Abstract: A method for aligning a text with a media material, an apparatus, and a storage medium are provided. The method includes: determining a set of anchor points in the text; performing following operations i) to v) repeatedly until all anchor points are removed from the set of anchor points or all media materials are removed from a set of media materials: i) ranking the anchor points in the set of anchor points, ii) selecting a target anchor point from the set of anchor points based on the ranked anchor points in the set, iii) determining, from the set of media materials, a media material matching a text segment starting from the target anchor point, iv) removing the target anchor point, and v) removing the media material matching the text segment starting from the target anchor point; and aligning the text segments with respective media matching materials.

Abstract: In response to a request to park an ADV in a parking lot, a set of parking-trajectories associated with a set of predetermined locations near one or more parking spots in the parking lot may be obtained, where the set of parking-trajectories was previously generated based on prior collected planning and control data of the parking lot. Each parking-trajectory of the set of parking-trajectories may correspond to one parking spot of the one or more parking spots in the parking lot. A parking-trajectory may be selected from the set of parking-trajectories based on a current location of the ADV. The ADV may be controlled to park in a corresponding parking spot according to the selected parking-trajectory.

Abstract: A computer-implement method for operating an ADV is disclosed. A starting point and an ending point of a route along which an ADV is to be driven is determined. Whether each of the starting point and the ending point is within a first driving area having a lane boundary or a second driving area as an open space is determined. The route is divided into a first route segment and a second route segment. Dependent upon whether the starting point or the ending point is within the first driving area or the second driving area, the method comprises operating in one of an on-lane mode or an open-space mode to plan a first trajectory for the first route segment and operating in one of the on-lane mode or the open-space mode to plan a second trajectory for the second route segment.

Abstract: Information recommendation system usually involve a multitask problem, which tries to predict not only users' click-through rate (CTR) but also the post-click conversion rate (CVR). At the same time, for multi-functional information systems, there are commonly multiple services for users, such as news feed, search engine, and product suggestions. The prediction/ranking model should be conducted in a multi-scene manner. In the present patent document, embodiments of a unified ranking model for such a multi-task and multi-scene problem are disclosed. The disclosed model explores independent and non-shared embeddings for each task and scene, which reduces the coupling between tasks and scenes. Therefore, new tasks or scenes may be added easily. Besides, a simplified network may be chosen beyond the embedding layer, which largely improves the ranking efficiency for various online services. Extensive offline and online experiments demonstrated the superiority of model embodiments.

Abstract: Deep neural networks (DNN) model quantization may be used to reduce storage and computation burdens by decreasing the bit width. Presented herein are novel cursor-based adaptive quantization embodiments. In embodiments, a multiple bits quantization mechanism is formulated as a differentiable architecture search (DAS) process with a continuous cursor that represents a possible quantization bit. In embodiments, the cursor-based DAS adaptively searches for a quantization bit for each layer. The DAS process may be accelerated via an alternative approximate optimization process, which is designed for mixed quantization scheme of a DNN model. In embodiments, a new loss function is used in the search process to simultaneously optimize accuracy and parameter size of the model. In a quantization step, the closest two integers to the cursor may be adopted as the bits to quantize the DNN together to reduce the quantization noise and avoid the local convergence problem.

Abstract: Modern deep neural network (DNN) models have many layers with a single layer potentially involving large matrix multiplications. Such heavy calculation brings challenges to deploy such DNN models on a single edge device, which has relatively limited computation resources. Therefore, multiple and even heterogeneous edge devices may be required for applications with stringent latency requirements. Disclosed in the present patent documents are embodiments of a model scheduling framework that schedules multiple models on a heterogeneous platform. Multiple-model heterogeneous computing is partitioned into a neural computation optimizer (NCO) part and a neural computation accelerator (NCA) part. The migration, transition, or transformation of DNN models from cloud to edge is handled by the NCO, while the deployment of the transformed DNN models on the heterogeneous platform is handled by the NCA. Such a separation of implementation simplifies task execution and improves the flexibility for the overall framework.

