Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating object predictions using a neural network. One of the methods includes receiving respective projections of a plurality of channels of input sensor data, wherein each channel of input sensor data represents different respective characteristics of electromagnetic radiation reflected off of one or more objects. Each of the projections of the plurality of channels of input sensor data are provided to a neural network subsystem trained to receive projections of input sensor data as input and to provide an object prediction as an output. At the output of the neural network subsystem, an object prediction that predicts a region of space that is likely to be occupied by an object is received.

Abstract: There is a need for more effective and efficient information deficiency processing. This need can be addressed by, for example, solutions for performing/executing label-based information deficiency processing. In one example, a method includes receiving a predictive input associated with a predictive entity based on the predictive input; determining a plurality of encoding probability values for the predictive entity based on the predictive input; determining a plurality of attention-based encoding vectors for the predictive entity based on the plurality of encoding probability values; determining an encoding deficiency prediction for the predictive entity, wherein the encoding deficiency prediction indicates a deficient subset of a plurality of encoding designations; and for each encoding designation in the deficient subset, performing a corresponding prediction-based action.

Abstract: Embodiments of the invention provide machine learning method and system. The method comprises: generating a group of sub-sequences based on a target sequence including n basic memory depth values, the group of sub-sequences includes at least one subset of composite sequences, and each composite sequence in any subset is generated based on an equal number of consecutive basic memory depth values (BMDV); determining weights of each sub-sequence, wherein initial weights for a composite sequence generated based on m BMDV are determined based on average of weights of at least two sub-sequences each having an equal number of BMDV which is less than and closest to m; determining weights of the target sequence based on an average of weights of at least two sub-sequences each having an equal number of BMDV which is closest to n; and solving the prediction problem based on weights of the target sequence.

Abstract: The amount of time required to train a neural network may be decreased by modifying the neural network to allow for greater parallelization of computations. The computations for cells of the neural network may be modified so that the matrix-vector multiplications of the cell do not depend on a previous cell and thus allowing the matrix-vector computations to be performed outside of the cells. Because the matrix-vector multiplications can be performed outside of the cells, they can be performed in parallel to decrease the computation time required for processing a sequence of training vectors with the neural network. The trained neural network may be applied to a wide variety of applications, such as performing speech recognition, determining a sentiment of text, determining a subject matter of text, answering a question in text, or translating text to another language.

Abstract: A generative adversarial neural network (GAN) learns a particular task by being shown many examples. In one scenario, a GAN may be trained to generate new images including specific objects, such as human faces, bicycles, etc. Rather than training a complex GAN having a predetermined topology of features and interconnections between the features to learn the task, the topology of the GAN is modified as the GAN is trained for the task. The topology of the GAN may be simple in the beginning and become more complex as the GAN learns during the training, eventually evolving to match the predetermined topology of the complex GAN. In the beginning the GAN learns large-scale details for the task (bicycles have two wheels) and later, as the GAN becomes more complex, learns smaller details (the wheels have spokes).

Abstract: A system for multitemporal data analysis is provided, comprising a directed computation graph service module configured to receive input data from a plurality of sources, analyze the input data to determine a best course of action for analyzing the input data, and split the input data for queueing to a general transformer service module or a decomposable service module based at least in part by analysis of the input data; a general transformer service module configured to receive data from the directed computation graph service module, and perform analysis on the received data; and a general transformer service module configured to receive data from directed computational graph module, and perform analysis on the received data.

Abstract: Computer implemented systems are described that implement a differentiable neural architecture search (DNAS) engine executing on one or more processors. The DNAS engine is configured with a stochastic super net defining a layer-wise search space having a plurality of candidate layers, each of the candidate layers specifying one or more operators for a neural network architecture. Further, the DNAS engine is configured to process training data to train weights for the operators in the stochastic super net based on a loss function representing a latency of the respective operator on a target platform, and to select a set of candidate neural network architectures from the trained stochastic super net. The DNAS engine may, for example, be configured to train the stochastic super net by traversing the layer-wise search space using gradient-based optimization of network architecture distribution.

Abstract: A neuromorphic synapse array includes a plurality of synaptic array cells being connected by circuitry such that the synaptic array cells are assigned to rows and columns of an array, the synaptic array cells respectively having a single polarity synapse weight, the rows respectively connected to respective input ends of the synaptic array cells, the columns respectively connected to respective output ends of the synaptic array cells, the synaptic array cells aligned in a column of the array being defined as operation column arrays and an array of current mirrors, each current mirror exhibiting a mirror ratio of N:1, wherein N is a number of columns of the synaptic array cells, respectively connected to the respective rows such that weights corresponding to all of the current mirrors are set to average weights of all of the synaptic array cells that are updated during a learning phase.

Abstract: A system and method for classifying products. A processor generates first and second instances of a first classifier, and trains the instances based on an input dataset. A second classifier is trained based on the input, where the second classifier is configured to learn a representation of a latent space associated with the input. A first supplemental dataset is generated in the latent space, where the first supplemental dataset is an unlabeled dataset. A first prediction is generated for labeling the first supplemental dataset based on the first instance of the first classifier, and a second prediction is generated for labeling the first supplemental dataset based on the second instance of the first classifier. Labeling annotations are generated for the first supplemental dataset based on the first prediction and the second prediction. A third classifier is trained based on at least the input dataset and the annotated first supplemental dataset.

