The present disclosure describes techniques of identifying optimal scheme of knowledge distillation (KD) for vision tasks. The techniques comprise configuring a search space by establishing a plurality of pathways between a teacher network and a student network and assigning an importance factor to each of the plurality of pathways; searching the optimal KD scheme by updating the importance factor and parameters of the student network during a process of training the student network; and performing KD from the teacher network to the student network by retraining the student network based at least in part on the optimized importance factors.

A method of jointly training of a transferable feature extractor network, an ordinal regressor network, and an order classifier network in an ordinal regression unsupervised domain adaption network by providing a source of labeled source images and unlabeled target images; outputting image representations from a transferable feature extractor network by performing a minimax optimization procedure on the source of labeled source images and unlabeled target images; training a domain discriminator network, using the image representations from the transferable feature extractor network, to distinguish between source images and target images; training an ordinal regressor network using a full set of source images from the transferable feature extractor network; and training an order classifier network using a full set of source images from said transferable feature extractor network.

A method for performing a crop yield estimation using a semi-supervised deep convolution neural network is provided. The method includes receiving monitoring data from a drone, wherein the monitoring data comprises a video of the crops captured by the drone; sampling the video by a predefined frame rate to obtain one or more images; inputting the images to a crop yield estimation model to obtain one or more result data, wherein the crop yield estimation model comprises a generator and a discriminator each comprising one or more DCNNs, and wherein the crop yield estimation model is trained by a semi-supervised learning method; and performing a quantity estimation and a quality estimation corresponding to the crops as shown in the images according to the one or more result data, so as to determine a total number and maturities of the crops respectively.

A system for assessing a data model includes a data receiver, a model receiver, and a model assessment device. The data receiver receives training data, historical data, and production data. The model receiver receives the data model associated with the historical data and trained using the training data. The historical data includes a first outcome of the data model provided based on an input feature in the production data. The model assessment device identifies a key feature in the production data relative to the input feature based on a target category in the historical data and a statistical distribution of the input feature in the production data. The model assessment device determines a second outcome of the data model based on the key feature. In response to the second outcome being different from the first outcome, the model assessment device determines a veracity score for assessing the data model.

A method for training an artificial intelligence intended to identify a predetermined object in the environment of an aircraft in flight. The method comprises steps of identifying at least one predetermined object in representations representing at least one predetermined object and its environment, establishing a training set and a validation set, the training set and the validation set comprising a plurality of representations from the representations representing at least one predetermined object, training the artificial intelligence with the training set and validating the artificial intelligence with the validation set. The artificial intelligence may then be used, in a method for assisting the landing of the aircraft, to identify a helipad where the landing operation may be performed. The artificial intelligence may also be used, in a method for avoiding a cable, to identify cables situated on or close to the trajectory of the aircraft.

Systems, methods and apparatus of integrated image sensing devices. In one example, a system includes a sensor that generates data. A memory device stores the generated data, and further stores a first portion of an artificial neural network (ANN). A host interface of the system is configured to communicate with a host system that stores a second portion of the ANN. The memory device can be stacked with the sensor. The memory device includes an inference engine configured to generate inference results using the stored data as input to the first portion of the ANN. The host interface is further configured to send the inference results to the host system for processing by the host system using the second portion of the ANN.

A method for image colorization includes receiving, from a camera, an input image including a plurality of input image pixels. One or more input interest pixels of the plurality of input image pixels are classified as corresponding to an object of interest. A display image is generated having a plurality of display image pixels each having pixel values based on relative temperature values of objects in a real-world environment, the display image pixels including display interest pixels corresponding to the input interest pixels. The display interest pixels are colorized with a color selected based on a recognized class of the object of interest to give a colorized display image, the selected color being independent of the relative temperature values of the object of interest. The colorized display image is displayed with the display interest pixels colorized with the selected color.

The present invention relates to methods and systems that utilize the production vehicles to develop new perception features related to new sensor hardware as well as new algorithms for existing sensors by using federated learning. To achieve this, the production vehicle's own worldview is post-processed and used as a reference, towards which the output of the software (SW) or hardware (HW) under development is compared. In case of a large discrepancy between the baseline worldview and perceived worldview by the module-under-test, the data is weakly annotated by the baseline worldview. Such weakly annotated data may subsequently be used to update the SW parameters of the “perception model” in the module-under-test in each individual vehicle, or to be transmitted to the “back-office” for off-board processing or more accurate annotations.

Embodiments described herein provide systems and methods for noise-robust contrastive learning. In view of the need for a noise-robust learning system, embodiments described herein provides a contrastive learning mechanism that combats noise by learning robust representations of the noisy data samples. Specifically, the training images are projected into a low-dimensional subspace, and the geometric structure of the subspace is regularized with: (1) a consistency contrastive loss that enforces images with perturbations to have similar embeddings; and (2) a prototypical contrastive loss augmented with a predetermined learning principle, which encourages the embedding for a linearly-interpolated input to have the same linear relationship with respect to the class prototypes. The low-dimensional embeddings are also trained to reconstruct the high-dimensional features, which preserves the learned information and regularizes the classifier.

An AI model evaluation method includes: obtaining an AI model and an evaluation data set, where the evaluation data set includes a plurality of pieces of evaluation data carrying labels that are used to indicate real results corresponding to the evaluation data; classifying the evaluation data in the evaluation data set based on a data feature to obtain an evaluation data subset; and calculating inference accuracy of the AI model on the evaluation data subset to obtain an evaluation result of the AI model on data whose value of the data feature meets the condition.