Abstract: In some embodiments, a method includes providing an indication of a first file having a first characteristic to a neural network and receiving a classification associated with the first file from the neural network. The method includes providing an indication of a second file having a second characteristic to the neural network and receiving a classification associated with the second file from the neural network. The method further includes calculating a shared importance value for each node from a set of nodes in the neural network. The shared importance value indicates an amount to which that node is used to produce both the classification associated with the first file and the classification associated with the second file. The method further includes modifying the neural network based on the shared importance for at least one node from the set of nodes.

Abstract: Systems and methods for determining predictive model types are provided. A method may include generating a predictive model for a web page of a web site, wherein the web page includes a configuration defining one or more objects presented with the web page, and wherein each object is associated with a predictive model. The method may include determining one or more predictive model types that are associated with the predictive model, determining one or more performance indicators that correspond to each determined predictive model type, wherein performance indicators represent one or more benefits to a website, selecting a predictive model type of the predictive model out of the one or more predictive model types, wherein the predictive model type is selected based on a performance indicator corresponding to the selected predictive model type, and determining a configuration of the web page using the selected predictive model type of the predictive model.

Abstract: In accordance with an embodiment, described herein is a system and method for predicting artists that create media content who are more likely to increase in popularity. Users are determined who requested playback of media content items associated with one or more generators of popular media content within a window of time. One or more early adopters are determined from these users based on a quantity of the one or more generators of popular media content whose media content items were requested for playback by the users. Artists that create media content who are more likely to increase in popularity than other artists that create media content are then predicted based on following further requested playback of media content items by the one or more early adopters.

Abstract: The present disclosure provides a bit quantization method of an artificial neural network. This method may include: (a) of selecting one parameter or one parameter group to be quantized in the artificial neural network; (b) a bit quantizing to reduce the data representation size for the selected parameter or parameter group to a unit of bits; (c) of determining whether the accuracy of the artificial neural network is equal to or greater than a predetermined target value; and (d) repeating steps (a) to (c) when the accuracy of the artificial neural network is equal to or greater than the target value.

Abstract: A method includes identifying a graph of a social network, the graph including nodes and edges, each edge connects two nodes, some of the plurality of nodes represent members of the social network, some edges of the plurality of edges represent a relationship between two associated nodes, creating a first taste profile for a first member node, the taste profile identifying a first entity of interest to the first member, identifying a second member node based on an absence of a second taste profile for the second member node, the second member node is connected to the first member node, determining that the second member node is connected to the first member node, creating a second taste profile for the second member node using the first taste profile, and providing a recommendation to a member associated with the second member node based on the created second taste profile.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing network inputs using an attention neural network that has one or more sparse attention sub-layers. Each sparse attention sub-layer is configured to apply a sparse attention mechanism that attends differently for input positions that are in a first proper subset of the input positions in the input to the sub-layer than for positions that are not in the first proper subset.

Abstract: Human knowledge may be injected in an explainable AI system in order to improve the model's generalization error, model accuracy, interpretability of the model, avoid or eliminate bias, while providing a path towards the integration of connectionist systems with symbolic logic in a combined AI system. Human knowledge injection may be implemented by harnessing the white-box nature of explainable/interpretable models. In one exemplary embodiment, a user applies intuition to model-specific cases or exceptions. In another embodiment, an explainable model may be embedded in workflow systems which enable users to apply pre-hoc and post-hoc operations. A third exemplary embodiment implements human-assisted focusing. An exemplary embodiment also presents a method to train and refine explainable or interpretable models without losing the injected knowledge defined by humans when applying gradient descent techniques.

Abstract: Disclosed is a data processing system that includes compile time logic configured to process a processing graph to generate a modified processing graph, which includes a plurality of forward processing nodes of a forward pass and a plurality of backward processing nodes of a backward pass. The data processing system also includes runtime logic configured with the compile time logic to execute the modified processing graph to generate, at a backward processing node of the plurality of backward processing nodes, a plurality of partial weight gradients, based on processing a corresponding plurality of gradient tiles of a gradient tensor, and generate, based on the plurality of partial weight gradients, a final weight gradient corresponding to the gradient tensor.

Abstract: Techniques are provided for acoustic event detection. A methodology implementing the techniques according to an embodiment includes extracting acoustic features from a received audio signal. The acoustic features may include, for example, one or more short-term Fourier transform frames, or other spectral energy characteristics, of the audio signal. The method also includes applying a trained classifier to the extracted acoustic features to identify and label acoustic event subparts of the audio signal and to generate scores associated with the subparts. The method further includes performing sequence decoding of the acoustic event subparts and associated scores to detect target acoustic events of interest based on the scores and temporal ordering sequence of the event subparts. The classifier is trained on acoustic event subparts that are generated through unsupervised subspace clustering techniques applied to training data that includes target acoustic events.

Abstract: Logic may quantize a primary neural network. Logic may generate, by a secondary neural network logic circuitry for a primary neural network logic circuitry, quantization parameters. The primary neural network logic circuitry may comprise a primary neural network with multiple layers trainable with an objective function. Each of the multiple layers of the primary neural network may comprise multiple tensors. The secondary neural network logic circuitry may comprise one or more secondary neural networks trainable with the objective function to output the quantization parameters to the tensors.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for knowledge-preserving sparse pruning on neural networks are described. An exemplary method includes obtaining a pre-trained machine learning model trained based on a plurality of general-purpose training data; training a task-specific machine learning model by tuning the pre-trained machine learning model based on a plurality of task-specific training data corresponding to a task; constructing a student network based on the task-specific machine learning model; simultaneously performing (1) knowledge distillation from the trained task-specific machine learning model as a teacher network to the student network and (2) network pruning on the student network; and obtaining the trained student network for serving the task.

