Abstract: An electronic apparatus for compressing a neural network model may acquire training data pairs based on an original, trained neural network model and train a compressed neural network model compressed from the original, trained neural network model using the acquired training data pairs.

Abstract: Described herein are systems and methods for training first and second neural network models. A system comprises a memory comprising instruction data representing a set of instructions and a processor configured to communicate with the memory and to execute the set of instructions. The set of instructions, when executed by the processor, cause the processor to set a weight in the second model based on a corresponding weight in the first model, train the second model on a first dataset, wherein the training comprises updating the weight in the second model and adjust the corresponding weight in the first model based on the updated weight in the second model.

Abstract: Methods, systems and media for random semi-structured row-wise pruning of filters of a convolutional neural network are described. Rows of weights are pruned from kernels of filters of a convolutional layer of a convolutional neural network according to a pseudo-randomly-generated row pruning mask. The convolutional neural network is trained to perform a particular task using the pruned filters that include the rows of weights that have not been pruned from the kernels of filters. The process may be repeated multiple times, with the best-performing row pruning mask being selected for use in pruning row weights from kernel filters when the trained convolutional neural network is deployed to processing system and used for an inference. Computation time may be decreased further with the use of multiple parallel hardware computation units of a processing system performing pipelined row-wise convolution.

Abstract: A data processing system for training a neural network, the data processing system comprising: a first set of one or more processing units running one model of the neural network, a second set of one or more processing units running another model of the neural network, a data storage, and an interconnect between the first set of one or more processing units, the second set of processing units and the data storage, wherein the data storage is configured to provide over the interconnect, training data to the first set of one or more processing units and the second set of one more processing units, wherein each of the first and second set of processing units is configured to, when performing the training, evaluate loss for the respective training iteration including a measure of the dissimilarity between the output values calculated based on the different modes running on the first and second set of processing units, wherein the dissimilarity measure is weighted in the evaluation of the loss in accordance with a

Abstract: Systems and methods related to dual-momentum gradient optimization with reduced memory requirements are described. An example method in a system comprising a gradient optimizer and a memory configured to store momentum values associated with a neural network model comprising L layers is described. The method includes retrieving from the memory a first set of momentum values and a second set of momentum values, corresponding to a layer of the neural network model, having a selected storage format. The method further includes converting the first set of momentum values to a third set of momentum values having a training format associated with the gradient optimizer and converting the second set of momentum values to a fourth set of momentum values having a training format associated with the gradient optimizer. The method further includes performing gradient optimization using the third set of momentum values and the fourth set of momentum values.

Abstract: Disclosed are a neural network training method, a neural network training device and an electronic device. The neural network training method includes: training a first neural network to be trained by using sample data; determining an indicator parameter of the first neural network in a current training process; determining an update manner corresponding to a preset condition if the indicator parameter meets the preset condition; and updating a parameter of a batch normalization layer in the first neural network based on the update manner. In this way, sparsing of a feature map output by a neural network is implemented, thereby reducing an amount of data to be transmitted and improving computation speed of a chip.

Abstract: Embodiments for training a neural network using sequential tasks are provided. A plurality of sequential tasks are received. For each task in the plurality of tasks a copy of the neural network that includes a plurality of layers is generated. From the copy of the neural network a task specific neural network is generated by performing an architectural search on the plurality of layers in the copy of the neural network. The architectural search identifies a plurality of candidate choices in the layers of the task specific neural network. Parameters in the task specific neural network that correspond to the plurality of candidate choices and that maximize architectural weights at each layer are identified. The parameters are retrained and merged with the neural network. The neural network trained on the plurality of sequential tasks is a trained neural network.

Abstract: Memory layout and conversion are disclosed to improve neural network (NN) inference performance. For one example, a NN selects a memory layout for a neural network (NN) among a plurality of different memory layouts based on thresholds derived from performance simulations of the NN. The NN stores multi-dimensional NN kernel computation data using the selected memory layout during NN inference. The memory layouts to be selected can be a channel, height, width, and batches (CHWN) layout, a batches, height, width and channel (NHWC) layout, and a batches, channel, height and width (NCHW) layout. If the multi-dimensional NN kernel computation data is not in the selected memory layout, the NN transforms the multi-dimensional NN kernel computation data for the selected memory layout.

