Abstract: In some aspects, a device may obtain point data from a lidar scanner. The point data may be associated with an angular subrange of a polar grid of the lidar scanner. The device may cause a transformer model to process the point data to identify a set of points based at least in part on the angular subrange, analyze the set of points based at least in part on a polar distance between the set of points and an origin of the polar grid, and indicate whether the set of points is associated with an object. The device may perform an action based at least in part on whether the transformer model indicates that the set of points is associated with the object. Numerous other aspects are described.

Abstract: A method for processing a video includes receiving a video as an input at a first layer of an artificial neural network (ANN). A first frame of the video is processed to generate a first label. Thereafter, the artificial neural network is updated based on the first label. The updating is performed while concurrently processing a second frame of the video. In doing so, the temporal inconsistency between labels is reduced.

Abstract: Certain aspects of the present disclosure provide techniques for provide a method, comprising: receiving input data for a layer of a neural network model; selecting a target code for the input data; and determining weights for the layer based on an autoencoder loss and the target code.

Abstract: Certain aspects of the present disclosure provide techniques for improved machine learning using gradient pruning, comprising computing, using a first batch of training data, a first gradient tensor comprising a gradient for each parameter of a parameter tensor for a machine learning model; identifying a first subset of gradients in the first gradient tensor based on a first gradient criteria; and updating a first subset of parameters in the parameter tensor based on the first subset of gradients in the first gradient tensor.

Abstract: Techniques for cross-domain adaptive learning are provided. A target domain feature extraction model is tuned from a source domain feature extraction model trained on a source data set, where the tuning is performed using a mask generation model trained on a target data set, and the tuning is performed using the target data set.

Abstract: An embodiment method includes performing first convolutional filtering on a first tensor constructed using a current frame and reference frames (or digital world reference images) of the current frame in a video, to generate a first estimated image of the current frame having a higher resolution than an image of the current frame. The method also includes performing second convolutional filtering on a second tensor constructed using the first estimated image and estimated reference images of the reference frames, to generate a second estimated image of the current having a higher resolution than the image of the current frame. The estimated reference images of the reference frames are reconstructed high resolution images of the reference images.

Abstract: Certain aspects of the present disclosure provide techniques for processing with an auto exiting machine learning model architecture, including processing input data in a first portion of a classification model to generate first intermediate activation data; providing the first intermediate activation data to a first gate; making a determination by the first gate whether or not to exit processing by the classification model; and generating a classification result from one of a plurality of classifiers of the classification model.

Abstract: Techniques are provided for determining consistency measures for image segmentation. For instance, a system can determine a first segmentation feature associated with a first segmentation mask of a first image frame. The system can determine a second segmentation feature associated with a second segmentation mask of a second image frame. The second segmentation feature corresponds to the first segmentation feature. The system can determine a first image feature of the first image frame that corresponds to the first segmentation feature and a second image feature of the second image frame that corresponds to the second segmentation feature. The system can determine a first similarity measurement between the first image feature and the second image feature. The system can further determine a temporal consistency measurement associated with the first image frame and the second image frame based at least in part on the first similarity measurement.

Abstract: A method applies a distance-based loss function to a boundary recognition model. The method classifies boundaries of an input with the boundary recognition model. The method also performs semantic segmentation based on the classifying of the boundaries, and outputting a segmentation map showing different classes of objects from the input, based on the semantic segmentation. The method may train an inverse transforming artificial neural network to predict a perspective transformation of an image so that the trained artificial neural network represents the distance-based loss function. The method may freeze weights of the inverse transforming artificial neural network, after training, to obtain the distance-based loss function. Training of the inverse transforming artificial neural network may include generating shifted, translated, and scaled versions of the image such that a ground truth comprises values corresponding to the amounts of shifting, translating, and scaling.

Abstract: Systems and techniques are described for performing supervised learning (e.g., semi-supervised learning, self-supervised learning, and/or mixed supervision learning) for optical flow estimation. For example, a method can include obtaining an image associated with a sequence of images and generating an occluded image. The occluded image can include at least one of the image with an occlusion applied to the image and a different image of the sequence of images with the occlusion applied. The method can include determining a matching map based at least on matching areas of the image and the occluded image and, based on the matching map, determining a loss term associated with an optical flow loss prediction associated with the image and the occluded image. The loss term may include a matched loss and/or other loss. Based on the loss term, the method can include training a network configured to determine an optical flow between images.

Abstract: Systems and techniques are described herein for performing optical flow estimation between one or more frames. For example, a process can include determining a subset of pixels of at least one of a first frame and a second frame, and generating a mask indicating the subset of pixels. The process can include determining, based on the mask, one or more features associated with the subset of pixels of at least the first frame and the second frame. The process can include determining optical flow vectors between the subset of pixels of the first frame and corresponding pixels of a second frame. The process can include generating an optical flow map for the second frame using the optical flow vectors.

Abstract: Certain aspects of the present disclosure provide techniques for machine learning using basis decomposition, comprising receiving a first runtime record, where the first runtime record includes RF signal data collected in a physical space; processing the first runtime record using a plurality of basis machine learning (ML) models to generate a plurality of inferences; aggregating the plurality of inferences to generate a prediction comprising a plurality of coordinates; and outputting the prediction, where the plurality of coordinates indicate a location of a physical element in a physical space.

Abstract: Certain aspects of the present disclosure provide techniques for performing tabular convolution, including performing a tabularization operation on input data to generate a tabularized representation of the input data and performing a convolution operation using the tabularized representation of the input data to generate a convolution output.

Abstract: Systems, methods, and non-transitory media are provided for providing spatiotemporal recycling networks (e.g., for video segmentation). For example, a method can include obtaining video data including a current frame and one or more reference frames. The method can include determining, based on a comparison of the current frame and the one or more reference frames, a difference between the current frame and the one or more reference frames. Based on the difference being below a threshold, the method can include performing semantic segmentation of the current frame using a first neural network. The semantic segmentation can be performed based on higher-spatial resolution features extracted from the current frame by the first neural network and lower-resolution features extracted from the one or more reference frames by a second neural network. The first neural network has a smaller structure and/or a lower processing cost than the second neural network.

Abstract: In one aspect of the present disclosure, a method includes: determining a number of loops for a convolution layer of an elastic bottleneck block; for each loop of the number of loops: loading a loop-specific set of convolution weights; performing a convolution operation using the loop-specific set of convolution-weights; and storing loop-specific convolution results in a local memory; and determining an output of the convolution layer based on a summation of loop-specific convolution results associated with each loop of the number of loops.

Abstract: Certain aspects of the present disclosure provide techniques for performing machine learning, including generating a set of basis masks for a convolution layer of a machine learning model, wherein each basis mask comprises a binary mask; determining a set of scaling factors, wherein each scaling factor of the set of scaling factors corresponds to a basis mask in the set of basis masks; generating a composite kernel based on the set of basis masks and the set of scaling factors; and performing a convolution operation based on the composite kernel.