Abstract: A mechanism is provided in a data processing system comprising a processor and a memory. The memory comprises instructions which are executed by the processor to cause the processor to be specifically configured to implement a recognizer module for detecting user input in a cognitive environment. The recognizer module receives sensor signals from at least one wearable device being worn by a user. The recognizer module analyzes the sensor signals using a machine learning model to determine at least one user input indicator describing user input activity of the user. The recognizer module communicates the at least one user input indicator to a cognitive system executing within the cognitive environment. The cognitive system performs at least one cognitive action based on the at least one user input indicator.

Abstract: One embodiment provides a method comprising receiving a training set comprising a plurality of data points, where a neural network is trained as a classifier based on the training set. The method further comprises, for each data point of the training set, classifying the data point with one of a plurality of classification labels using the trained neural network, and recording neuronal activations of a portion of the trained neural network in response to the data point. The method further comprises, for each classification label that a portion of the training set has been classified with, clustering a portion of all recorded neuronal activations that are in response to the portion of the training set, and detecting one or more poisonous data points in the portion of the training set based on the clustering.

Abstract: Method and apparatus for detecting vibrational and/or acoustic transfers in a mechanical system A method and apparatus for detecting vibro-acoustic transfers in a mechanical system are provided. The method comprises: while operating the mechanical system, acquiring, at each of multiple input points, an input signal indicative of a mechanical load acting on the input point, and acquiring, at a response point, a response signal indicative of a mechanical response; training a neural network device using the input signals acquired at the input points and using the response signal acquired at the response point; and, for each of the input points: providing only the input signal acquired at the respective input point to the trained neural network device and obtaining, from the neural network device, a contribution signal indicative of a predicted contribution of the respective input signal to the response signal.

Abstract: The invention is directed to a hybrid architecture that comprises a stacked autoencoder and a deep echo state layer for temporal pattern discovery in high-dimensional sequence data. The stacked autoencoder plays a preprocessing role that exploits spatial structure in data and creates a compact representation. The compact representation is then fed to the echo state layer in order to generate a short-term memory of the inputs. The output of the network may be trained to generate any target output.

Abstract: A method includes identifying a graph of a social network, the graph including nodes and edges, each edge connects two nodes, some of the plurality of nodes represent members of the social network, some edges of the plurality of edges represent a relationship between two associated nodes, creating a first taste profile for a first member node, the taste profile identifying a first entity of interest to the first member, identifying a second member node based on an absence of a second taste profile for the second member node, the second member node is connected to the first member node, determining that the second member node is connected to the first member node, creating a second taste profile for the second member node using the first taste profile, and providing a recommendation to a member associated with the second member node based on the created second taste profile.

Abstract: Methods and system for converting a plurality of weights of a filter of a Deep Neural Network (DNN) in a first number format to a second number format, the second number format having less precision than the first number format, to enable the DNN to be implemented in hardware logic.

Abstract: A system configured for learning new trained concepts used to retrieve content relevant to the concepts learned. The system may comprise one or more hardware processors configured by machine-readable instructions to obtain one or more digital media items. The one or more hardware processors may be further configured to obtain an indication conveying a concept to be learned from the one or more digital media items. The one or more hardware processors may be further configured to receive feedback associated with individual ones of the one or more digital media items. The one or more hardware processors may be configured to obtain individual neural network representations for the individual ones of the one or more digital media items. The one or more hardware processors may be configured to determine a trained concept based on the feedback and the neural network representations of the one or more digital media items.

Abstract: A synchrophasor measurement-based disturbance identification method is described considering different penetration levels of renewable energy. A differential Teager-Kaiser energy operator (dTKEO)-based algorithm is first utilized to improve multiple-disturbances detection accuracy. Then, feature extractions via the integrated additive angular margin (AAM) loss and the long short-term memory (LSTM) network is described. This enables one to deal with intra-class similarity and inter-class variance of disturbances when high penetration renewable energy occurs. With the extracted features, a multi-stage weighted summing (MSWS) loss-based criterion is described for adaptive data window determination and fast disturbance pre-classification. Finally, the re-identification model based on feature similarity is established to identify unknown disturbances, a challenge for existing machine learning algorithms.

Abstract: Mechanisms are provided to implement a medical coding engine to perform medical coding using a neural network architecture that leverages hierarchical semantics between medical concepts. The medical coding engine configures a medical coding neural network to comprise an first layer of nodes comprising preferred terminology (PT) nodes, a second layer comprising lowest level terminology (LLT) nodes, and a third layer comprising weighted values for each connection between each PT node and each LLT node forming a PT node/LLT node connection. Responsive to receiving an adverse event from a cognitive system, a PT node is identified in the first layer associated with a citation from the adverse event. One or more nodes are identified from the second layer based on the identification PT node and a weight associated with the PT node/LLT node connection. A medical code associated with each the one or more LLT nodes is then output.