Abstract: A computer system and method for machine inductive learning on a graph is provided. In the inductive learning computational approach, an iterative approach is used for sampling a set of seed nodes and then considering their k-degree (hop) neighbors for aggregation and propagation. The approach is adapted to enhance privacy of edge weights by adding noise during a forward pass and a backward pass step of an inductive learning computational approach. Accordingly, it becomes more technically difficult for a malicious user to attempt to reverse engineer the edge weight information. Applicants were able to experimentally validate that acceptable privacy costs could be achieved in various embodiments described herein.

Abstract: A non-transitory, computer-readable recording medium stores therein a reinforcement learning program that uses a value function and causes a computer to execute a process comprising: estimating first coefficients of the value function represented in a quadratic form of inputs at times in the past than a present time and outputs at the present time and the times in the past, the first coefficients being estimated based on inputs at the times in the past, the outputs at the present time and the times in the past, and costs or rewards that corresponds to the inputs at the times in the past; and determining second coefficients that defines a control law, based on the value function that uses the estimated first coefficients and determining input values at times after estimation of the first coefficients.

Abstract: Described is a system for learning and predicting key phrases. The system learns based on a dataset of historical forecasting questions, their associated time-series data for a quantity of interest, and associated keyword sets. The system learns the optimal policy of actions to take given the associated keyword sets and the optimal set of keywords which are predictive of the quantity of interest. Given a new forecasting question, the system extracts an initial keyword set from a new forecasting question, which are perturbed to generate an optimal predictive key-phrase set. Key-phrase time-series data are extracted for the optimal predictive key-phrase set, which are used to generate a forecast of future values for a value of interest. The forecast can be used for a variety of purposes, such as advertising online.

Abstract: Examples of systems and method described herein provide for the processing of image codes (e.g., a binary embedding) at a memory die. Such images codes may generated by various endpoint computing devices, such as Internet of Things (IoT) computing devices, Such devices can generate a Hamming processing command, having an image code of the image, to compare that representation of the image to other images (e.g., in an image dataset) to identify a match or a set of neural network results. Advantageously, examples described herein may be used in neural networks to facilitate the processing of datasets, so as to increase the rate and amount of processing of such datasets. For example, comparisons of image codes can be performed on a memory die itself, like a memory die of a NAND memory device.

Abstract: Long-short term memory (LSTM) cells on spiking neuromorphic hardware are provided. In various embodiments, such systems comprise a spiking neurosynaptic core. The neurosynaptic core comprises a memory cell, an input gate operatively coupled to the memory cell and adapted to selectively admit an input to the memory cell, and an output gate operatively coupled to the memory cell an adapted to selectively release an output from the memory cell. The memory cell is adapted to maintain a value in the absence of input.

Abstract: The present disclosure provides systems, methods and computer-readable media for optimizing the neural architecture search for the automated machine learning process. In one aspect, neural architecture search method including selecting a neural architecture for training as part of an automated machine learning process; collecting statistical parameters on individual nodes of the neural architecture during the training; determining, based on the statistical parameters, active nodes of the neural architecture to form a candidate neural architecture; and validating the candidate neural architecture to produce a trained neural architecture to be used in implemented an application or a service.

Abstract: A method for producing a strategy for a robot. The method includes the following steps: initializing the strategy and an episode length; repeated execution of the loop including the following steps: producing a plurality of further strategies as a function of the strategy; applying the plurality of the further strategies for the length of the episode length; ascertaining respectively a cumulative reward, which is obtained in the application of the respective further strategy; updating the strategy as a function of a second plurality of the further strategies that obtained the greatest cumulative rewards. After each execution of the loop, the episode length is increased. A computer program, a device for carrying out the method, and a machine-readable memory element on which the computer program is stored, are also described.

Abstract: A neural processing system includes a first frontend module, a second frontend module, a first backend module, and a second backend module. The first frontend module executes a feature extraction operation using a first feature map and a first weight, and outputs a first operation result and a second operation result. The second frontend module executes the feature extraction operation using a second feature map and a second weight, and outputs a third operation result and a fourth operation result. The first backend module receives an input of the first operation result provided from the first frontend module and the fourth operation result provided from the second frontend module via a second bridge to sum up the first operation result and the fourth operation result.

Abstract: A method of pruning a convolutional neural network, comprising at least one of determining a number of channels (N) between a network input and a network output, constructing N lookup tables, each lookup table matched to a respective channel and pruning filters in the convolutional neural network to create a shortcut between the network input and the network output based on the N lookup tables.

