Hrl Laboratories, LLC

The Boeing Company

Soheil Kolouri

Heiko Hoffmann

Charles E Martin

Mohammad Rostami

Kyungnam Kim

Praveen K Pilly

Rajan Bhattacharyya

David W Payton

Shankar R Rao

Amir M Rahimi

Qin Jiang

Adour Vahe Kabakian

Brian N Limketkai

Michael D Howard

Nigel D Stepp

Nicholas A Ketz

Yuri Owechko

Kang-Yu Ni

Hyukseong Kwon

Jiejun Xu

Samuel D Johnson

Hyun (Tiffany) J Kim

Wei L Yi

Christian Lebiere

Kangyu Ni

Alexander J Gabourie

Zachary Murez

Phillip E Pope

A method for training a controller to control a robotic system includes: receiving a neural network of an original controller for the robotic system based on origin data samples from an origin domain and labels in a label space, the neural network including encoder and classifier parameters, the neural network being trained to: map an input data sample from the origin domain to a feature vector in a feature space using the encoder parameters; and assign a label of the label space to the input data sample using the feature vector based on the classifier parameters; updating the encoder parameters to minimize a dissimilarity, in the feature space, between: origin feature vectors computed from the origin data samples; and target feature vectors computed from target data samples from a target domain; and updating the controller with the updated encoder parameters to control the robotic system in the target domain.

Systems and methods for few-shot transfer learning

Described is a system for continual adaptation of a machine learning model implemented in an autonomous platform. The system adapts knowledge previously learned by the machine learning model for performance in a new domain. The system receives a consecutive sequence of new domains comprising new task data. The new task data and past learned tasks are forced to share a data distribution in an embedding space, resulting in a shared generative data distribution. The shared generative data distribution is used to generate a set of pseudo-data points for the past learned tasks. Each new domain is learned using both the set of pseudo-data points and the new task data. The machine learning model is updated using both the set of pseudo-data points and the new task data.

Systems and methods for unsupervised continual learning

Described is a method for extraction of a region of interest (ROI) from a composite synthetic aperture radar (SAR) phase history data. The method comprising receiving, with a system comprising a processor, the composite SAR phase history data of a plurality of backscattered return signals produced by a SAR system illuminating a scene with a SAR beam. The method also comprises obtaining a location of a first ROI within the scene and extracting from the composite SAR phase history data a first component SAR phase history data corresponding to the ROI at the location of the ROI.

System for extraction of a region of interest (ROI) from a composite synthetic aperture radar (SAR) system phase history

Disclosed is a synthetic aperture radar (SAR) system for target recognition with complex range profile. The SAR system comprising a memory, a recurrent neural network (RNN), a multi-layer linear network in signal communication the RNN, and a machine-readable medium on the memory. The machine-readable medium is configured to store instructions that, when executed by the RNN, cause the SAR system to perform various operations. The various operation comprise: receiving raw SAR data associated with observed views of a scene, wherein the raw SAR data comprises information captured via the SAR system; radio frequency (RF) preprocessing the received raw SAR data to produce a processed raw SAR data; converting the processed raw SAR data to a complex SAR range profile data; processing the complex SAR range profile data with the RNN having RNN states; and mapping the RNN states to a target class with the multi-layer linear network.

Complex recurrent neural network for Synthetic Aperture Radar (SAR) target recognition

A computing system including a processor configured to train a synthetic aperture radar (SAR) classifier neural network. The SAR classifier neural network is trained at least in part by, at a SAR encoder, receiving training SAR range profiles that are tagged with respective first training labels, and, at an image encoder, receiving training two-dimensional images that are tagged with respective second training labels. Training the SAR classifier neural network further includes, at a shared encoder, computing shared latent representations based on the SAR encoder outputs and the image encoder outputs, and, at a classifier, computing respective classification labels based on the shared latent representations. Training the SAR classifier neural network further includes computing a value of a loss function based on the plurality of first training labels, the plurality of second training labels, and the plurality of classification labels and performing backpropagation based on the value of the loss function.

Synthetic aperture radar classifier neural network

Described is a system for object detection that is robust to adversarial attacks. An initial hypothesis of an identity of an object in an input image is generated using a sparse convolutional neural network (CNN) and a distribution aware classifier. A foveated hypothesis verification process is performed for identifying a region of the input image that supports the initial hypothesis. Using a part-based classifier, an identity of a part of the object in the region of the input image is predicted. An attack probability for the predicted identity of the part, and the initial hypothesis is updated based on the predicted identity of the part and the attack probability. The foveated hypothesis verification process and updating of hypotheses is performed until a hypothesis reaches a certainty threshold. The object is labeled based on the hypothesis that reached the certainty threshold.

Process to make machine object detection robust to adversarial attacks

Described is a system for automatically identifying chemical properties of a molecule. A chemical representation of a molecular structure is converted into atomic features and an adjacency matrix. The atomic features and the adjacency matrix are processed with a neural network, resulting in neural activations corresponding to each atom in the molecular structure. The system determines a probability for each atom quantifying its relevance for a given chemical characteristic. The probabilities are displayed as a graphical representation on the molecular structure, and groups of atoms are identified for the given chemical characteristic from the graphical representation. The identified groups of atoms for the given chemical characteristic are stored in a database, and a new molecule having the given chemical characteristic is designed based on the stored identified groups of atoms.

