Abstract: Examples provide a self-supervised language model for document-to-document similarity scoring and ranking long documents of arbitrary length in an absence of similarity labels. In a first stage of a two-staged hierarchical scoring, a sentence similarity matrix is created for each paragraph in the candidate document. A sentence similarity score is calculated based on the sentence similarity matrix. In the second stage, a paragraph similarity matrix is constructed based on aggregated sentence similarity scores associated with the first candidate document. A total similarity score for the document is calculated based on the normalize the paragraph similarity matrix for each candidate document in a collection of documents. The model is trained using a masked language model and intra-and-inter document sampling. The documents are ranked based on the similarity scores for the documents.

Abstract: A diagnostic tool for deep learning similarity models and image classifiers provides valuable insight into neural network decision-making. A disclosed solution generates a saliency map by: receiving a baseline image and a test image; determining, with a convolutional neural network (CNN), a first similarity between the baseline image and the test image; based on at least determining the first similarity, determining, for the test image, a first activation map for at least one CNN layer; based on at least determining the first similarity, determining, for the test image, a first gradient map for the at least one CNN layer; and generating a first saliency map as an element-wise function of the first activation map and the first gradient map. Some examples further determine a region of interest (ROI) in the first saliency map, cropping the test image to an area corresponding to the ROI, and determine a refined similarity score.

Abstract: The disclosure herein describes a system for interpreting text-based similarity between a seed item and a recommended item selected by a pre-trained language model from a plurality of candidate items based on semantic similarities between the seed item and the recommended item. The system analyzes similarity scores and contextual paragraph representations representing text-based descriptions of the seed item and recommended item to generate gradient maps and word scores representing the text-based descriptions. A model for interpreting text-based similarity utilizes the calculated gradients and word scores to match words from the seed item description with words in the recommended item description having similar semantic meaning. The word-pairs having the highest weight are identified by the system as the word-pairs having the greatest influence over the selection of the recommended item from the candidate items by the original pre-trained language model.

Abstract: The disclosure herein describes a system and method for attentive sentence similarity scoring. A distilled sentence embedding (DSE) language model is trained by decoupling a transformer language model using knowledge distillation. The trained DSE language model calculates sentence embeddings for a plurality of candidate sentences for sentence similarity comparisons. An embedding component associated with the trained DSE language model generates a plurality of candidate sentence representations representing each candidate sentence in the plurality of candidate sentences which are stored for use in analyzing input sentences associated with queries or searches. A representation is created for the selected sentence. This selected sentence representation is used with the plurality of candidate sentence representations to create a similarity score for each candidate sentence-selected sentence pair.

Abstract: Machine learning multiple features of an item depicted in images. Upon accessing multiple images that depict the item, a neural network is used to machine train on the plurality of images to generate embedding vectors for each of multiple features of the item. For each of multiple features of the item depicted in the images, in each iteration of the machine learning, the embedding vector is converted into a probability vector that represents probabilities that the feature has respective values. That probability vector is then compared with a value vector representing the actual value of that feature in the depicted item, and an error between the two vectors is determined. That error is used to adjust parameters of the neural network used to generate the embedding vector, allowing for the next iteration in the generation of the embedding vectors. These iterative changes continue thereby training the neural network.

Abstract: Solutions for localizing relevant objects in multi-object images include receiving a multi-object image; detecting a plurality of detected objects within the multi-object image; generating a primary heatmap for the multi-object image, the primary heatmap having at least one region of interest; determining a relevant detected object corresponding to a region of interest in the primary heatmap; determining an irrelevant detected object not corresponding to a region of interest in the primary heatmap; and indicating the relevant detected object as an output result but not indicating the irrelevant detected object as an output result. Some examples identify a plurality of objects that are visually similar to the relevant object and displaying the visually similar objects to a user, for example as recommendations of alternative catalog items on an e-commerce website. Some examples are able to identify a plurality of relevant objects and display multiple sets of visually similar objects.

