Abstract: The present disclosure advantageously provides a system including a memory, a processor, and a circuitry to execute one or more mixed precision layers of an artificial neural network (ANN), each mixed precision layer including high-precision weight filters and low precision weight filters. The circuitry is configured to perform one or more calculations on an input feature map having a plurality of input channels (cin) using the high precision weight filters to create a high precision output feature map having a first number of output channels (k), perform one or more calculations on the input feature map using the low precision weight filters to create a low precision output feature map having a second number of output channels (cout?k), and concatenate the high precision output feature map and the low precision output feature map to create a unified output feature map having a plurality of output channels (cout).

Abstract: A BGC prediction system identifies candidate biosynthetic gene clusters (BGCs) within genomes using machine-learned models, such as a shallow neural network and recurrent neural network (RNN). A set of domains within a genome sequence are identified, each domain corresponds to a set of domain identifiers. A shallow neural network block is applied to each set of domain identifiers to produce a set of vectors. An RNN block is applied to the set of vectors to produce a BGC class score for each domain. The RNN block was trained using an identified set of positive vectors, which represents known BGCs, and a synthesized set of negative vectors, which is unlikely to represent BGCs. Candidate BGCs are selected by averaging BGC class scores across genes within a domain and comparing the average BGC class scores to a threshold. The candidate BGCs are provided for display on a user interface.

Abstract: Mechanisms are provided for implementing a generative adversarial network (GAN) based restoration system. A first neural network of a generator of the GAN based restoration system is trained to generate an artificial audio spectrogram having a target damage characteristic based on an input audio spectrogram and a target damage vector. An original audio recording spectrogram is input to the trained generator, where the original audio recording spectrogram corresponds to an original audio recording and an input target damage vector. The trained generator processes the original audio recording spectrogram to generate an artificial audio recording spectrogram having a level of damage corresponding to the input target damage vector. A spectrogram inversion module converts the artificial audio recording spectrogram to an artificial audio recording waveform output.

Abstract: A learning device includes an input section and a learning section. The input section obtains training data in which an image containing an image of equipment of a vehicle, first information indicating a skeleton location of a specific part of a vehicle-occupant, and second information indicating a state of the vehicle-occupant with respect to the equipment are associated with each other. The learning section forms a model such that an estimating device can obtain the first information and the second information associated with the image containing the image of the equipment.

Abstract: Computing technology for managing support requests are provided. The technology includes a processor executable application programming interface (API) that receives a support case indicating a problem associated with a device. The API utilizes a training model to predict a problem category for the support case. The training model predicts the problem category based on a feature extracted from information included in the support case. The training model further identifies a plurality of proximate support cases based on a distance between the support case and the proximate support cases within a virtual space assigned to the predicted problem category; determines relevance of each proximate support case to the support case; and outputs a resolution code for the support case based on the determined relevance of each proximate support case.

Abstract: A sparse convolutional neural network accelerator system that dynamically and efficiently identifies fine-grained parallelism in sparse convolution operations. The system determines matching pairs of non-zero input activations and weights from the compacted input activation and weight arrays utilizing a scalable, dynamic parallelism discovery unit (PDU) that performs a parallel search on the input activation array and the weight array to identify reducible input activation and weight pairs.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for providing a machine learning model that is trained to perform a machine learning task. In one aspect, a method comprises receiving a request to train a machine learning model on a set of training examples; determining a set of one or more meta-data values characterizing the set of training examples; using a mapping function to map the set of meta-data values characterizing the set of training examples to data identifying a particular machine learning model architecture; selecting, using the particular machine learning model architecture, a final machine learning model architecture for performing the machine learning task; and training a machine learning model having the final machine learning model architecture on the set of training examples.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for reinforcement learning with scheduled auxiliary tasks. In one aspect, a method includes maintaining data specifying parameter values for a primary policy neural network and one or more auxiliary neural networks; at each of a plurality of selection time steps during a training episode comprising a plurality of time steps: receiving an observation, selecting a current task for the selection time step using a task scheduling policy, processing an input comprising the observation using the policy neural network corresponding to the selected current task to select an action to be performed by the agent in response to the observation, and causing the agent to perform the selected action.

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: An indication of semantic relationships among classes is obtained. A neural network whose loss function is based at least partly on the semantic relationships is trained. The trained neural network is used to identify one or more classes to which an input observation belongs.

Abstract: A risk management method, system, and non-transitory computer readable medium, include a data analyzing circuit configured to analyze user data, site data, and equipment data to map prior behavior types to an event on a site, a relationship determining circuit configured to determine a relationship between the mapped data and the event based on behaviors exhibited by the user and an impact on a performance factor and a risk factor, and a reinforcement learning circuit configured to use reinforcement learning to learn the performance factor to the risk factor ratio to optimize an overall site productivity.

Abstract: Systems and methods are disclosed for sorting elements, such as bids in an auction environment or marketplace for the display of an advertisement on a web page. According to one implementation, a plurality of elements may be received over a network. For example, the plurality of elements may comprise a plurality of bids, each of which may include, for example, a bid price, a bid allocation, and a bid target. The received set of elements may be recursively partitioned into a plurality of element subsets. For example, the element set may be partitioned about a pivot selected from a sample set of elements chosen from the element set. A subset of elements may be identified from the element set as eligible elements (e.g., bids eligible for selection as a winning bid). Once the set of eligible elements has been identified, the recursive partitioning of elements may be terminated.

