Abstract: Methods, systems, apparatuses, devices, and computer program products are described. A system may input a first audio stream (e.g., audio recording) and a corresponding text sting into a machine learning model. The first audio stream and the text string may correspond to a first identity (e.g., person). Based on an output of the machine learning model, the system may generate a second audio stream associated with a second identity and mimics the first audio steam. For example, the second audio stream may be a generated recording of the second identity speaking the first text string. In addition, the system may generate a video depicting the second identity speaking the first text string (e.g., the second audio stream) based on combining the second audio stream with some image or previous video of the second identity. For example, the system may generate the video based on generating a head motion sequence.

Abstract: Systems and methods are provided for detecting, at a face detection model of a server, a face of a person having a first identity in an image. A face recognition model of the server may detect an identity vector for the detected face of the person in the image. A new identity vector may be generated based on at least the detected identity vector. A revised image may be generated using the generated new identity vector and the image to be output for display, where the face of the person having the first identity in the image has a second identity in the revised image based on the generated new identity vector. The server may transmit the revised image.

Abstract: A recommendation service access a data matrix listing of products associated with product profiles, the data matrix having product entries that store sparse historical electronic activity. For a target product it is determined which other products should be used to boost the historical electronic activity of the target product based on a first subset of product profiles that share product characteristics with the target product. Similarity scores are computed between the product profile of the target product and the first subset of product profiles to identify a second subset of one or more products having a similarity score above a scoring threshold. The historical electronic activity of the target product is boosted using the historical electronic activity of the other products in the second subset. Association values are calculated between the target product and the other products in the second subset by based on the boosted activity.

Abstract: A recommendation service access a data matrix listing of products associated with product profiles, the data matrix having product entries that store sparse historical electronic activity. For a target product it is determined which other products should be used to boost the historical electronic activity of the target product based on a first subset of product profiles that share product characteristics with the target product. Similarity scores are computed between the product profile of the target product and the first subset of product profiles to identify a second subset of one or more products having a similarity score above a scoring threshold. The historical electronic activity of the target product is boosted using the historical electronic activity of the other products in the second subset. Association values are calculated between the target product and the other products in the second subset by based on the boosted activity.

Abstract: Methods, computer readable media, and devices for exceeding the limits of visual-linguistic multi-task learning are disclosed. One method may include identifying a multi-modal multi-task classification dataset including a plurality of data examples, creating a transformer machine learning model to predict a plurality of categorical attributes of a product, and training the transformer machine learning model based on the multi-modal multi-task classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity.

Abstract: Systems and methods are provided for generating textual embeddings by tokenizing text data and generating vectors to be provided to a transformer system, where the textual embeddings are vector representations of semantic meanings of text that is part of the text data. The vectors may be averaged for every token of the generated textual embeddings and concatenating average output activations of two layers of the transformer system. Image embeddings may be generated with a convolutional neural network (CNN) from image data, wherein the image embeddings are vector representations of the images that are part of the image data. The textual embeddings and image embeddings may be combined to form combined embeddings to be provided to the transformer system.

Abstract: Systems and methods are provided for generating textual embeddings by tokenizing text data and generating vectors to be provided to a transformer system, where the textual embeddings are vector representations of semantic meanings of text that is part of the text data. The vectors may be averaged for every token of the generated textual embeddings and concatenating average output activations of two layers of the transformer system. Image embeddings may be generated with a convolutional neural network (CNN) from image data, wherein the image embeddings are vector representations of the images that are part of the image data. The textual embeddings and image embeddings may be combined to form combined embeddings to be provided to the transformer system.

Abstract: Methods, computer readable media, and devices for estimating product attribute preferences are disclosed. One method may include identifying a set of users, a set of products offered to users of the set of users, and a set of product attributes associated with products in the set of products, creating a product embedding matrix, an attribute embedding matrix, a user interaction matrix, a product attribute matrix, and a user attribute matrix, assigning an attribute weight to each product attribute, assigning, for each user, a user attribute weight for each product attribute, and displaying the set of products to a user in a ranked order based on the attribute weights and the user attribute weights assigned to the user.