Abstract: A method can be used to process an initial set of data through a convolutional neural network that includes a convolution layer followed by a pooling layer. The initial set is stored in an initial memory along first and second orthogonal directions. The method includes performing a first filtering of the initial set of data by the convolution layer using a first sliding window along the first direction. Each slide of the first window produces a first set of data. The method also includes performing a second filtering of the first sets of data by the pooling layer using a second sliding window along the second direction.

Abstract: The technology relates to predicting that an object is going to enter into a trajectory of a vehicle. This may include receiving sensor data identifying a first location of the object in an environment of the vehicle at a first point in time and receiving sensor data identifying a second location of the object in the environment at a second point in time. In addition, a boundary of the trajectory is determined by defining at least a two-dimensional area through which the vehicle is expected to travel in the future. A first distance between the boundary and the first location and a second distance between the trajectory and the second location are determined. The first distance and the second distance are used to determine that the object is going to enter into the trajectory at a future point in time.

Abstract: Apparatus and associated methods relate to predicting various transient output waveforms at a receiver's output after an initial neural network model is trained by a receiver's transient input waveform and a corresponding transient output waveform. In an illustrative example, the machine learning model may include an adaptive-ordered auto-regressive moving average external input based on neural networks (NNARMAX) model designed to mimic the performance of a continuous time linear equalization (CTLE) mode of the receiver. A Pearson Correlation Coefficient (PCC) score may be determined to select numbers of previous inputs and previous outputs to be used in the neural network model. In other examples, corresponding bathtub characterizations and eye diagrams may be extracted from the predicted transient output waveforms. Providing a machine learning model may, for example, advantageously predict various data patterns without knowing features or parameters of the receiver or related channels.

Abstract: A method of training a learning network is described. The method includes generating a first estimate of a gradient for the learning network and generating subsequent estimates of the gradient using a feedback network. The feedback network generates improved perturbations for the subsequent gradient estimates. Gradient estimates include the first estimate of the gradient and the subsequent estimates of the gradient. The method also includes using the gradient estimates to determine weights in the learning network. The improved perturbations may include lower variance perturbations.

Abstract: Systems and methods are described for a decision-making process including actions characterized by stochastic availability, provide an Markov decision process (MDP) model that includes a stochastic action set based on the decision-making process, compute a policy function for the MDP model using a policy gradient based at least in part on a function representing the stochasticity of the stochastic action set, identify a probability distribution for one or more actions available at a time period using the policy function, and select an action for the time period based on the probability distribution.

Abstract: Systems and methods for predicting system device failure are provided. The method includes representing device failure related data associated with the devices from a predetermined domain by temporal graphs for each of the devices. The method also includes extracting vector representations based on temporal graph features from the temporal graphs that capture both temporal and structural correlation in the device failure related data. The method further includes predicting, based on the vector representations and device failure related metrics in the predetermined domain, one or more of the devices that is expected to fail within a predetermined time.

Abstract: A neural network learning mechanism has a device which perturbs analog neurons to measure an error which results from perturbations at different points within the neural network and modifies weights and biases to converge to a target.

Abstract: In one embodiment, a method includes receiving, from a client system associated with a user of an online social network, data indicating that the user is located at a first geographic location at a first time; accessing a first embedding representing a first place-entity corresponding to the first geographic location; accessing multiple second embeddings representing multiple respective second place-entities each corresponding to a second geographic location; calculating, a similarity metric between the embedding representing the first place-entity and each of the embeddings representing the second place-entities; ranking each of the second place-entities based on their calculated similarity metrics; and sending, to the client system, information associated with one or more second geographic locations corresponding to one or more second place-entities having a ranking greater than a threshold ranking.

Abstract: Recurrent neural networks, and methods therefor, are provided with diagonal and programming fluctuation to find energy global minima. The method may include storing the matrix of weights in memory cells of a crossbar array of a recursive neural network prior to operation of the recursive neural network; altering the weights according to a probability distribution; setting the weights to non-zero values in at least one of the memory cells in a diagonal of the memory cells in the crossbar array; and operating the recursive neural network.

Abstract: As deep learning application domains grow, a deeper understanding of the relationships between training set size, computational scale, and model accuracy improvements is extremely beneficial. Presented herein are large-scale empirical study of error and model size growth as training sets grow. Embodiments of a methodology for this measurement are introduced herein as well as embodiments for predicting other metrics, such as compute-related metrics. It is shown herein that power-law may be used to represent deep model relationships, such as error and training data size. It is also shown that model size scales sublinearly with data size. These scaling relationships have significant implications on deep learning research, practice, and systems. They can assist model debugging, setting accuracy targets, and decisions about data set growth. They can also guide computing system design and underscore the importance of continued computational scaling.

Abstract: Systems and methods are provided for selecting an optimized data model architecture subject to resource constraints. One or more resource constraints for target deployment are identified, and random model architectures are generated from a set of model architecture production rules subject to the one or more resource constraints. Each random model architecture is defined by randomly chosen values for one or more meta parameters and one or more layer parameters. One or more of the random model architectures are adaptively refined to improve performance relative to a metric, and the refined model architecture with the best performance relative to the metric is selected.