Abstract: An approach is provided for processing received feature vectors whose indices are in an original hierarchical index space. A transformation of each mode of the feature vectors is generated and includes a pooling operation, a transformation operation, and an unpooling operation. The pooling operation aggregates and collapses the feature vectors to produce a coarse-grained feature set having a collapsed index space. The transformation operation transforms the coarse-grained feature set into a transformed feature set that preserves the collapsed index space. The unpooling operation generates a final feature set in the original hierarchical index space by reverting pooling from the transformed feature set. The feature vectors are transformed without any pooling or unpooling to produce an additional set of transformed feature vectors. A summation that preserves the original hierarchical index space is generated by summing the final feature set and the additional set of transformed feature vectors.

Abstract: Computer-implemented methods for performing a data processing task on a data set with a pretrained language model are provided. Aspects include obtaining the data set and a type of the data processing task to be performed on the data set, generating a prompt by inputting data from the data set into a template, and inputting the prompt into an encoder of a pretrained language model. Aspects also include obtaining, from the encoder, a set of prompt embeddings and a set of token embeddings, inputting the set of prompt embeddings into a trained neural network, and obtaining, from the trained neural network, a prefix vector. Aspects further include inputting a set of extended embeddings that are created by appending the set of token embeddings to the prefix vector into a decoder of the pretrained language model, obtaining, from the decoder, an output, and modifying the data set based on the output.

Abstract: A method, computer system, and a computer program product are provided. A visual inspection machine learning model is trained using a generative adversarial network. Within the generative adversarial network a vector bypass is implemented. By transmitting a vector embedding representation of an unlabeled image through the vector bypass, the vector embedding representation is transmitted around the visual inspection machine learning model and to a generator to assist with image reconstruction.

Abstract: A computer-implemented method for semantic segmentation includes constructing a co-occurrence table that includes co-occurrences of predictions of a pre-trained model for base classes and labels for novel classes from the pre-trained model for base classes and from training data with novel classes. Classifiers are trained that associated with a base class and that classify an input into the base class and one of the novel classes that have co-occurrences with the base class according to the co-occurrence. A prediction is fused from the pre-trained model and the trained classifiers to obtain a final prediction result as a fully labeled image.

Abstract: Described are techniques for establishing a multi-cloud environment. A multi-cloud environment is modeled using a game theoretic model. Furthermore, a probability distribution over information technology costs for both the cloud buyers and cloud providers is created. Additionally, an estimate of the expected utility (measure of how much benefit a party receives or is expected to receive) for each party in a single-cloud environment is calculated. A mechanism (centralized broker) for the multi-cloud environment is then created that collects payments from the cloud buyers and makes payments to the cloud providers using the expected utility for each party in the single-cloud environment and the probability distribution over information technology costs for both the cloud buyers and the cloud providers as parameters for the game theoretic model, such as by applying an automated mechanism design approach to the game theoretic model.

Abstract: A computer-implemented method for modifying a current policy using reinforcement learning (RL) includes the following operations. A number, corresponding to an inputted sample size, of Markov Decision Processes (MDPs) defining an environment are sampled. For each of the sampled MDPs, behavior data for the current policy is collected, a quantile function of return with the current policy is determined using the collected behavior data, and a current weight is generated by updating a weight for a particular sampled MDP using the quantile function of return for the particular sampled MDP. The policy is modified based upon the weights for each of the sampled MDPs. The current weights are generated by minimizing a conditional value of at risk (CVaR) of a return of the current policy, and the policy is modified to maximize a weighted average of the CVaR of the return with the current weights.

Abstract: A method for performing offline distributional reinforcement learning. The method includes randomly sampling a dataset comprising historical training data between an agent and an environment to generate a minibatch of the historical training data; updating a plurality of predictors, using non-uniform underestimation, based on the minibatch; and updating a policy using the updated plurality of predictors and the minibatch.

Abstract: A method for a determinantal Point Process-based prediction includes obtaining, using a hardware processor, a training data set stored on one or more computer readable storage mediums operably coupled to the hardware processor, training an asymmetric kernel of a Determinantal Point Process (DPP) from a training data set by calculating an inverse matrix of a sum of the asymmetric kernel and an identity matrix in a recursive manner to reduce time and computational resources utilized, and determining a prediction model by training the asymmetric kernel as at least part of a prediction model to make a prediction.

Abstract: A method is provided for choosing an action of an agent in a first team that competes against a second team. The method includes determining an action, based on first, second and third types of local payoff matrices. The method further includes performing the action. The determining step includes representing, by the first type of local payoff matrices, a payoff to the first team due to a pairwise interaction between agent teammates of the first team. The determining step further includes representing, by the second type of local payoff matrices, a payoff to the first team due to a pairwise interaction between agent opponents from the first team and the second team. The determining step also includes representing, by the third type of local payoff matrices, a payoff to the first team due to a pairwise interaction between agent teammates of the second team.

Abstract: A computer-implemented method, a computer program product, and a computer processing system are provided for online learning for a Dynamic Boltzmann Machine (DyBM) with hidden units. The method includes imposing, by a processor device, limited connections in the DyBM where (i) a current observation x[t] depends only on latest hidden units h[t-1/2] and all previous observations x[<t] and (ii) the latest hidden units h[t-1/2] depend on all the previous observations x[<t] while being independent of older hidden units h[t-1/2]. The method further includes computing, by the processor device, gradients of an objective function. The method also includes optimizing, by the processor device, the objective function in polynomial time using a stochastic Gradient Descent algorithm applied to the gradients.