Abstract: Aspects of the technology described herein use acoustic features of a music track to capture information for a recommendation system. The recommendation can work without analyzing label data (e.g., genre, artist) or usage data for a track. For each audio track, a descriptor is generated that can be used to compare the track to other tracks. The comparisons between track descriptors result in a similarity measure that can be used to make a recommendation. In this process, the audio descriptors are used directly to form a track-to-track similarity measure between tracks. By measuring the similarity between a track that a user is known to like and an unknown track, a decision can be made whether to recommend the unknown track to the user.

Abstract: The disclosure herein describes a system and method for attentive sentence similarity scoring. A distilled sentence embedding (DSE) language model is trained by decoupling a transformer language model using knowledge distillation. The trained DSE language model calculates sentence embeddings for a plurality of candidate sentences for sentence similarity comparisons. An embedding component associated with the trained DSE language model generates a plurality of candidate sentence representations representing each candidate sentence in the plurality of candidate sentences which are stored for use in analyzing input sentences associated with queries or searches. A representation is created for the selected sentence. This selected sentence representation is used with the plurality of candidate sentence representations to create a similarity score for each candidate sentence-selected sentence pair.

Abstract: Machine learning multiple features of an item depicted in images. Upon accessing multiple images that depict the item, a neural network is used to machine train on the plurality of images to generate embedding vectors for each of multiple features of the item. For each of multiple features of the item depicted in the images, in each iteration of the machine learning, the embedding vector is converted into a probability vector that represents probabilities that the feature has respective values. That probability vector is then compared with a value vector representing the actual value of that feature in the depicted item, and an error between the two vectors is determined. That error is used to adjust parameters of the neural network used to generate the embedding vector, allowing for the next iteration in the generation of the embedding vectors. These iterative changes continue thereby training the neural network.

Abstract: Aspects of the technology described herein use acoustic features of a music track to capture information for a recommendation system. The recommendation can work without analyzing label data (e.g., genre, artist) or usage data for a track. For each audio track, a descriptor is generated that can be used to compare the track to other tracks. The comparisons between track descriptors result in a similarity measure that can be used to make a recommendation. In this process, the audio descriptors are used directly to form a track-to-track similarity measure between tracks. By measuring the similarity between a track that a user is known to like and an unknown track, a decision can be made whether to recommend the unknown track to the user.

Abstract: A diagnostic tool for deep learning similarity models and image classifiers provides valuable insight into neural network decision-making. A disclosed solution generates a saliency map by: receiving a baseline image and a test image; determining, with a convolutional neural network (CNN), a first similarity between the baseline image and the test image; based on at least determining the first similarity, determining, for the test image, a first activation map for at least one CNN layer; based on at least determining the first similarity, determining, for the test image, a first gradient map for the at least one CNN layer; and generating a first saliency map as an element-wise function of the first activation map and the first gradient map. Some examples further determine a region of interest (ROI) in the first saliency map, cropping the test image to an area corresponding to the ROI, and determine a refined similarity score.

Abstract: The disclosure herein describes a recommendation system utilizing a specialized domain-specific language model for generating cold-start recommendations in an absence of user-specific data based on a user-selection of a seed item. A generalized language model is trained using a domain-specific corpus of training data, including title and description pairs associated with candidate items in a domain-specific catalog. The language model is trained to distinguish between real title-description pairs and fake title-description pairs. The trained language model analyzes the title and description of the seed item with the title and description of each candidate item in the catalog to create a hybrid set of scores. The set of scores includes similarity scores and classification scores for the seed item title with each candidate item description and title. The scores are utilized by the model to identify candidate items maximizing similarity with the seed item for cold-start recommendation to a user.

Abstract: A recommender system that represents items in a catalog by first feature vectors in a first vector space based on first characteristics of the items and second feature vectors in a second vector space based on second characteristics of the items different from the first characteristics and maps a feature vector defined in the first vector space for an item to a vector in the second vector space to provide recommendations based on the item.

