Abstract: Hardware and neural architecture co-search may be performed by operations including obtaining a specification of a function and a plurality of hardware design parameters. The hardware design parameters include a memory capacity, a number of computational resources, a communication bandwidth, and a template configuration for performing neural architecture inference. The operations further include determining, for each neural architecture among a plurality of neural architectures, an overall latency of performance of inference of the neural architecture by an accelerator within the hardware design parameters. Each neural architecture having been trained to perform the function with an accuracy. The operations further include selecting, from among the plurality of neural architectures, a neural architecture based on the overall latency and the accuracy.

Abstract: A computer-implemented method includes employing a dynamic Boltzmann machine (DyBM) to solve a maximum likelihood of generalized normal distribution (GND) of time-series datasets. The method further includes acquiring the time-series datasets transmitted from a source node to a destination node of a neural network including a plurality of nodes, learning, by the processor, a time-series generative model based on the GND with eligibility traces, and, performing, by the processor, online updating of internal parameters of the GND based on a gradient update to predict updated times-series datasets generated from non-Gaussian distributions.

Abstract: Preparation and execution of quantized scaling may be performed by operations including obtaining an original array and a scaling factor representing a ratio of a size of the original array to a size of a scaled array, determining, for each column of the scaled array, a horizontal coordinate of each of two nearest elements in the horizontal dimension of the original array, and, for each row of the scaled array, a vertical coordinate of each of two nearest elements in the vertical dimension of the original array, calculating, for each row of the scaled array and each column of the scaled array, a linear interpolation coefficient, converting each value of the original array from a floating point number into a quantized number, converting each linear interpolation coefficient from a floating point number into a fixed point number, storing, in a memory, the horizontal coordinates and vertical coordinates as integers, the values as quantized numbers, and the linear interpolation coefficients as fixed point numbers

Abstract: Deep reinforcement learning of cooperative neural networks can be performed by obtaining an action and observation sequence including a plurality of time frames, each time frame including action values and observation values. At least some of the observation values of each time frame of the action and observation sequence can be input sequentially into a first neural network including a plurality of first parameters. The action values of each time frame of the action and observation sequence and output values from the first neural network corresponding to the at least some of the observation values of each time frame of the action and observation sequence can be input sequentially into a second neural network including a plurality of second parameters. An action-value function can be approximated using the second neural network, and the plurality of first parameters of the first neural network can be updated using backpropagation.

Abstract: Neural network accelerator hardware-specific division of inference may be performed by operations including obtaining a computational graph and a hardware chip configuration. The operations also include dividing inference of the plurality of layers into a plurality of groups. Each group includes a number of sequential layers based on an estimate of duration and energy consumption by the hardware chip to perform inference of the neural network by performing the mathematical operations on activation data, sequentially by layer, of corresponding portions of layers of each group. The operations further include generating instructions for the hardware chip to perform inference of the neural network, sequentially by group, of the plurality of groups.

Abstract: Neural network accelerator hardware-specific division of inference may be performed by operations including obtaining a computational graph and a hardware chip configuration. The operations also include dividing inference of the plurality of layers into a plurality of groups. Each group includes a number of sequential layers based on an estimate of duration and energy consumption by the hardware chip to perform inference of the neural network by performing the mathematical operations on activation data, sequentially by layer, of corresponding portions of layers of each group. The operations further include generating instructions for the hardware chip to perform inference of the neural network, sequentially by group, of the plurality of groups.

Abstract: Neural network inference may be performed by configuration of a device including a plurality of convolution modules, a plurality of adder modules, an accumulation memory, and a convolution output interconnect control module configured to open and close convolution output interconnects among a plurality of convolution output interconnects connecting the plurality of convolution modules, the plurality of adder modules, and the accumulation memory. Inference may be performed while the device is configured according to at least one convolution output connection scheme whereby each convolution module has no more than one open direct connection through the plurality of convolution output interconnects to the accumulation memory or one of the plurality of adder modules. The device includes a convolution output interconnect control module to configure the plurality of convolution output interconnects according to the at least one convolution output connection scheme.

Abstract: A computer-implement method and an apparatus are provided for neural network reinforcement learning. The method includes obtaining, by a processor, an action and observation sequence. The method further includes inputting, by the processor, each of a plurality of time frames of the action and observation sequence sequentially into a plurality of input nodes of a neural network. The method also includes updating, by the processor, a plurality of parameters of the neural network by using the neural network to approximate an action-value function of the action and observation sequence.

Abstract: Regularization of neural networks. Neural networks can be regularized by obtaining an original neural network having a plurality of first-in-first-out (FIFO) queues, each FIFO queue located between a pair of nodes among a plurality of nodes of the original neural network, generating at least one modified neural network, the modified neural network being equivalent to the original neural network with a modified length of at least one FIFO queue, evaluating each neural network among the original neural network and the at least one modified neural network, and determining which neural network among the original neural network and the at least one modified neural network is most accurate, based on the evaluation.

Abstract: A computer-implement method and an apparatus are provided for neural network reinforcement learning. The method includes obtaining, by a processor, an action and observation sequence. The method further includes inputting, by the processor, each of a plurality of time frames of the action and observation sequence sequentially into a plurality of input nodes of a neural network. The method also includes updating, by the processor, a plurality of parameters of the neural network by using the neural network to approximate an action-value function of the action and observation sequence.

