a favourite pastime may be discovered.

[0069] In addition to the profile of the end user being updated 316, cluster data is also updated 318. For example, positive feedback to a particular image, or category of images, can increase the weighting which is applied to that particular image or category of images for future end users of that cluster. As described above, this can increase the efficiency of the tool in identifying passion points or key interests of the individual.

[0070] The end user can receive several batches of images within a session. Each batch of images can be impacted by the previous batch of images. Wherein the profile is updated following feedback to a batch of images, a new secure script can be generated. The therapy tool, based on the secure script, can then select images which are related to the images which previously received positive feedback. In such a way, the interests and passion points of the end user can be confirmed. Weighting can also be applied to the selection of the end users, wherein continuously selecting images within a specific category can indicate a strong preference or interest in the specific category. The strength of preferences can be recorded in the profile of the end user, and used for selection of images via the secure script supplied to the therapy tool. Similarly, the strength of a preference can be indicated in the application and viewed on the computing device. Random images can also be shown to confirm that a positive reaction was specific to the displayed image. Additionally, random images can also be shown in order to give the end user the opportunity to indicate a wide variety of preferences. The updated profile can be stored for future sessions.

[0071] As described above, matching the cluster data to the meta data of images is improved with an increased number of end users’ data. At least initially, cluster data can be developed from alternative sources, other than the application. For example, individuals could be polled on their general interests and this data added to the therapy tool.

[0072] Figure 4 illustrates an example of a library. The library can be applied by any of the examples of the therapy tool 100 described herein. The library is a storage facility, and in some examples is a remote storage facility such as a cloud storage system. The library comprises images for selection by the therapy tool 100 and providing to the client device. The library can comprise images which are available for selection and presentation to all end users as well as images which are only available for selection and presentation to specific end users. The profile of the user can be assigned an identification which does not include any personal details. The therapy tool can identify a profile through this identification to ensure that images specific to that identification are only shown to the related profile. In some cases an end user is able to acquire a copy of any of the images in a digital, print, framed, canvas, book or other format.

[0073] The library can receive inputs from a plurality of sources. For example, the end user or carer can upload images to the library from the client device 110, or a separate computing device. Wherein the library is a remote storage facility access can also be provided to the family, friends and connections of the end user 102 such that they can also upload images to the library. The library can also include stock images, or any open access images. Wherein the end user has a social media account, the application can link with the social media account and retrieve images. Similarly, data from the social media account can also be retrieved and included in the profile of the end user. For example, if an end user regularly interacts with a particular account through their social media account this could be considered a point of interest, or passion point, of the user.

[0074] In some examples, the therapy tool comprises a generative Artificial Intelligence (Al) model. The generative Al model can generate new images, based on the end user’s profile and feedback. In some examples, the library provides an input to the generative Al model. This input can be both publicly available images, and the images specific to the end user. Based on the images provided to the generative Al model by the library, the generative Al model generates additional images which are added to the library and therefore available to the therapy tool 100 for providing to the end user 102.

[0075] In some examples, the Al generative tool can receive inputs of both personal and public images such that personal and public images can be combined to create new images. For example, an image of an individual could be combined with an image of a destination from a travel blog to create a new image of the individual at the destination. Importantly, the new image is only made available to the specific end user.

[0076] In some examples, the therapy tool comprises a memory. The memory can store the selections made by the end user as well as a record of the end user’s temperament before during and after a plurality of sessions. For example, the memory can store passion points of the end user, indicated by their preferences during a session. Wherein the enduser has selected images comprising boats, the memory can store that the end user likes to be shown images of boats.

[0077] In some examples, the therapy tool 100 can use the inputs from a plurality of users to improve the images provided to the end user. For example, the therapy tool can identify patterns in the preferences of specific age groups, genders or nationalities. In such a way, images having a higher likelihood of receiving a positive response from the end user can be prioritised over other images. For example, the therapy tool may identify that men of a certain age and ethnicity, in general, share a particular passion point. Wherein the therapy tool is faced with a new user who is male and of that age and ethnicity, the therapy tool will show images related to that particular passion point during the session. In such a way, the more users which use the therapy tool, the more efficient the tool becomes at identifying patterns between users and identifying images likely to receive a positive response.

