Rupesh Kumar Srivastava

From captions to visual concepts and back

This paper presents a novel approach for automatically generating image descriptions: visual detectors, language models, and multimodal similarity models learnt directly from a dataset of image captions. We use multiple instance learning to train visual detectors for words that commonly occur in captions, including many different parts of speech such as nouns, verbs, and adjectives. The word detector outputs serve as conditional inputs to a maximum-entropy language model. The language model learns from a set of over 400,000 image descriptions to capture the statistics of word usage. We capture global semantics by re-ranking caption candidates using sentence-level features and a deep multimodal similarity model. Our system is state-of-the-art on the official Microsoft COCO benchmark, producing a BLEU-4 score of 29.1%. When human judges compare the system captions to ones written by other people on our held-out test set, the system captions have equal or better quality 34% of the time.

2013 Special Issue: First experiments with PowerPlay

Like a scientist or a playing child, POWERPLAY (Schmidhuber, 2011) not only learns new skills to solve given problems, but also invents new interesting problems by itself. By design, it continually comes up with the fastest to find, initially novel, but eventually solvable tasks. It also continually simplifies or compresses or speeds up solutions to previous tasks. Here we describe first experiments with POWERPLAY. A self-delimiting recurrent neural network SLIM RNN (Schmidhuber, 2012) is used as a general computational problem solving architecture. Its connection weights can encode arbitrary, self-delimiting, halting or non-halting programs affecting both environment (through effectors) and internal states encoding abstractions of event sequences. Our POWERPLAY-driven SLIM RNN learns to become an increasingly general solver of self-invented problems, continually adding new problem solving procedures to its growing skill repertoire. Extending a recent conference paper (Srivastava, Steunebrink, Stollenga, & Schmidhuber, 2012), we identify interesting, emerging, developmental stages of our open-ended system. We also show how it automatically self-modularizes, frequently re-using code for previously invented skills, always trying to invent novel tasks that can be quickly validated because they do not require too many weight changes affecting too many previous tasks.

An Evolutionary Algorithm Based Approach to Design Optimization Using Evidence Theory

For problems involving uncertainties in design variables and parameters, a bi-objective evolutionary algorithm (EA) based approach to design optimization using evidence theory is proposed and implemented in this paper. In addition to a functional objective, a plausibility measure of failure of constraint satisfaction is minimized. Despite some interests in classical optimization literature, this is the first attempt to use evidence theory with an EA. Due to EA’s flexibility in its operators, non-requirement of any gradient, its ability to handle multiple conflicting objectives, and ease of parallelization, evidence-based design optimization using an EA is promising. Results on a test problem and a couple of engineering design problems show that the modified evolutionary multi-objective optimization (EMO) algorithm is capable of finding a widely distributed trade-off frontier showing different optimal solutions corresponding to different levels of plausibility failure limits. Furthermore, a single-objective evidence based EA is found to produce better optimal solutions than a previously reported classical optimization procedure. The use of a GPU based parallel computing platform demonstrates EA’s performance enhancement around 160 to 700 times in implementing plausibility computations. Handling uncertainties of different types are getting increasingly popular in applied optimization studies and this EA based study should motivate further studies in handling uncertainties.

An evolutionary based Bayesian design optimization approach under incomplete information

Design optimization in the absence of complete information about uncertain quantities has been recently gaining consideration, as expensive repetitive computation tasks are becoming tractable due to the invention of faster and parallel computers. This work uses Bayesian inference to quantify design reliability when only sample measurements of the uncertain quantities are available. A generalized Bayesian reliability based design optimization algorithm has been proposed and implemented for numerical as well as engineering design problems. The approach uses an evolutionary algorithm (EA) to obtain a trade-off front between design objectives and reliability. The Bayesian approach provides a well-defined link between the amount of available information and the reliability through a confidence measure, and the EA acts as an efficient optimizer for a discrete and multi-dimensional objective space. Additionally, a GPU-based parallelization study shows computational speed-up of close to 100 times in a simulated scenario wherein the constraint qualification checks may be time consuming and could render a sequential implementation that can be impractical for large sample sets. These results show promise for the use of a parallel implementation of EAs in handling design optimization problems under uncertainties.

Bayesian reliability analysis under incomplete information using evolutionary algorithms

During engineering design, it is often difficult to quantify product reliability because of insufficient data or information for modeling the uncertainties. In such cases, one needs a reliability estimate when the functional form of the uncertainty in the design variables or parameters cannot be found. In this work, a probabilistic method to estimate the reliability in such cases is implemented using Non-Dominated Sorting Genetic Algorithm-II. The method is then coupled with an existing RBDO method to solve a problem with both epistemic and aleatory uncertainties.

Continually adding self-invented problems to the repertoire: First experiments with POWERPLAY

Pure scientists do not only invent new methods to solve given problems. They also invent new problems. The recent POWERPLAY framework formalizes this type of curiosity and creativity in a new, general, yet practical way. To acquire problem solving prowess through playing, POWERPLAY-based artificial explorers by design continually come up with the fastest to find, initially novel, but eventually solvable problems. They also continually simplify or speed up solutions to previous problems. We report on results of first experiments with POWERPLAY. A self-delimiting recurrent neural network (SLIM RNN) is used as a general computational architecture to implement the systems solver. Its weights can encode arbitrary, self-delimiting, halting or non-halting programs affecting both environment (through effectors) and internal states encoding abstractions of event sequences. In open-ended fashion, our POWERPLAY-driven RNNs learn to become increasingly general problem solvers, continually adding new problem solving procedures to the growing repertoire, exhibiting interesting developmental stages.

An EA-based approach to design optimization using evidence theory

For problems involving uncertainties in design variables and parameters, a bi-objective evolutionary algorithm (EA) based approach to design optimization using evidence theory is proposed and implemented in this paper. In addition to a functional objective, a plausibility measure of failure of constraint satisfaction is minimized. Despite some interests in classical optimization literature, such a consideration in EA is rare. Due to EAs flexibility in its operators, non-requirement of any gradient, its ability to handle multiple conflicting objectives, and ease of parallelization, evidence-based design optimization using an EA is promising. Results on a test problem and a couple of engineering design problems show that the modified evolutionary multi-objective optimization (EMO) algorithm is capable of finding a widely distributed trade-off frontier showing different optimal solutions corresponding to different levels of plausibility failure limits. Furthermore, a single-objective evidence based EA is found to produce better optimal solutions than a previously reported classical optimization procedure. Handling uncertainties of different types are getting increasingly popular in applied optimization studies and more such studies using EAs will make EAs more useful and pragmatic in practical optimization problem-solving tasks.