Estimating the features to be extracted from an image for classification tasks are sometimes difficult, especially if images are related to a particular kind of noise. The aim of this paper is to propose a neural network architecture named Guided-Generative Network (GGN) to extract refined information that allows to correctly quantify the noise present in a sliding window of images. GNN tends to find the desired features to address such a problem in order to emit a detection criterion of this noise. The proposed GGN is applied on photorealistic images which are rendered by Monte-Carlo methods by evaluating a large number of samples per pixel. An insufficient number of samples per pixel tends to result in residual noise which is very noticeable to humans. This noise can be reduced by increasing the number of samples, as proven by Monte-Carlo theory, but this involves considerable computational time. Finding the right number of samples needed for human observers to perceive no noise is still an open problem. The results obtained show that GGN can correctly solve the problem without prior knowledge of the noise while being competitive with existing methods.

Choosing automatically the right algorithm using problem descriptors is a classical component of combinatorial optimization. It is also a good tool for making evolutionary algorithms fast, robust and versatile. We present Shiwa, an algorithm good at both discrete and continuous, noisy and noise-free, sequential and parallel, black-box optimization. Our algorithm is experimentally compared to competitors on YABBOB, a BBOB comparable testbed, and on some variants of it, and then validated on several real world testbeds.

Space-filling designs such as Low Discrepancy Sequence (LDS), Latin Hypercube Sampling (LHS) and Jittered Sampling (JS) were proposed for fully parallel hyperparameter search, and were shown to be more effective than random and grid search. We prove that LHS and JS outperform random search only by a constant factor. Consequently, we introduce a new sampling approach based on the reshaping of the search distribution, and we show both theoretically and numerically that it leads to significant gains over random search. Two methods are proposed for the reshaping: Recentering (when the distribution of the optimum is known), and Cauchy transformation (when the distribution of the optimum is unknown). The proposed methods are first validated on artificial experiments and simple real-world tests on clustering and Salmon mappings. Then we demonstrate that they drive performance improvement in a wide range of expensive artificial intelligence tasks, namely attend/infer/repeat, video next frame segmentation forecasting and progressive generative adversarial networks.

Multi-Modal Optimization problems are widespread and can be solved using numerous methods, such as niching, sharing or clearing. In this paper, we are interested in algorithms based on restart strategies, where the searching point is restarted at another initial position when an optimum is found. Previous works show that the choice of these initial positions greatly impacts the performance of the algorithm but is not easy to make. In this paper, we propose a new restart strategy, based on reinforcement learning. Our algorithm subdivides the search space and uses a Multi-Armed Bandit technique to choose the successive restart positions. We experiment this algorithm on various functions and on a modified Hump function with more complex local areas. Our results show significant improvements over previous algorithms, such as the Quasi-Random restart with Decreasing Step-size algorithm.

This paper is devoted to noisy optimization in case of a noise with standard deviation as large as variations of the fitness values, specifically when the variance does not decrease to zero around the optimum. We focus on comparing methods for choosing the number of resamplings. Experiments are performed on the differential evolution algorithm. By mathematical analysis, we design a new rule for choosing the number of resamplings for noisy optimization, as a function of the dimension, and validate its efficiency compared to existing heuristics.

We study mathematically and experimentally the convergence rate of differential evolution and particle swarm optimization for simple unimodal functions. Due to parallelization concerns, the focus is on lower bounds on the runtime, i.e. upper bounds on the speed-up, as a function of the population size. Two cases are particularly relevant: A population size of the same order of magnitude as the dimension and larger population sizes. We use the branching factor as a tool for proving bounds and get, as upper bounds, a linear speed-up for a population size similar to the dimension, and a logarithmic speed-up for larger population sizes. We then propose parametrizations for differential evolution and particle swarm optimization that reach these bounds.

International audience