Abstract: An electronic clinical decision support (CDS) device (10) employs a trained CDS algorithm (30) that operates on values of a set of covariates to output a prediction of a medical condition. The CDS algorithm was trained on a training data set (22). The CDS device includes a computer (12) that is programmed to provide a user interface (62) for completing clinical survey questions using the display and the one or more user input devices. Marginal probability distributions (42) for the covariates of the set of covariates are generated from the completed clinical survey questions. The trained CDS algorithm is adjusted for covariate shift using the marginal probability distributions. A prediction of the medical condition is generated for a medical subject using the trained CDS algorithm adjusted for covariate shift (50) operating on values for the medical subject of the covariates of the set of covariates.

Abstract: Autonomous vehicles may utilize neural networks for image classification in order to navigate infrastructures and foreign environments, using context dependent transfer learning adaptation. Techniques include receiving a transferable output layer from the infrastructure, which is a model suitable for the infrastructure and the local environment. Sensor data from the autonomous vehicle may then be passed through the neural network and classified. The classified data can map to an output of the transferable output layer, allowing the autonomous vehicle to obtain particular outputs for particular context dependent inputs, without requiring further parameters within the neural network.

Abstract: A neuromorphic package device includes a systolic array package and a controller. The systolic array package includes neuromorphic chips arranged in a systolic array along a first direction and a second direction. The controller communicates with a host controls the neuromorphic chips. Each of the neuromorphic chips sequentially transfers weights of a plurality layers of a neural network system in the first direction to store the weights. A first neuromorphic chip performs a calculation based on stored weights therein and an input data received in the second direction, and provides a result of the calculation to at least one of a second neuromorphic chip and a third neuromorphic chip which are adjacent to the first neuromorphic chip. The at least one of the second and third neuromorphic chips performs a calculation based on a provided result of the calculation and stored weights therein.

Abstract: A neuromorphic device includes a memory cell array that includes first memory cells corresponding to a first address and storing first weights and second memory cells corresponding to a second address and storing second weights, and a neuron circuit that includes an integrator summing first read signals from the first memory cells and an activation circuit outputting a first activation signal based on a first sum signal of the first read signals output from the integrator.

Abstract: A system and method for operating a neural network. In some embodiments, the neural network includes a variational autoencoder, and the training of the neural network includes training the variational autoencoder with a plurality of samples of a first random variable; and a plurality of samples of a second random variable, the plurality of samples of the first random variable and the plurality of samples of the second random variable being unpaired, the training of the neural network including updating weights in the neural network based on a first loss function, the first loss function being based on a measure of deviation from consistency between: a conditional generation path from the first random variable to the second random variable, and a conditional generation path from the second random variable to the first random variable.

Abstract: The present disclosure relates to a method for allocating resources of an accelerator to two or more neural networks for execution. The two or more neural networks may include a first neural network and a second neural network. The method comprises analyzing workloads of the first neural network and the second neural network, wherein the first neural network and second neural network each includes multiple computational layers, evaluating computational resources of the accelerator for executing each computational layer of the first and second neural networks, and scheduling computational resources of the accelerator to execute one computational layer of the multiple computation layers of the first neural network and to execute one or more computational layers of the multiple computational layers of the second neural network.

Abstract: A micro-processor circuit and a method of performing neural network operation are provided. The micro-processor circuit is suitable for performing neural network operation. The micro-processor circuit includes a parameter generation module, a compute module and a truncation logic. The parameter generation module receives in parallel a plurality of input parameters and a plurality of weight parameters of the neural network operation. The parameter generation module generates in parallel a plurality of sub-output parameters according to the input parameters and the weight parameters. The compute module receives in parallel the sub-output parameters. The compute module sums the sub-output parameters to generate a summed parameter. The truncation logic receives the summed parameter. The truncation logic performs a truncation operation based on the summed parameter to generate a plurality of output parameters of the neural network operation.

Abstract: A method of training an action selection neural network to perform a demonstrated task using a supervised learning technique. The action selection neural network is configured to receive demonstration data comprising actions to perform the task and rewards received for performing the actions. The action selection neural network has auxiliary prediction task neural networks on one or more of its intermediate outputs. The action selection policy neural network is trained using multiple combined losses, concurrently with the auxiliary prediction task neural networks.