[0078] In some examples, the application on the client device includes gamification. Said gamification can be tailored to the physical and cognitive abilities of the end user. The gamification can be based on the profile of the user. The games can be a spot the difference game or a memory based game with the themes identified as passion points of the end user. For example, the gamification could be a game of pairs where the user has to select matching images of boats. In some examples, a reward will be provided to the end user after successful completion of the game. The reward could be a new, previously unseen image, or an amination related to the passion point behind the game. For some users, the gamification could include physical challenges such as walk 200 steps today, or challenge another user to a fitness or strength challenge. In such a way, the app can encourage positive changes in activity levels of end users both in terms of physical and mental activity. Further, particularly where the game involves additional users, this has the additional benefit of social interaction which has been shown to reduce withdrawal and symptoms of depression.

[0079] Another way in which the application can provide gamification is physical games within a residential setting. For example, a fishing game could be introduced within a care home. Further, where individuals have expressed an interest in baking via the therapy tool, baking activities can be introduced.

[0080] FIG. 5 is an example of an initial session of an end user with the application. The end user provides, optionally via the carer, their basic category data including that they are female, aged 75, born in Glasgow, and like bird watching and baking. The application then generalises this information and provides a secure script to the therapy tool. The end user is assigned to a Glasgow cluster, a female cluster, a 70-80 cluster and a specific cluster where all three of these clusters overlap. The therapy tool 100 then selects a plurality of images from the library. The indicated preference of bird watching is matchedwith the key word “birds”, and at least one picture of a bird is selected. Similarly, the indicated preference of baking is matched with the key word “cake” and at least one picture of a cake is selected. Based on previous data points, the therapy tool 100 matches the specific Glasgow, 70-80 and female cluster to a key word of “sewing” and accordingly selects at least one corresponding image from the library. The therapy tool 100 also selects a random image, which in this example is a picture of a camel. Additional images are selected to create a bundle of around fifty images. The selected images are supplied to the computing device for viewing by the end user. The end user then provides feedback on the images, according to any of the embodiments described herein.

[0081] FIG. 6 is an example of the processing of feedback from the session described in relation to FIG. 5. In this example, at the end of the interaction, the end user selects the images of the camel and one of the sewing images for further discussion. The application prompts the carer to ask the end user a question, in this example the question is “why did you select these images”? Through asking these questions, the carer learns that the end user worked in the Singer Sewing factory and that they have a granddaughter living in Dubai. These fact points are then added to the profile of the end user for future sessions. The carer has also successfully learned more about their end user and encouraged a personalised social interaction. In such a way, further conversations can be initiated based on these topics. Note that the example of the Singer Sewing factory is one example only and is not intended to be limiting.

[0082] As shown in FIG. 6 the feedback from the end user has two key impacts. The first, is the updating of the end user profile which in the subsequent session or interaction informs the selection of further images for displaying to the end user. The second impact is that, due to the positive response from the end user to the sewing image, the weighting of the “sewing” preference for the clusters which the end user was assigned to is increased. In such a way, future end users belonging to these clusters will be more likely than other clusters to receive images related to sewing.

[0083] FIG. 7 illustrates various components of an example computing device 1010 in which examples of a therapy tool 100 or client device 110 are implemented in some examples. Where the computing device 1010 is used to implement a client device it is in the form of a hand held computer such as a tablet computer, smart phone, smart watch or other hand held computer. Where the computing device 1010 is used to implement the therapy tool 100 it is a web server, communications network node, compute server, personal computer or other computing entity.

[0084] The computing device 1010 comprises one or more processors 1002 which are microprocessors, controllers or any other suitable type of processors for processing computer executable instructions to control the operation of the device in order to perform the methods of figures 1 to 6. In some examples, for example where a system on a chiparchitecture is used, the processors 1002 include one or more fixed function blocks (also referred to as accelerators) which implement a part of the method of figures 1 to 6 in hardware (rather than software or firmware). That is, the methods described herein are implemented in any one or more of software, firmware, hardware. The computing device has a data store holding vectors, categories, information about end users, images, keywords or other data. The computing device has either an application for interacting with a therapy tool, or therapy tool functionality. The computing device 1010 optionally has an encoder 1020 and optionally has a decoder 1022. Platform software comprising an operating system 1006 or any other suitable platform software is provided at the computing-based device to enable application software 1008 to be executed on the device. Although the computer storage media (memory 1012) is shown within the computing-based device 1010 it will be appreciated that the storage is, in some examples, distributed or located remotely and accessed via a network or other communication link (e.g. using interface 1004).