System and method for discovering chemically active compounds of a molecule

Described is a system for continual learning using experience replay. In operation, the system receives a plurality of tasks sequentially, from which a current task is fed to an encoder. The current task has data points associated with the current task. The encoder then maps the data points into an embedding space, which reflects the data points as discriminative features. A decoder then generates pseudo-data points from the discriminative features, which are provided back to the encoder. The discriminative features are updated in the embedding space based on the pseudo-data points. The encoder then learns (updates) a classification of a new task by matching the new task with the discriminative features in the embedding space.

System and method for continual learning using experience replay

Described is a system for learning object labels for control of an autonomous platform. Pseudo-task optimization is performed to identify an optimal pseudo-task for each source model of one or more source models. An initial target network is trained using the optimal pseudo-task. Source image components are extracted from source models, and an attribute dictionary of attributes is generated from the source image components. Using zero-shot attribution distillation, the unlabeled target data is aligned with the source models similar to the unlabeled target data. The unlabeled target data are mapped onto attributes in the attribute dictionary. A new target network is generated from the mapping, and the new target network is used to assign an object label to an object in the unlabeled target data. The autonomous platform is controlled based on the object label.

Process to learn new image classes without labels

Described is a system for adapting a deep convolutional neural network (CNN). A deep CNN is first trained on an annotated source image domain. The deep CNN is adapted to a new target image domain without requiring new annotations by determining domain agnostic features that map from the annotated source image domain and a target image domain to a joint latent space, and using the domain agnostic features to map the joint latent space to annotations for the target image domain.

Domain adaption learning system

A memory circuit configured to perform multiply-accumulate (MAC) operations for performance of an artificial neural network includes a series of synapse cells arranged in a cross-bar array. Each cell includes a memory transistor connected in series with a memristor. The memory circuit also includes input lines connected to the source terminal of the memory transistor in each cell, output lines connected to an output terminal of the memristor in each cell, and programming lines coupled to a gate terminal of the memory transistor in each cell. The memristor of each cell is configured to store a conductance value representative of a synaptic weight of a synapse connected to a neuron in the artificial neural network, and the memory transistor of each cell is configured to store a threshold voltage representative of a synaptic importance value of the synapse connected to the neuron in the artificial neural network.

Neuromorphic memory circuit and method of neurogenesis for an artificial neural network

Described is a system for detecting backdoor attacks in deep convolutional neural networks (CNNs). The system compiles specifications of a pretrained CNN into an executable model, resulting in a compiled model. A set of Universal Litmus Patterns (ULPs) are fed through the compiled model, resulting in a set of model outputs. The set of model outputs are classified and used to determine presence of a backdoor attack in the pretrained CNN. The system performs a response based on the presence of the backdoor attack.

System and method for detecting backdoor attacks in convolutional neural networks

Described is a system for continuously predicting and adapting optimal strategies for attacker elicitation. The system includes a global bot controlling processor unit and one or more local bot controlling processor units. The global bot controlling processor unit includes a multi-layer network software unit for extracting attacker features from diverse, out-of-band (OOB) media sources. The global controlling processing unit further includes an adaptive behavioral game theory (GT) software unit for determining a best strategy for eliciting identifying information from an attacker. Each local bot controlling processor unit includes a cognitive model (CM) software unit for estimating a cognitive state of the attacker and predicting attacker behavior. A generative adversarial network (GAN) software unit predicts the attacker's strategies. The global bot controlling processor unit and the one or more local bot controlling processor units coordinate to predict the attacker's next action and use the prediction to disrupt an attack.

Continuously habituating elicitation strategies for social-engineering-attacks (CHESS)

Described is a system for transferring learned knowledge from an electro-optical (EO) domain to a synthetic-aperture-radar (SAR) domain. The system uses a measured similarity between the EO domain and the SAR domain to train a model for classifying SAR images using knowledge previously learned from the electro-optical (EO) domain. Using the trained model, a SAR image is processed to determine regions of interest in the SAR image. A region of interest is classified to determine whether the region of interest corresponds to an object of interest, and classified regions of interest that contain the object of interest are output. The object of interest is displayed on a visualization map, and the visualization map is automatically updated to reflect a change in position of the object of interest.

System and method for transferring electro-optical (EO) knowledge for synthetic-aperture-radar (SAR)-based object detection

Described is a stripmap SAR system on a vehicle comprising an antenna that is fixed and directed outward from the side of the vehicle, a SAR sensor, a storage, and a computing device. The computing device comprises a memory, one or more processing units, and a machine-readable medium on the memory. The machine-readable medium stores instructions that, when executed by the one or more processing units, cause the stripmap SAR system to perform various operations. The operations comprise: receiving stripmap range profile data associated with observed views of a scene; transforming the received stripmap range profile data into partial circular range profile data; comparing the partial circular range profile data to a template range profile data of the scene; and estimating registration parameters associated with the partial circular range profile data relative to the template range profile data to determine a deviation from the template range profile data.

Growing community of inventors

Calabasas, CA, United States of America

Soheil Kolouri