Abstract: Machine learning multiple features of an item depicted in images. Upon accessing multiple images that depict the item, a neural network is used to machine train on the plurality of images to generate embedding vectors for each of multiple features of the item. For each of multiple features of the item depicted in the images, in each iteration of the machine learning, the embedding vector is converted into a probability vector that represents probabilities that the feature has respective values. That probability vector is then compared with a value vector representing the actual value of that feature in the depicted item, and an error between the two vectors is determined. That error is used to adjust parameters of the neural network used to generate the embedding vector, allowing for the next iteration in the generation of the embedding vectors. These iterative changes continue thereby training the neural network.

Abstract: Computerized searching for an item based on a prior viewed item. A displayed item is identified as a query input item to be used in searching for a target item. That input item has an associated set of embedding vectors each representing a respective feature of the input item. Target features of the search are then identified based on the input item. For each feature in the target item that is desired to be the same as the input item, an embedding vector for the input item is accessed as the vector for that feature in the search. For each feature in the target item that is desired to be different than the input item, a special vector associated with that desired value and feature is accessed for that feature in the search. These accessed vectors are then compared against target items to find close matches.

Abstract: The disclosure herein describes a system and method for attentive sentence similarity scoring. A distilled sentence embedding (DSE) language model is trained by decoupling a transformer language model using knowledge distillation. The trained DSE language model calculates sentence embeddings for a plurality of candidate sentences for sentence similarity comparisons. An embedding component associated with the trained DSE language model generates a plurality of candidate sentence representations representing each candidate sentence in the plurality of candidate sentences which are stored for use in analyzing input sentences associated with queries or searches. A representation is created for the selected sentence. This selected sentence representation is used with the plurality of candidate sentence representations to create a similarity score for each candidate sentence-selected sentence pair.

Abstract: The disclosure herein describes a recommendation system utilizing a specialized domain-specific language model for generating cold-start recommendations in an absence of user-specific data based on a user-selection of a seed item. A generalized language model is trained using a domain-specific corpus of training data, including title and description pairs associated with candidate items in a domain-specific catalog. The language model is trained to distinguish between real title-description pairs and fake title-description pairs. The trained language model analyzes the title and description of the seed item with the title and description of each candidate item in the catalog to create a hybrid set of scores. The set of scores includes similarity scores and classification scores for the seed item title with each candidate item description and title. The scores are utilized by the model to identify candidate items maximizing similarity with the seed item for cold-start recommendation to a user.

Abstract: In one embodiment, an audio signal for an audio track is received and segmented into a plurality of segments of the audio signal. The plurality of segments of audio are input into a classification network that is configured to predict output values based on a plurality of genre and mood combinations formed from different combinations of a plurality of genres and a plurality of moods. The classification network predicts a set of output values for the plurality of segments, each of the set of output values corresponding to one or more the plurality of genre and mood combinations. One or more of the plurality of genre and mood combinations are assigned to the audio track based on the set of output values for one or more of the plurality of segments.

Abstract: Aspects disclosed herein may utilize neural embedding techniques to model session activity. A dataset may be collected from on online market place, such as an app store. The data set may include one or more user sessions comprising sequential click actions and/or item purchases. Models may be generated to represent session activity and, therefore, may be utilized for contextual recommendations of apps in an online app store. As such, the various aspects disclosed herein may also generate purchase predictions based on click-purchase relations in a sequence. The item similarities and purchase predictions may be used to provide real-time aid to users navigating an online marketplace.

Abstract: A playlist generator that utilizes multiple data sources to rank each track within a set of candidate tracks to enable selection of candidate tracks according to the ranking. Candidate tracks are each scored according to one or more features, such as acoustic similarity and/or similar usage patterns of the candidate track or artist of the candidate track to a current or previously played track or artist. Each feature is weighted according to historical listening patterns surrounding a user-selected playlist seed artist. The weighting may also be further corrected according to historical listening patterns of the particular user. When historical usage data related to a particular seed artist is limited, more generalized historical usage data related to a higher level in a genre hierarchy may be used.