[0085] The computing-based device 1010 also comprises a display device 1014 which may be separate from or integral to the computing-based device 1010. The display may provide a graphical user interface and / or may display images.

[0086] The term “subset” means some but not all elements of a plurality of elements and does not include the empty set.

[0087] Any reference to 'an' item refers to one or more of those items. The term 'comprising' is used herein to mean including the method blocks or elements identified, but that such blocks or elements do not comprise an exclusive list and an apparatus may contain additional blocks or elements and a method may contain additional operations or elements. Furthermore, the blocks, elements and operations are themselves not impliedly closed.

[0088] The steps of the methods described herein may be carried out in any suitable order, or simultaneously where appropriate. The arrows between boxes in the figures show one example sequence of method steps but are not intended to exclude other sequences or the performance of multiple steps in parallel. Additionally, individual blocks may be deleted from any of the methods without departing from the spirit and scope of the subject matter described herein. Aspects of any of the examples described above may be combined with aspects of any of the other examples described to form further examples without losing the effect sought. Where elements of the figures are shown connected by arrows, it will be appreciated that these arrows show just one example flow of communications (including data and control messages) between elements. The flow between elements may be in either direction or in both directions.

[0089] Where the description has explicitly disclosed in isolation some individual features, any apparent combination of two or more such features is considered also to be disclosed,to the extent that such features or combinations are apparent and capable of being carried out based on the present specification as a whole in the light of the common general knowledge of a person skilled in the art, irrespective of whether such features or combinations of features solve any problems disclosed herein. In view of the foregoing description it will be evident to a person skilled in the art that various modifications may be made within the scope of the invention.

Claims

CLAIMS1. A client device comprising : a display ; a processor ; a neural network encoder; a memory, the memory storing instructions which when executed on the processor implement a method of: receiving data about an end user; generating a secure script using the received data by encoding the received data using the neural network encoder; sending the secure script to a generative machine learning model; receiving a response from the generative machine learning model, the response comprising an image or a vector encoding an image; and displaying the image at the display.

2. The client device of claim 1 wherein the image or the vector encoding the image have been computed by the generative machine learning model in response to the secure script.

3. The client device of claim 1 or claim 2 wherein sending the secure script to the generative machine learning model comprises sending the secure script in plaintext form.

4. The client device of any of claims 1 to 3 wherein the neural network encoder is configured to encode the received data into an embedding space which is the same embedding space as the embedding space of the generative machine learning model.

5. The client device of any of claims 1 to 4 comprising a decoder configured to decode the vector encoding the image into an image.

6. The client device of any of claims 1 to 5 wherein receiving the response from the generative machine learning model is contingent on receiving information at the client device about outcome of all or part of a game.

7. The client device of any of claims 1 to 6 comprising: a camera and emotion recognition functionality; or a graphical user interface arranged to receive the data about the end user by receiving any one or more of: a rating of a temperament of the end user, a rating of a level of cognitive ability of the end user, an age category, a hobby, a topic of interest.

8. A computer-implemented method comprising: receiving a secure script from an end user display device, the secure script comprising information about the end user; using the secure script to generate a response from a generative machine learning model, the response comprising an image or a vector encoding an image; and sending the response to the end user device to trigger display of the image at the end user device.

9. The method of claim 8 wherein the secure script comprises information about the end user in the form of category data.

10. The method of claim 8 or claim 9 wherein the secure script comprises a vector encoding of the information about the end user.

11. The method of claim 10 wherein the vector encoding of the information about the end user is in an embedding space of the generative machine learning model.

12. The method of any of claims 8 to 11 wherein the secure script comprises an encoding of an image.

13. The method of any of claims 8 to 12 wherein the secure script is formed using one or more key words from a library of key words.

14. The method of claim 6 wherein at least one of the key words is selected based on the category data of the end user.

15. An apparatus comprising: a processor; a memory storing instructions which when executed on the processor implement a method comprising: receiving a secure script from an end user display device, the secure script comprising information about the end user encoded as a vector; using the secure script to generate a response from a generative machine learning model by processing the vector using the generative machine learning model, the response comprising an image or a vector encoding an image; and sending the response to the end user device to trigger display of the image at the end user device.

Images