Cognizant Technology Solutions U.S. Corporation

Evolv Technology Solutions, Inc.

Cognizant Technology Solutions U.S. Corportion

Risto Miikkulainen

Elliot Meyerson

Hormoz Shahrzad

Neil Iscoe

Jason Zhi Liang

Babak Hodjat

Santiago Gonzalez

Xin Qiu

Myles Brundage

Nigel Duffy

Daniel Edward Fink

Philip Michael Long

Aditya Rawal

Olivier Francon

A process for evolving a data augmentation policy for application to sample data from a dataset for us in training a neural network to perform a predetermined task is described. An initial population of candidate data augmentation policy models, each model including multiple nodes which are distinct data augmentation operations and multiple edges which have weight values representing a probability related to action by a second node on input data from a first node. The models in the population are evaluated by applying to the sample data and at least partially training the neural network using the augmented sample dataset. A fitness is determined based on the results of the training and models are selected either as final policy models or for reproduction and repeating of the evolution process until a final model is selected.

System and method for evolved data augmentation and selection

A process for optimizing loss functions includes progressively building better sets of parameters for loss functions represented as multivariate Taylor expansions in accordance with an iterative process. The optimization process is built upon CMA-ES. At each generation (i.e., each CMA-ES iteration), a new set of candidate parameter vectors is sampled. These candidate parameter vectors are sampled from a multivariate Gaussian distribution representation that is modeled by the CMA-ES covariance matrix and the current mean vector. The candidates are then each evaluated by training a model (neural network) using the candidates and determining a fitness value for each candidate against a validation data set.

Loss function optimization using Taylor series expansion

An evolutionary AutoML framework called LEAF optimizes hyperparameters, network architectures and the size of the network. LEAF makes use of both evolutionary algorithms (EAs) and distributed computing frameworks. A multiobjective evolutionary algorithm is used to maximize the performance and minimize the complexity of the evolved networks simultaneously by calculating the Pareto front given a group of individuals that have been evaluated for multiple objectives.

Multiobjective coevolution of deep neural network architectures

The technology disclosed relates to a webinterface production and deployment system. In particular, it relates to a presentation module that applies a selected candidate individual to a presentation database to determine frontend element values corresponding to dimension values identified by the selected candidate individual, and which presents toward a user a funnel having the determined frontend element values.

Presentation module for webinterface production and deployment system

User-driven exploration functionality, referred to herein as a Scratchpad, is a post-learning extension for machine learning systems. For example, in ESP, consisting of the Predictor (a surrogate model of the domain) and Prescriptor (a solution generator model), the Scratchpad allows the user to modify the suggestions of the Prescriptor, and evaluate each such modification interactively with the Predictor. Thus, the Scratchpad makes it possible for the human expert and the AI to work together in designing better solutions. This interactive exploration also allows the user to conclude that the solutions derived in this process are the best found, making the process trustworthy and transparent to the user.

Framework for interactive exploration, evaluation, and improvement of AI-generated solutions

The technology disclosed relates to webinterface generation and testing to promote a predetermined target user behavior. In particular, the technology disclosed stores a candidate database having a population of candidate individuals. Each of the candidate individuals identify respective values for a plurality of hyperparameters of the candidate individual. The hyperparameters describe topology of a respective neural network and coefficients for interconnects of the respective neural network. The technology disclosed writes a preliminary pool of candidate individuals into the candidate individual population. The technology disclosed tests each of the candidate individuals in the candidate individual population. The technology disclosed adds to the candidate individual population new individuals based on the testing. The technology disclosed repeats the candidate testing and the addition of the new individuals.

Webinterface generation and testing using artificial neural networks

A multi-task (MTL) process is adapted to the single-task learning (STL) case, i.e., when only a single task is available for training. The process is formalized as pseudo-task augmentation (PTA), in which a single task has multiple distinct decoders projecting the output of the shared structure to task predictions. By training the shared structure to solve the same problem in multiple ways, PTA simulates the effect of training towards distinct but closely-related tasks drawn from the same universe. Training dynamics with multiple pseudo-tasks strictly subsumes training with just one, and a class of algorithms is introduced for controlling pseudo-tasks in practice.

System and method for pseudo-task augmentation in deep multitask learning

A genetic loss function optimization (hereafter 'GLO') process uses a genetic algorithm to construct candidate loss functions as trees. The process takes the best candidate loss functions from this set and optimizes the coefficients thereof using covariance-matrix adaptation evolutionary strategy (hereafter “CMA-ES”), resulting in new loss functions showing substantial improvements in accuracy, convergence speed, and data requirements.

