Abstract: A training server acquires an input color image of an agricultural field, detects one or more foliage regions in the input color image, and generates output binary mask images of foliage mask indicating one or more foliage regions and a soil region. The training server further generates an augmented color image by combining pixels of the soil region adjusted for soil hue, with pixels of the one or more foliage regions unaltered from the acquired input color image in the RGB color space. The training server then utilizes the generated augmented color image in training of a crop detection (CD) neural network model.

Abstract: A training server includes one or more processors configured to determine a plurality of crop image data variation classifications representative of real-world variations in physical appearance of a crop plant as well as a surrounding area around the crop plant. A first set of input color images is selected from first training dataset and a plurality of different image level augmentation operations are executed to obtain an augmented set of color images. Noisy images are identified and filtered from a second training dataset and a third training dataset comprising noise filtered images from the second training dataset is obtained. The third training dataset is split into a plurality of different classes for data balancing across the plurality of different classes and a neural network model in a first stage is trained on the third training dataset.

Abstract: Techniques are generally described for modeling seasonal relevance for online search. A first query or item comprising text is received. A language model may be used to generate a plurality of token embeddings representing the text. In some cases, the first neural network may predict a seasonal relevance vector for the first query or item based on the plurality of token embeddings, the seasonal relevance vector being predicted based at least in part on a similarity between the text of the first query or item and at least one other query/item for which monthly concentration data is available. The seasonal relevance vector, including a respective seasonal relevance score for each month of a year, may be output.

Abstract: A camera apparatus includes control circuitry configured to acquire an input color image of an agricultural field, detect one or more foliage regions, and generate output binary mask images of foliage mask indicating one or more foliage regions and a soil region. The control circuitry is configured to convert the input color image to a Hue, Saturation, Lightness (HSV) color space to obtain an HSV image. Thereafter, the control circuitry is configured to selectively adjust a hue component and convert back to the RGB color space to obtain a soil region-adjusted RGB image. Furthermore, generate an augmented color image by combining pixels of the soil region, with pixels of the one or more foliage regions and utilize the generated augmented color image in training of a crop detection (CD) neural network model to learn a plurality of different types of soil and a range of color variation of soil.

Abstract: A camera apparatus includes one or more processors configured to determine a plurality of crop image data variation classifications representative of real-world variations in physical appearance of a crop plant as well as a surrounding area around the crop plant. Furthermore, select a first set of input color images from first training dataset comprising a plurality of different field-of-views (FOVs). Thereafter, execute plurality of different image level augmentation operations to obtain an augmented set of color images, identify and filter noisy images from a second training dataset based on a predefined set of image parameters. After that, train neural network model in a first stage on a third training dataset, re-determine new crop image data variation classifications and re-select new color images representative of the new crop image data variation classifications to further train the neural network model in a second stage to detect one or more crop plants.

Abstract: System and methods are provided that can address cold-start problems in database keyword searches. The search system generates machine-learned values for new item and queries based on historical signals for already existing item and queries. The values are used as input in a ranking model to rank search results for a user query. The initial values for the new item query pairs predict user engagement with the new item query pairs based on historical data for existing item query pairs and increase the visibility of new items to accumulate user interaction data for the new items. After additional user interactions are received, the values are updated using a Bayesian formula.

Abstract: A system implements a key value memory network including a key matrix with key vectors learned from training static feature data and time-series feature data, a value matrix with value vectors representing time-series trends, and an input layer to receive, for a target entity, input data comprising a concatenation of static feature data of the target entity, time-specific feature data, and time-series feature data for the target entity. The key value memory network also includes an entity-embedding layer to generate an input vector from the input data, a key-addressing layer to generate a weight vector indicating similarities between the key vectors and the input vector, a value-reading layer to compute a context vector from the weight and value vectors, and an output layer to generate predicted time-series data for a target metric of the target entity by applying a continuous activation function to the context vector and the input vector.

Abstract: A system implements a key value memory network including a key matrix with key vectors learned from training static feature data and time-series feature data, a value matrix with value vectors representing time-series trends, and an input layer to receive, for a target entity, input data comprising a concatenation of static feature data of the target entity, time-specific feature data, and time-series feature data for the target entity. The key value memory network also includes an entity-embedding layer to generate an input vector from the input data, a key-addressing layer to generate a weight vector indicating similarities between the key vectors and the input vector, a value-reading layer to compute a context vector from the weight and value vectors, and an output layer to generate predicted time-series data for a target metric of the target entity by applying a continuous activation function to the context vector and the input vector.

Abstract: This disclosure involves using key-value memory networks to predict time-series data. For instance, a computing system retrieves, for a target entity, static feature data and target time-series feature data. The computing system can normalize the target time-series feature data based on a normalization scale. The computing system also generates input data by, for example, concatenating the static feature data, the normalized time-series feature data, and time-specific feature data. The computing system generates predicted time-series data for the target metric of the target entity by applying a key-value memory network to the input data. The key-value memory network can include a key matrix learned from training static feature data and training time-series feature data, a value matrix representing time-series trends, and an output layer with a continuous activation function for generating predicted time-series data.

Abstract: System and methods are provided that can address cold-start problems in database keyword searches. The search system generates machine-learned values for new items based on historical signals for already existing items. These initial values are generated at the time of new item's inclusion in the search index. The values are used as input in a ranking model to rank search results for a user query. The initial values for the new items predict user engagement with the new items based on historical data for existing items and increase the visibility of new items to accumulate user interaction data for the new items.

Abstract: This disclosure involves using key-value memory networks to predict time-series data. For instance, a computing system retrieves, for a target entity, static feature data and target time-series feature data. The computing system can normalize the target time-series feature data based on a normalization scale. The computing system also generates input data by, for example, concatenating the static feature data, the normalized time-series feature data, and time-specific feature data. The computing system generates predicted time-series data for the target metric of the target entity by applying a key-value memory network to the input data. The key-value memory network can include a key matrix learned from training static feature data and training time-series feature data, a value matrix representing time-series trends, and an output layer with a continuous activation function for generating predicted time-series data.

Abstract: Non-limiting examples of the present disclosure describe query classification to identify appropriateness of a query. A query may be received by at least one processing device. A deep neural network (DNN) model may be applied to evaluate the query. A vector representation may be generated for query based on application of the DNN model, where the DNN model is trained to classify queries according to a plurality of categories of appropriateness. The DNN model may be utilized to classify the query in a category of appropriateness based on analysis of the vector representation. In one example, auto-complete suggestions for the query may be filtered based on the classification of the category of appropriateness. In another example, classification of the query may be provided to an entry point. In yet another example, a response to the query is managed based on the classification of the query. Other examples are also described.

Abstract: Non-limiting examples of the present disclosure describe query classification to identify appropriateness of a query. A query may be received by at least one processing device. A deep neural network (DNN) model may be applied to evaluate the query. A vector representation may be generated for query based on application of the DNN model, where the DNN model is trained to classify queries according to a plurality of categories of appropriateness. The DNN model may be utilized to classify the query in a category of appropriateness based on analysis of the vector representation. In one example, auto-complete suggestions for the query may be filtered based on the classification of the category of appropriateness. In another example, classification of the query may be provided to an entry point. In yet another example, a response to the query is managed based on the classification of the query. Other examples are also described.