Abstract: In a hardware-implemented approach for operating a neural network system, a neural network system is provided comprising a controller, a memory, and an interface connecting the controller to the memory, where the controller comprises a processing unit configured to execute a neural network and the memory comprises a neuromorphic memory device with a crossbar array structure that includes input lines and output lines interconnected at junctions via electronic devices. The electronic devices of the neuromorphic memory device are programmed to incrementally change states by coupling write signals into the input lines based on: write instructions received from the controller and write vectors generated by the interface. Data is retrieved from the neuromorphic memory device, according to a multiply-accumulate operation, by coupling read signals into one or more of the input lines of the neuromorphic memory device based on: read instructions from the controller and read vectors generated by the interface.

Abstract: Methods, systems, and non-transitory computer readable storage media are disclosed for training a machine-learning model utilizing batchwise weighted loss functions and scaled padding based on source density. For example, the disclosed systems can determine a density of words or phrases in digital content from a digital content source that indicate an affinity towards one or more content classes. In some embodiments, the disclosed systems can use the determined source density to split digital content from the source into segments and pad the segments with padding characters based on the source density. The disclosed systems can also generate document embeddings using the padded segments and then train the machine-learning model using the document embeddings. Furthermore, the disclosed system can use batchwise weighted cross entropy loss for applying different class weightings on a per-batch basis during training of the machine-learning model.

Abstract: The ability to rapidly identify symmetry and anti-symmetry is an essential attribute of intelligence. Symmetry perception is a central process in human vision and may be key to human 3D visualization. While previous work in understanding neuron symmetry perception has concentrated on the neuron as an integrator, the invention provides the coincidence detecting property of the spiking neuron can be used to reveal symmetry density in spatial data. A synchronized symmetry-identifying spiking artificial neural network enables layering and feedback in the network. The network of the invention can identify symmetry density between sets of data and present a digital logic implementation demonstrating an 8×8 leaky-integrate-and-fire symmetry detector in a field-programmable gate array. The efficiency of spiking neural networks can be harnessed to rapidly identify symmetry in spatial data with applications in image processing, 3D computer vision, and robotics.

Abstract: An online concierge system trains a user interaction model to predict a probability of a user performing an interaction after one or more content items are displayed to the user. This provides a measure of an effect of displaying content items to the user on the user performing one or more interactions. The user interaction model is trained from displaying content items to certain users of the online concierge system and withholding display of the content items to other users of the online concierge system. To train the user interaction model, the user interaction model is applied to labeled examples identifying a user and value based on interactions the user performed after one or more content items were displayed to the user and interactions the user performed when one or more content items were not used.

Abstract: System configurations and techniques for implementation of a neural network in neuromorphic hardware with use of external memory resources are described herein. In an example, a system for processing spiking neural network operations includes: a plurality of neural processor clusters to maintain neurons of the neural network, with the clusters including circuitry to determine respective states of the neurons and internal memory to store the respective states of the neurons; and a plurality of axon processors to process synapse data of synapses in the neural network, with the processors including circuitry to retrieve synapse data of respective synapses from external memory, evaluate the synapse data based on a received spike message, and propagate another spike message to another neuron based on the synapse data. Further details for use and access of the external memory and processing configurations for such neural network operations are also disclosed.

Abstract: An information processing apparatus that performs deep learning using a neural network includes a memory, and an arithmetic processing device that performs a process for layers of the neural network in a predetermined direction. The process for the layers includes: pre-determining a decimal point position of a fixed-point number of an intermediate data obtained by an operation of each of the layers; performing the arithmetic operation for each layer with the pre-determined decimal point position to obtain the intermediate data and acquiring first statistical information of a distribution of bits of the intermediate data; determining a decimal point position of the intermediate data based on the statistical information; and performing the arithmetic operation for each layer with the determined decimal point position again when the difference of the determined decimal point position and the pre-determined decimal point position is greater than a threshold value.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for implementing a reservoir computing neural network. In one aspect there is provided a reservoir computing neural network comprising: (i) a brain emulation sub-network, and (ii) a prediction sub-network. The brain emulation sub-network is configured to process the network input in accordance with values of a plurality of brain emulation sub-network parameters to generate an alternative representation of the network input. The prediction sub-network is configured to process the alternative representation of the network input in accordance with values of a plurality of prediction sub-network parameters to generate the network output. The values of the brain emulation sub-network parameters are determined before the reservoir computing neural network is trained and are not adjusting during training of the reservoir computing neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for augmenting neural networks with an external memory. One of the methods includes providing an output derived from the neural network output for the time step as a system output for the time step; maintaining a current state of the external memory; determining, from the neural network output for the time step, memory state parameters for the time step; updating the current state of the external memory using the memory state parameters for the time step; reading data from the external memory in accordance with the updated state of the external memory; and combining the data read from the external memory with a system input for the next time step to generate the neural network input for the next time step.