System and method for loss function metalearning for faster, more accurate training, and smaller datasets

A computer-implemented method optimizing genetic algorithms for finding solutions to a provided problem is described. The method implements a multi-arm bandit algorithm to determine performance scores for candidate individuals from a candidate pool in dependence on successes and failures of the one or more candidates. The method evolves the candidate individuals in the candidate pool by performing evolution steps including: determining a fitness score for each of the candidate individuals in the candidate pool in dependence on the performance scores for the candidate individuals, discarding candidate individuals from the candidate pool in dependence upon their assigned performance measure, and adding, to the candidate pool, a new candidate individual procreated from candidate individuals remaining in the candidate pool after the discarding of the candidate individuals. This evolution is repeated evolve the candidate individuals in the candidate pool and one or more candidate individuals from the candidate pool is selected based on best neighborhood performance measures, wherein the selected winning candidate individual is a solution to the provided problem.

Enhancing evolutionary optimization in uncertain environments by allocating evaluations via multi-armed bandit algorithms

Roughly described, the technology disclosed provides a so-called machine-learned conversion optimization (MLCO) system that uses artificial neural networks and evolutionary computations to efficiently identify most successful webpage designs in a search space without testing all possible webpage designs in the search space. The search space is defined based on webpage designs provided by marketers. Neural networks are represented as genomes. Neural networks map user attributes from live user traffic to different dimensions and dimension values of output funnels that are presented to the users in real time. The genomes are subjected to evolutionary operations like initialization, testing, competition, and procreation to identify parent genomes that perform well and offspring genomes that are likely to perform well.

Webinterface presentation using artificial neural networks

A surrogate-assisted evolutionary optimization method, ESP, discovers decision strategies in real-world applications. Based on historical data, a surrogate is learned and used to evaluate candidate policies with minimal exploration cost. Extended into sequential decision making, ESP is highly sample efficient, has low variance, and low regret, making the policies reliable and safe. As an unexpected result, the surrogate also regularizes decision making, making it sometimes possible to discover good policies even when direct evolution fails.

Process and system including an optimization engine with evolutionary surrogate-assisted prescriptions

A residual estimation with an I/O kernel ('RIO') framework provides estimates of predictive uncertainty of neural networks, and reduces their point-prediction errors. The process captures neural network ('NN') behavior by estimating their residuals with an I/O kernel using a modified Gaussian process (&#x201c;GP&#x201d;). RIO is applicable to real-world problems, and, by using a sparse GP approximation, scales well to large datasets. RIO can be applied directly to any pretrained NNs without modifications to model architecture or training pipeline.

Quantifying the predictive uncertainty of neural networks via residual estimation with I/O kernel

A process for training and sharing generic functional modules across multiple diverse (architecture, task) pairs for solving multiple diverse problems is described. The process is based on decomposing the general multi-task learning problem into several fine-grained and equally-sized subproblems, or pseudo-tasks. Training a set of (architecture, task) pairs then corresponds to solving a set of related pseudo-tasks, whose relationships can be exploited by shared functional modules. An efficient search algorithm is introduced for optimizing the mapping between pseudo-tasks and the modules that solve them, while simultaneously training the modules themselves.

System and method for implementing modular universal reparameterization for deep multi-task learning across diverse domains

A process for discovering optimal Generative Adversarial Networks (GAN) includes jointly optimizing the three functions of a GANs process including (i) a real component of a discriminator network's loss that is a function of D(x), wherein D(x) is the discriminator network's output for a real sample from an input dataset; (ii) a synthetic component of the discriminator network's loss that is a function of D(G(z)), wherein D(G(z)) is the discriminator network's output for a generator network's synthetic samples z from a latent distribution; and (iii) a generator network's loss which is a function of D(G(z)), with the discriminator network's total loss being the sum of components (i) and (ii). And separately optimizing each of the three functions in accordance with a process for evolving candidate loss functions having tree format by a genetic algorithm and optimizing selected candidate loss functions by optimizing set of coefficients of each of the one or more best candidate loss functions, wherein the set of coefficients is represented by a vector with dimensionality equal to the number of nodes in the best candidate loss function's tree.

Generative adversarial network optimization

Roughly described, an evolutionary data mining system includes at least two processing units, each having a pool of candidate individuals in which each candidate individual has a fitness estimate and experience level. A first processing unit tests candidate individuals against training data, updates an individual's experience level, and assigns each candidate to one of multiple layers of the candidate pool based on the individual's experience level. Individuals within the same layer of the same pool compete with each other to remain candidates. The first processing unit selects a set of candidates to retain based on the relative novelty of their responses to the training data. The first processing unit reports successful individuals to the second processing unit, and receives individuals for further testing from the second processing unit. The second processing unit selects individuals to retain based on their fitness estimate.

Growing community of inventors

Stanford, CA, United States of America

Risto Miikkulainen