Abstract: A computer-implemented method comprising: receiving data associated with an operational control problem; formulating the operation control problem as an optimization problem; recursively generating a sequence of policies of operational control associated with the operational control problem, wherein each subsequent policy in the sequence is constructed by modifying one or more actions at a single state in a preceding policy in the sequence, and wherein the modifying monotonically changes a risk associated with the subsequent policy; constructing, from the sequence of policies, an optimal solution path, wherein each vertex on the optimal solution path represents an optimal solution to the operational control problem; calculating a ratio of reward to risk for each of the vertices on the path; and selecting one of the policies in the sequence to apply to the operational control problem, based, at least in part, on the calculated ratios.

Abstract: A method, a computer program product, and a system of adversarial semi-supervised one-shot training using a data stream. The method includes receiving a data stream based on an observation, wherein the data stream includes unlabeled data and labeled data. The method also includes training a prediction model with the labeled data using stochastic gradient descent based on a classification loss and an adversarial term and training a representation model with the labeled data and the unlabeled data based on a reconstruction loss and the adversarial term. The adversarial term is a cross-entropy between the middle layer output data from the models. The classification loss is a cross-entropy between the labeled data and an output from the prediction model. The method further includes updating a discriminator with middle layer output data from the prediction model and the representation model and based on a discrimination loss, and discarding the data stream.

Abstract: A neuromorphic chip includes synaptic cells including respective resistive devices, axon lines, dendrite lines and switches. The synaptic cells are connected to the axon lines and dendrite lines to form a crossbar array. The axon lines are configured to receive input data and to supply the input data to the synaptic cells. The dendrite lines are configured to receive output data and to supply the output data via one or more respective output lines. A given one of the switches is configured to connect an input terminal to one or more input lines and to changeably connect its one or more output terminals to a given one or more axon lines.

Abstract: An n-steps-ahead value of time-series data is estimated by a prediction model configured to output a sum of discounted m-th order differences of adjacent time steps at each time step, wherein an m-th order difference at a corresponding time step is discounted by using a discount factor such that an m-th order difference discount increases as a time step of the m-th order difference increases.

Abstract: In an approach, a processor obtains a target base strategy for selecting actions of a target agent. A processor obtains an adversarial base strategy for selecting adversarial actions of an adversarial agent. A processor calculates, for each candidate action among a plurality of candidate actions of the target agent, a risk measure of the candidate action based on the adversarial base strategy and a payoff to the target agent in a case where the target agent takes the candidate action and the adversarial agent takes an adversarial action based on the adversarial base strategy. A processor generates a target strategy by adjusting the target base strategy based on the risk measure for each candidate action.

Abstract: A computer implemented method determines a policy for risk sensitive decisions. A computer system receives state and action pairs. The computer system, with initial probabilistic discounted entropic risk measure values for the state and action pairs, determines in a recursive manner current probabilistic discounted entropic risk measure values for the state and action pairs based on a risk factor until the current probabilistic discounted entropic risk measure values reach a desired level. The current probabilistic discounted entropic risk measure values are the initial probabilistic discounted entropic risk measure values for a next determination.

Abstract: A neuromorphic chip includes synaptic cells including respective resistive devices, axon lines, dendrite lines and switches. The synaptic cells are connected to the axon lines and dendrite lines to form a crossbar array. The axon lines are configured to receive input data and to supply the input data to the synaptic cells. The dendrite lines are configured to receive output data and to supply the output data via one or more respective output lines. A given one of the switches is configured to connect an input terminal to one or more input lines and to changeably connect its one or more output terminals to a given one or more axon lines.

Abstract: A cumulative reward of a target system type is predicted by training a prediction model by performing an iteration for each time step. The iteration includes recursively updating a matrix by using the weighted difference between an eligibility trace of a current time step and an eligibility trace of a previous time step, recursively updating a vector by using a reward of a subsequent time step and the eligibility trace of the current time step, and recursively updating an eligibility trace of a subsequent time step by using a feature vector of the subsequent time step. Each feature vector is an encoded representation of a state of a training system of the target system type at a corresponding time step. The matrix and the vector are output as the prediction model for estimating the cumulative reward of a target time step of a target system of the target system type.

Abstract: A computer-implemented method is provided for machine prediction. The method includes forming, by a hardware processor, a Convolutional Dynamic Boltzmann Machine (C-DyBM) by extending a non-convolutional DyBM with a convolutional operation. The method further includes generating, by the hardware processor using the convolution operation of the C-DyBM, a prediction of a future event at time t from a past patch of time-series of observations. The method also includes performing, by the hardware processor, a physical action responsive to the prediction of the future event at time t.

Abstract: An apparatus for implementing a computing system to predict preferences includes at least one processor device operatively coupled to a memory. The at least one processor device is configured to calculate a parameter relating to a density of a prior distribution at each sample of a set of samples associated with the prior distribution. The at least one parameter including a distance from each sample to at least one neighboring sample. The at least one processor device is further configured to estimate, for the plurality of samples, at least one differential entropy of at least one posterior distribution associated with at least one observation based on the parameter relating to the density of the prior distribution at each sample and the likelihood of observation for each sample. The estimation is performed without sampling the at least one posterior distribution to reduce consumption of resources of the computing system.