Abstract: A computer-implemented method, computer program product, and system are provided for learning mapping information between different modalities of data. The method includes mapping, by a processor, high-dimensional modalities of data into a low-dimensional manifold to obtain therefor respective low-dimensional embeddings through at least a part of a first network. The method further includes projecting, by the processor, each of the respective low-dimensional embeddings to a common latent space to obtain therefor a respective one of separate latent space distributions in the common latent space through at least a part of a second network. The method also includes optimizing, by the processor, parameters of each of the networks by minimizing a distance between the separate latent space distributions in the common latent space using a variational lower bound. The method additionally includes outputting, by the processor, the parameters as the mapping information.

Abstract: Methods and a system are provided for detecting object pose. A method includes training, by a processor, a first autoencoder (AE) to generate synthetic output images based on synthetic input images. The method further includes training, by the processor, a second AE to generate synthetic output images, similar to the synthetic output images generated by the first AE, based on real input images. The method also includes training, by the processor, a neural network (NN) to detect the object pose using the synthetic output images generated by the first and second AEs. The method additionally includes detecting and outputting, by the processor, a pose of an object in a real input test image by inputting the real input test image to the second AE to generate a synthetic image therefrom, and inputting the synthetic image to the NN to generate an NN output indicative of the pose of the object.

Abstract: A computer-implemented method is provided for autonomously making continuous trading decisions for assets using a first eligibility trace enabled Neural Network (NN). The method includes pretraining the first eligibility trace enabled NN, using asset price time series data, to generation predictions of future asset price time series data. The method further includes initializing a second eligibility trace enabled NN for reinforcement learning using learned parameters of the first eligibility trace enabled NN. The method also includes augmenting state information of the second eligibility trace enabled NN for reinforcement learning using an output from the first eligibility trace enabled NN. The method additionally includes performing continuous actions for trading assets at each of multiple time points.

Abstract: A computer-implemented method includes employing a dynamic Boltzmann machine (DyBM) to solve a maximum likelihood of generalized normal distribution (GND) of time-series datasets. The method further includes acquiring the time-series datasets transmitted from a source node to a destination node of a neural network including a plurality of nodes, learning, by the processor, a time-series generative model based on the GND with eligibility traces, and, performing, by the processor, online updating of internal parameters of the GND based on a gradient update to predict updated times-series datasets generated from non-Gaussian distributions.

Abstract: A computer-implemented method includes employing a dynamic Boltzmann machine (DyBM) to predict a higher-order moment of time-series datasets. The method further includes acquiring the time-series datasets transmitted from a source node to a destination node of a neural network including a plurality of nodes, learning, by the processor, a time-series generative model based on the DyBM with eligibility traces, and obtaining, by the processor, parameters of a generalized auto-regressive heteroscedasticity (GARCH) model to predict a time-varying second-order moment of the times-series datasets.

Abstract: A computer-implemented method, computer program product, and system are provided for learning mapping information between different modalities of data. The method includes mapping, by a processor, high-dimensional modalities of data into a low-dimensional manifold to obtain therefor respective low-dimensional embeddings through at least a part of a first network. The method further includes projecting, by the processor, each of the respective low-dimensional embeddings to a common latent space to obtain therefor a respective one of separate latent space distributions in the common latent space through at least a part of a second network. The method also includes optimizing, by the processor, parameters of each of the networks by minimizing a distance between the separate latent space distributions in the common latent space using a variational lower bound. The method additionally includes outputting, by the processor, the parameters as the mapping information.

Abstract: A computer-implemented method, computer program product, and system are provided for estimating a reward in reinforcement learning. The method includes preparing a state prediction model trained to predict a state for an input using visited states in expert demonstrations performed by an expert. The method further includes inputting an actual state observed by an agent in reinforcement learning into the state prediction model to calculate a predicted state. The method also includes estimating a reward in the reinforcement learning based, at least in part, on similarity between the predicted state and an actual state observed by the agent.

Abstract: Methods and a system are provided for detecting object pose. A method includes training, by a processor, a first autoencoder (AE) to generate synthetic output images based on synthetic input images. The method further includes training, by the processor, a second AE to generate synthetic output images, similar to the synthetic output images generated by the first AE, based on real input images. The method also includes training, by the processor, a neural network (NN) to detect the object pose using the synthetic output images generated by the first and second AEs. The method additionally includes detecting and outputting, by the processor, a pose of an object in a real input test image by inputting the real input test image to the second AE to generate a synthetic image therefrom, and inputting the synthetic image to the NN to generate an NN output indicative of the pose of the object.

Abstract: Adaptive exploration in deep reinforcement learning may be performed by inputting a current time frame of an action and observation sequence sequentially into a function approximator, such as a deep neural network, including a plurality of parameters, the action and observation sequence including a plurality of time frames, each time frame including action values and observation values, approximating a value function using the function approximator based on the current time frame to acquire a current value, updating an action selection policy through exploration based on an ?-greedy strategy using the current value, and updating the plurality of parameters.

Abstract: Deep reinforcement learning of cooperative neural networks can be performed by obtaining an action and observation sequence including a plurality of time frames, each time frame including action values and observation values. At least some of the observation values of each time frame of the action and observation sequence can be input sequentially into a first neural network including a plurality of first parameters. The action values of each time frame of the action and observation sequence and output values from the first neural network corresponding to the at least some of the observation values of each time frame of the action and observation sequence can be input sequentially into a second neural network including a plurality of second parameters. An action-value function can be approximated using the second neural network, and the plurality of first parameters of the first neural network can be updated using backpropagation.