Abstract: Systems and methods are provided for generating textual embeddings by tokenizing text data and generating vectors to be provided to a transformer system, where the textual embeddings are vector representations of semantic meanings of text that is part of the text data. The vectors may be averaged for every token of the generated textual embeddings and concatenating average output activations of two layers of the transformer system. Image embeddings may be generated with a convolutional neural network (CNN) from image data, wherein the image embeddings are vector representations of the images that are part of the image data. The textual embeddings and image embeddings may be combined to form combined embeddings to be provided to the transformer system.

Abstract: Systems and methods are provided for determining pairwise scores for items in an electronic catalog that is stored in a storage device communicatively coupled to a server, and determining when a pair of items in the electronic catalog is valid. A sum of the pairwise scores by activity type for a plurality of activity types may be determined, and the pairwise scores for the plurality of activity types may be summed. A sorted list of items that compliment an anchor item of the electronic catalog may be generated. Items from the generated sorted list may be filtered, and an item ranking in the filtered list may adjusted. Consistency between recommendations of the items in the adjusted ranking list may be determined, final recommendations of the items based on the determined consistency between recommendations may be selected and transmitted for display.

Abstract: Methods, computer readable media, and devices for exceeding the limits of visual-linguistic multi-task learning are disclosed. One method may include identifying a multi-modal multi-task classification dataset including a plurality of data examples, creating a transformer machine learning model to predict a plurality of categorical attributes of a product, and training the transformer machine learning model based on the multi-modal multi-task classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity.

Abstract: Systems and methods are provided for receiving, at a server, a selection of an anchor product from an electronic catalog stored in at least one storage device communicatively coupled to the server, and vectorizing at least one of text and images associated with the selected anchor product and other products in the catalog. At least one of key words may be determined from text data and key images from image data for each product of the catalog. Vectors may be formed from at least one of the keywords and key images, and concatenating the separate vectors together to form final vectors for the products. A similarity search may be performed using the final vectors to determine a group of similar products from the vectorized products of the catalog. Selected products that are within a same slot as the anchor product may be labelled in batch.

Abstract: Systems and methods are provided for generating textual embeddings by tokenizing text data and generating vectors to be provided to a transformer system, where the textual embeddings are vector representations of semantic meanings of text that is part of the text data. The vectors may be averaged for every token of the generated textual embeddings and concatenating average output activations of two layers of the transformer system. Image embeddings may be generated with a convolutional neural network (CNN) from image data, wherein the image embeddings are vector representations of the images that are part of the image data. The textual embeddings and image embeddings may be combined to form combined embeddings to be provided to the transformer system.

Abstract: Systems and methods are provided for determining pairwise scores for items in an electronic catalog that is stored in a storage device communicatively coupled to a server, and determining when a pair of items in the electronic catalog is valid. A sum of the pairwise scores by activity type for a plurality of activity types may be determined, and the pairwise scores for the plurality of activity types may be summed. A sorted list of items that compliment an anchor item of the electronic catalog may be generated. Items from the generated sorted list may be filtered, and an item ranking in the filtered list may adjusted. Consistency between recommendations of the items in the adjusted ranking list may be determined, final recommendations of the items based on the determined consistency between recommendations may be selected and transmitted for display.

Abstract: Systems and methods are provided for receiving, at a server, a selection of an anchor product from an electronic catalog stored in at least one storage device communicatively coupled to the server, and vectorizing at least one of text and images associated with the selected anchor product and other products in the catalog. At least one of key words may be determined from text data and key images from image data for each product of the catalog. Vectors may be formed from at least one of the keywords and key images, and concatenating the separate vectors together to form final vectors for the products. A similarity search may be performed using the final vectors to determine a group of similar products from the vectorized products of the catalog. Selected products that are within a same slot as the anchor product may be labelled in batch.

Abstract: A method of automated search result sorting includes applying a search request to a data set to generate search results, and generating a mean score value and an uncertainty value for each item in the search results by mathematically combining values of characteristics associated with the respective item, taking into account a weighting factor of the characteristic. The method further includes producing, for each of the items, a single score that is a function of the mean score value of that item and the uncertainty value of that item. This includes selecting, as the single score for each item, a value in accord with a probability distribution that is centered about the mean score value for the item and that has a standard deviation defined by the uncertainty value for that item. The method presents items from the search results ordered in accord with the respective single scores produced for them.