Abstract: Deep neural network (DNN) models have been widely used for user-relevance content prediction. Presented herein are embodiments of a new user-relevance framework, which may be referred as Gating-Enhanced Multi-task Neural Networks (GemNN) embodiments. Neural network-based multi-task learning model embodiments herein predict user engagement with content in a coarse-to-fine manner, which gradually reduces content candidates and allows parameter sharing from upstream tasks to downstream tasks to improve the training efficiency. Also, in one or more embodiments, a gating mechanism was introduced between embedding layers and multi-layer perceptions to learn feature interactions and control the information flow fed to MLP layers. Tested embodiments demonstrated considerable improvements over prior approaches.

Abstract: Modem deep neural network (DNN) models have many layers with a single layer potentially involving large matrix multiplications. Such heavy calculation brings challenges to deploy such DNN models on a single edge device, which has relatively limited computation resources. Therefore, multiple and even heterogeneous edge devices may be required for applications with stringent latency requirements. Disclosed in the present patent documents are embodiments of a model scheduling framework that schedules multiple models on a heterogeneous platform. Two different approaches, model first scheduling (MFS) and hardware first scheduling (HFS), are presented to allocate a group of models for a service into corresponding heterogeneous edge devices, including CPU, VPU and GPU. Experimental results prove the effectiveness of the MFS and HFS methods for improving the inference speed of single and multiple AI-based services.

Abstract: Image super-resolution (SR) refers to the process of recovering high-resolution (HR) images from low-resolution (LR) inputs. Blind image SR is a more challenging task which involves unknown blurring kernels and characterizes the degradation process from HR to LR. In the present disclosure, embodiments of a variational autoencoder (VAE) are leveraged to train a kernel autoencoder for more accurate degradation representation and more efficient kernel estimation. In one or more embodiments, a kernel-agnostic loss is used to learn more robust kernel features in the latent space from LR inputs without using ground-truth kernel references. In addition, attention-based adaptive pooling is introduced to improve kernel estimation accuracy, and spatially non-uniform kernel features are passed into SR restoration resulting in additional kernel estimation error tolerance.

Abstract: A starting point, a set of one or more way points, and a destination point of a route along which the ADV is to be driven are determined. All lane segments near the starting point, the set of way points, and the destination point within a predetermined threshold distance are determined respectively. A set of route candidates are determined with an A-star (A*) searching algorithm based on a set of nodes representing all the lane segments near the starting point, the set of way points, and the destination point respectively. The route is selected from the set of route candidates based on respective costs of the set of route candidates. The ADV is being controlled to drive along the selected route autonomously.

Abstract: Described herein are embodiments of a reinforcement learning based large-scale multi-objective ranking system. Embodiments of the system may be used for optimizing short-video recommendation on a video sharing platform. Multiple competing ranking objective and implicit selection bias in user feedback are the main challenges in real-world platform. In order to address those challenges, multi-gate mixture of experts (MMoE) and soft actor critic (SAC) are integrated together into a MMoE_SAC system. Experiment results demonstrate that embodiments of the MMoE_SAC system may greatly reduce a loss function compared to systems only based on single strategies.

Abstract: The disclosure describes various embodiments for detecting an unexpected control state of an autonomous driving system. According to an embodiment, an exemplary method of detecting an unexpected control state of an autonomous driving system include the operations of generating environmental data of a vehicle; determining, by the autonomous driving system, a first control state based on the environmental data of the vehicle; determining, by a reference model, a second control state based on the environmental data, wherein the reference model defines at least one scenario each corresponding to a plurality of expected control states and a state switching condition, and in each of the expected control states corresponding to the scenario, an action of the vehicle in the scenario obeys a traffic rule; and determining the unexpected control state of the autonomous driving system by comparing the first control state with the second control state.

Abstract: A computer-implement method for operating an ADV is disclosed. A first trajectory for the ADV to drive along is planned. The ADV is to autonomously drive along the first trajectory. An obstacle in an affected region of the ADV is detected based on sensor data obtained from a plurality of sensors mounted on the ADV. An expected residence time of the obstacle in the affected region is determined. Whether to plan a second trajectory or to wait for the obstacle to leave the affected region is determined based on the expected residence time of the obstacle in the affected region. A second trajectory for the ADV to drive along is planned and the ADV is to autonomously drive along the second trajectory, or the ADV is to wait for the obstacle to leave the affected region and to autonomously drive along the first trajectory afterwards.