Abstract: A semiconductor circuit and an operating method for the same are provided. The method includes the following steps. A memory circuit is operated during a first timing to obtain a first memory state signal S1. The memory circuit is operated during a second timing after the first timing to obtain a second memory state signal S2. A difference between the first memory state signal S1 and the second memory state signal S2 is calculated to obtain a state difference signal SD. A calculating is performed to obtain an un-compensated output data signal OD relative with an input data signal ID and the second memory state signal S2. The state difference signal SD and the un-compensated output data signal OD are calculated to obtain a compensated output data signal OD?.

Abstract: Techniques for manipulation of a recursive neural network using random access memory are disclosed. Neural network descriptor information and weight matrices are stored in a semiconductor random access memory device which includes neural network processing logic. The network descriptor information and weight matrices comprise a trained neural network functionality. An input matrix is obtained for processing on the memory device. The trained neural network functionality is executed on the input matrix, which includes processing data for a first layer from the neural network descriptor information to set up the processing logic; manipulating the input matrix using the processing logic and at least one of the weight matrices; caching results of the manipulating in a storage location of the memory device; and processing recursively the results that were cached through the processing logic. Additional data for additional layers is processed until the neural network functionality is complete.

Abstract: An AV is described herein. The AV includes a lidar sensor system. The AV additionally includes a computing system that executes a road surface analysis component to determine, based upon lidar sensor data, whether a road surface feature is present on or in a roadway in a travel path of the AV. The AV can be configured to initiate a mitigation maneuver responsive to determining that the road surface feature is present. Performing the mitigation maneuver causes the AV to avoid the road surface feature or decelerate prior to reaching the road surface feature, thereby improving the apparent quality or comfort of the ride to a passenger of the AV.

Abstract: A method for providing a training data set used for training a signal classification neural network is provided. The method includes generating at least one first virtual waveform primitive comprising a predetermined signal level and at least one second virtual waveform primitive comprising a signal edge. The training data set is formed and comprises a predetermined number of generated virtual waveform primitives including first virtual waveform primitives and second virtual waveform primitives. Each virtual waveform primitive comprises a sequence of time and amplitude discrete values. The training data set is used for training the signal classification neural network.

Abstract: Disclosed herein are a convolutional layer acceleration unit, an embedded system having the convolutional layer acceleration unit, and a method for operating the embedded system. The method for operating an embedded system, the embedded system performing an accelerated processing capability programmed using a Lightweight Intelligent Software Framework (LISF), includes initializing and configuring, by a parallelization managing function entity (FE), entities present in resources for performing mathematical operations in parallel, and processing in parallel, by an acceleration managing FE, the mathematical operations using the configured entities.

Abstract: Techniques for phased deployment of machine learning models are described. Customers can call a training API to initiate model training, but then must wait while the training completes before the model can be used to perform inference. Depending on the type of model, machine learning algorithm being used for training, size of the training dataset, etc. this training process may take hours or days to complete. This leads to significant downtime where inference requests cannot be served. Embodiments improve upon existing systems by providing phased deployment of custom models. For example, a simple, less accurate model, can be provided synchronously in response to a request for a custom model. At the same time, one or more machine learning models can be trained asynchronously in the background. When the machine learning model is ready for use, the customers' traffic and jobs can be transferred over to the better model.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for detecting intermediate reinforcement learning goals. One of the methods includes obtaining a plurality of demonstration sequences, each of the demonstration sequences being a sequence of images of an environment while a respective instance of a reinforcement learning task is being performed; for each demonstration sequence, processing each image in the demonstration sequence through an image processing neural network to determine feature values for a respective set of features for the image; determining, from the demonstration sequences, a partitioning of the reinforcement learning task into a plurality of subtasks, wherein each image in each demonstration sequence is assigned to a respective subtask of the plurality of subtasks; and determining, from the feature values for the images in the demonstration sequences, a respective set of discriminative features for each of the plurality of subtasks.

