Google Inc.

Cognizant Technology Solutions U.S. Corporation

Columbia University

Evolv Technology Solutions, Inc.

Sentient Technologies (barbados) Limited

Philip Michael Long

Nigel Duffy

Rocco Servedio

Diego Guy M Legrand

Risto Miikkulainen

Hormoz Shahrzad

Roger N Anderson

Vineet Gupta

Albert Boulanger

Olivier Francon

Hanie Sedghi

Petr Tsatsin

Diego Legrand

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a convolutional neural network using a regularization scheme. One of the methods includes repeatedly performing the following operations: obtaining a kernel of a particular convolutional layer; applying a Fourier transform to the kernel; generating a decomposition using singular-value decomposition (SVD); generating a regularized diagonal matrix; generating a recomposition; applying an inverse Fourier transform to the recomposition; and training the convolutional neural network on training inputs.

Regularizing the training of convolutional neural networks

A method for finding a solution to a problem is provided. The method includes storing candidate individuals in a candidate pool and evolving the candidate individuals by performing steps including (i) testing each of the candidate individuals to obtain test results, (ii) assigning a performance measure to the tested candidate individuals, (iii) discarding candidate individuals from the candidate pool in dependence upon their assigned performance measure, and (iv) adding, to the candidate pool, a new candidate individual procreated from candidate individuals remaining in the candidate pool. The method further includes selecting, as the winning candidate individual, a candidate individual having a best neighborhood performance measure, where the neighborhood performance measure of a particular candidate individual is given by the performance measures of (i) the particular candidate individual and (ii) K neighborhood candidate individuals which are nearest in the candidate pool to the particular candidate individual, and where K&#x3e;0.

Method and system for finding a solution to a provided problem by selecting a winner in evolutionary optimization of a genetic algorithm

The technology disclosed introduces a concept of training a neural network to create an embedding space. The neural network is trained by providing a set of K+2 training documents, each training document being represented by a training vector x, the set including a target document represented by a vector x, a favored document represented by a vector x, and K&#x3e;1 unfavored documents represented by vectors x, each of the vectors including input vector elements, passing the vector representing each document set through the neural network to derive an output vectors y, yand y, each output vector including output vector elements, the neural network including adjustable parameters which dictate an amount of influence imposed on each input vector element to derive each output vector element, adjusting the parameters of the neural network to reduce a loss, which is an average over all of the output vectors yof [D(y,y)&#x2212;D(y, y)].

Content embedding using deep metric learning algorithms

Roughly described, a system for user identification of a desired document. A database identifies a catalog of documents in an embedding space, in which the distance between documents corresponds to a measure of their dissimilarity. The system presents an initial collection of the documents toward the user from an initial candidate space which is part of the embedding space, then in response to iterative user input, refines the candidate space and subsequent collections of documents presented toward the user. The initial collection is determined using a weighted cost-based iterative addition to the initial collection of documents from the initial candidate space, trading off between two sub-objectives of representativeness and diversity.

Selection of initial document collection for visual interactive search

A method for identifying a desired document is provided to include calculating a Prior probability score for each document of a candidate list including a portion of documents of an embedding space, the Prior probability score indicating a preliminary probability, for each document of the candidate list, that the document is the desired document, and identifying an initial (i=0) collection of N0&#x3e;1 candidate documents from the candidate list in dependence on the calculated Prior probability scores, the initial collection of candidate documents having fewer documents than the candidate list. The method further includes, for each i'th iteration in a plurality of iterations, beginning with a first iteration (i=1) and in response to user selection of an i'th selected document from the (i&#x2212;1)'th collection of candidate documents, identifying an i'th collection of Ni&#x3e;1 candidate documents from the candidate list in dependence on Posterior probability scores.

Bayesian visual interactive search

A computationally efficient method and apparatus is disclosed for machine learning an unknown, n-dimensional &#x3b3;-margin halfspace to accuracy 1&#x2212;&#x3b5; in the presence of malicious noise, when the noise rate is allowed to be as high as A smooth boosting module repeatedly calls a weak learner module that generates candidate classifiers based on a majority vote of randomly-generated classifiers.

Method and apparatus for machine learning

A method and apparatus is disclosed for machine learning an unknown n-dimensional &#x3b3;-margin halfspace. The method and apparatus can be implemented using parallel processing. In one embodiment, the method includes generating a random d-dimensional projection of the set of examples. The random projection is generated by rounding the examples, taking a projection and then rounding the results. A set of m examples are drawn from the d-dimensional projection. Linear programs are then solved with m constraints based on the m examples using multiple stages of Newton's method. Parallel algorithms for linear algebra can be used to carry out at least some stages of Newton's method.

Method and apparatus for machine learning using a random projection

Boosting algorithms are provided for accelerated machine learning in the presence of misclassification noise. In an exemplary embodiment, a machine learning method having multiple learning stages is provided. Each learning stage may include partitioning examples into bins, choosing a base classifier for each bin, and assigning an example to a bin by counting the number of positive predictions previously made by the base classifier associated with the bin.

Growing community of inventors

Palo Alto, CA, United States of America

Philip Michael Long