Abstract: Deep neural network (DNN) models have been widely used for user-relevance content prediction. Presented herein is a new user-relevance framework, embodiments of which may be referred as Gating-Enhanced Multi-task Neural Networks (GemNN). In one or more, neural network-based multi-task learning model embodiments herein predict user engagement with content in a coarse-to-fine manner, which gradually reduces content candidates and allows parameter sharing from upstream tasks to downstream tasks to improve the training efficiency. Also, in one or more embodiments, a gating mechanism was introduced between embedding layers and multi-layer perceptions to learn feature interactions and control the information flow fed to MLP layers. Tested embodiments demonstrated considerable improvements over prior approaches.

Abstract: A method, apparatus, and system for filtering a point cloud generated by a LIDAR device in an autonomous vehicle is disclosed. A point cloud comprising a plurality of points is generated based on outputs of the LIDAR device. The point cloud is filtered to remove a first set of points in the plurality of points that correspond to noise based on one or more of: point intensity measurements, distances between points, or a combination thereof. Perception data is generated based on the filtered point cloud. Operations of the autonomous vehicle are controlled based on the perception data.

Abstract: In some implementations, a method is provided. The method includes determining a first set of data acquisition characteristics of a first sensor of an autonomous driving vehicle. The method also includes determining a second set of data acquisition characteristics of a second sensor of the autonomous driving vehicle. The method further includes synchronizing a first data acquisition time of the first sensor and a second data acquisition time of the second sensor, based on the first set of data acquisition characteristics and the second set of data acquisition characteristics. The first sensor obtains first sensor data at the first data acquisition time. The second sensor obtains second sensor data at the second data acquisition time.

Abstract: According to one embodiment, a DP accelerator includes one or more execution units (EUs) configured to perform data processing operations in response to an instruction received from a host system coupled over a bus. The DP accelerator includes a time unit (TU) coupled to the security unit to provide timestamp services. The DP accelerator includes a security unit (SU) configured to establish and maintain a secure channel with the host system to exchange commands and data associated with the data processing operations, where the security unit includes a secure storage area to store a private root key associated with the DP accelerator, where the private root key is utilized for authentication. The SU includes a random number generator to generate a random number, and a cryptographic engine to perform cryptographic operations on data exchanged with the host system over the bus using a session key derived based on the random number.

Abstract: Described herein are systems and methods that improve the success rate of relocalization and eliminate the ambiguity of false relocalization by exploiting motions of the sensor system. In one or more embodiments, during a relocalization process, a snapshot is taken using one or more visual sensors and a single-shot relocalization in a visual map is implemented to establish candidate hypotheses. In one or more embodiments, the sensors move in the environment, with a movement trajectory tracked, to capture visual representations of the environment in one or more new poses. As the visual sensors move, the relocalization system tracks various estimated localization hypotheses and removes false ones until one winning hypothesis. Once the process is finished, the relocalization system outputs a localization result with respect to the visual map.

Abstract: A method for processing autonomous driving simulation data. The method includes: determining a type of a message transmitted between a simulation system and an auto driving system (ADS); determining a data acquisition mode based on the type of the message; obtaining a data stream transmitted between the simulation system and the ADS based on the data acquisition mode; and determining performance of the ADS based on the data stream.

Abstract: Embodiments of the present disclosure provide a method and apparatus for aligning a paragraph and a video. The method may include: acquiring a commentary and a candidate material resource set corresponding to the commentary, a candidate material resource being a video or an image; acquiring a matching degree between each paragraph in the commentary and each candidate material resource in the candidate material resource set; and determining a candidate material resource sequence corresponding to the each paragraph in the commentary based on the matching degrees between the paragraphs in the commentary and the candidate material resources, playing durations of the candidate material resources and text lengths of the paragraphs in the commentary, an image playing duration being a preset image playing duration.