Abstract: Methods, devices, systems, and instructions for adaptive quantization in an artificial neural network (ANN) calculate a distribution of ANN information; select a quantization function from a set of quantization functions based on the distribution; apply the quantization function to the ANN information to generate quantized ANN information; load the quantized ANN information into the ANN; and generate an output based on the quantized ANN information. Some examples recalculate the distribution of ANN information and reselect the quantization function from the set of quantization functions based on the resampled distribution if the output does not sufficiently correlate with a known correct output. In some examples, the ANN information includes a set of training data. In some examples, the ANN information includes a plurality of link weights.

Abstract: Techniques for facilitating incremental static program analysis based on machine learning techniques are provided. In one example, a system comprises a feature component that, in response to an update to a computer program, generates feature vector data representing the update, wherein the feature vector data comprises feature data representing a feature of the update derived from an abstract state of the computer program, and wherein the abstract state is based on a mathematical model of the computer program that is generated in response to static program analysis of the computer program. The system can further comprise a machine learning component that employs a classifier algorithm to identify an affected portion of the mathematical model that is affected by the update. The system can further comprise an incremental analysis component that incrementally applies the static program analysis to the computer program based on the affected portion.

Abstract: First, the system obtains time-series sensor data. Next, the system identifies missing values in the time-series sensor data, and fills in the missing values through interpolation. The system then divides the time-series sensor data into a training set and an estimation set. Next, the system trains an inferential model on the training set, and uses the inferential model to replace interpolated values in the estimation set with inferential estimates. If there exist interpolated values in the training set, the system switches the training and estimation sets. The system trains a new inferential model on the new training set, and uses the new inferential model to replace interpolated values in the new estimation set with inferential estimates. The system then switches back the training and estimation sets. Finally, the system combines the training and estimation sets to produce preprocessed time-series sensor data, wherein missing values are filled in with imputed values.

Abstract: In one embodiment, a matrix operation may be performed using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations. The plurality of input matrices may be partitioned into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements. The plurality of input partitions may be distributed among a plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements. A plurality of partial matrix operations may be performed using the plurality of processing elements, and partial matrix data may be transmitted between the plurality of processing elements while performing the plurality of partial matrix operations. A result of the matrix operation may be determined based on the plurality of partial matrix operations.

Abstract: A method includes providing input data to a plurality of base models to generate a plurality of intermediate outputs. The base models are non-linear in that different base models are specialized differently such that the different base models are complementary to one another. Each of the base models is generated using a different base classification algorithm in a multi-layered machine learning system. The method also includes processing the intermediate outputs using a fusion model to generate a final output associated with the input data. The fusion model is generated using a meta classification algorithm in the multi-layered machine learning system. The method may also include training the classification algorithms, where training data used by each of at least one of the base classification algorithms is selected based on an uncertainty associated with at least one other of the base classification algorithms.

Abstract: Herein, horizontally scalable techniques efficiently configure machine learning algorithms for optimal accuracy and without informed inputs. In an embodiment, for each particular hyperparameter, and for each epoch, a computer processes the particular hyperparameter. An epoch explores one hyperparameter based on hyperparameter tuples. A respective score is calculated from each tuple. The tuple contains a distinct combination of values, each of which is contained in a value range of a distinct hyperparameter. All values of a tuple that belong to the particular hyperparameter are distinct. All values of a tuple that belong to other hyperparameters are held constant. The value range of the particular hyperparameter is narrowed based on an intersection point of a first line based on the scores and a second line based on the scores. A machine learning algorithm is optimally configured from repeatedly narrowed value ranges of hyperparameters. The configured algorithm is invoked to obtain a result.

Abstract: A method for training a machine learning model includes: receiving, by a computer system including a processor and memory, a training data set including imbalanced data; computing, by the computer system, a label density fX(x) in the training data set, computing, by the computer system, a weight function w(x) including a term that is inversely proportional to the label density; weighting, by the computer system, a loss function (x, {circumflex over (x)}) in accordance with the weight function to generate a weighted loss function w(x, {circumflex over (x)}); training, by the computer system, a continuous machine learning model in accordance with the training data set and the weighted loss function w(x, {circumflex over (x)}); and outputting, by the computer system, the trained continuous machine learning model.

Abstract: Systems and methods are disclosed for inquiry-based deep learning. In one implementation, a first content segment is selected from a body of content. The content segment includes a first content element. The first content segment is compared to a second content segment to identify a content element present in the first content segment that is not present in the second content segment. Based on an identification of the content element present in the first content segment that is not present in the second content segment, the content element is stored in a session memory. A first question is generated based on the first content segment. The session memory is processed to compute an answer to the first question. An action is initiated based on the answer. Using deep learning, content segments can be encoded into memory. Incremental questioning can serve to focus various deep learning operations on certain content segments.