Abstract: An apparatus for providing similar contents, using a neural network, includes a memory storing instructions, and a processor configured to execute the instructions to obtain a plurality of similarity values between a user query and a plurality of images, using a similarity neural network, obtain a rank of each the obtained plurality of similarity values, and provide, as a most similar image to the user query, at least one among the plurality of images that has a respective one among the plurality of similarity values that corresponds to a highest rank among the obtained rank of each of the plurality of similarity values. The similarity neural network is trained with a divergence neural network for outputting a divergence between a first distribution of first similarity values for positive pairs, among the plurality of similarity values, and a second distribution of second similarity values for negative pairs, among the plurality of similarity values.

Abstract: Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency, such as accuracy of learning, accuracy of prediction, speed of learning, performance of learning, and energy efficiency of learning. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Each compute element has processing resources and memory resources. Each router enables communication via wavelets with at least nearest neighbors in a 2D mesh. Stochastic gradient descent, mini-batch gradient descent, and continuous propagation gradient descent are techniques usable to train weights of a neural network modeled by the processing elements. Reverse checkpoint is usable to reduce memory usage during the training.

Abstract: A system comprises a memory that stores computer-executable components; and a processor, operably coupled to the memory, that executes the computer-executable components. The system includes a receiving component that receives a corpus of data; a relation extraction component that generates noisy knowledge graphs from the corpus; and a training component that acquires global representations of entities and relation by training from output of the relation extraction component.

Abstract: A near memory system is provided for the calculation of a layer in a machine learning application. The near memory system includes an array of memory cells for storing an array of filter weights. A multiply-and-accumulate circuit couples to columns of the array to form the calculation of the layer.

Abstract: Weighted population code in neuromorphic systems is provided. According to an embodiment, a plurality of input values is received. For each of the plurality of values, a plurality of spikes is generated. Each of the plurality of spikes has an associated weight. A consumption time is determined for each of the plurality of spikes. Each of the plurality of spikes is sent for consumption at its consumption time.

Abstract: Various embodiments are provided for automating decision making for a neural architecture search by one or more processors in a computing system. One or more specifications may be automatically selected for a dataset, tasks, and one or more constraints for a neural architecture search. The neural architecture search may be performed based on the one or more specifications. A deep learning model may be suggested, predicted, and/or configured for the dataset, the tasks, and the one or more constraints based on the neural architecture search.

Abstract: A disclosed neural network processing system includes a host computer system, a RAMs coupled to the host computer system, and neural network accelerators coupled to the RAMs, respectively. The host computer system is configured with software that when executed causes the host computer system to write input data and work requests to the RAMS. Each work request specifies a subset of neural network operations to perform and memory locations in a RAM of the input data and parameters. A graph of dependencies among neural network operations is built and additional dependencies added. The operations are partitioned into coarse grain tasks and fine grain subtasks for optimal scheduling for parallel execution. The subtasks are scheduled to accelerator kernels of matching capabilities. Each neural network accelerator is configured to read a work request from the respective RAM and perform the subset of neural network operations on the input data using the parameters.

Abstract: Systems and methods for fine tuning a trained artificial neural network (ANN) are provided. An example method may include receiving a description of the neurons, a first set of first parameters for the neurons and a second set of second parameters for the neurons; acquiring a plurality of inputs to the neurons, the inputs including first inputs associated with the first set of first parameters and second inputs associated with the second set of second parameters; obtaining first values correlating the first inputs and the second inputs; obtaining second values correlating the first inputs and the second inputs being weighted partially by the first parameters or the second parameters; and determining, based on the first values and the second values, a third set of third parameters to minimize a distance between neurons outputs determined based on the first parameters and neurons outputs determined based the third parameters.

Abstract: A machine learning management apparatus identifies a maximum prediction performance score amongst a plurality of prediction performance scores corresponding to a plurality of models generated by executing each of a plurality of machine learning algorithms. As for a first machine learning algorithm having generated a model corresponding to the maximum prediction performance score, the machine learning management apparatus determines a first training dataset size to be used when the first machine learning algorithm is executed next time based on the maximum prediction performance score, first estimated prediction performance scores, and first estimated runtimes.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network used to select actions performed by a reinforcement learning agent interacting with an environment. In one aspect, a method includes maintaining a replay memory, where the replay memory stores pieces of experience data generated as a result of the reinforcement learning agent interacting with the environment. Each piece of experience data is associated with a respective expected learning progress measure that is a measure of an expected amount of progress made in the training of the neural network if the neural network is trained on the piece of experience data. The method further includes selecting a piece of experience data from the replay memory by prioritizing for selection pieces of experience data having relatively higher expected learning progress measures and training the neural network on the selected piece of experience data.