Vol 27, No 2 (2018)

The current state-of-the-art in Deep Learning (DL) based artificial intelligence (AI) is reviewed. A special emphasis is made to compare the level of a concrete AI system with human abilities to show what remains to be done to achieve human level AI. Several estimates are proposed for comparison of the current “intellectual level” of AI systems with the human level. Among them is relation of Shannon’s estimate for lower bound on human word perplexity to recent progress in natural language AI modeling. Relations between the operation of DL constructions and principles of live neural information processing are discussed. The problem of AI risks and benefits is also reviewed based on arguments from both sides.

This study considers solving the inverse problem of determination of salt or ionic composition of multi-component solutions of inorganic salts by their Raman spectra using artificial neural networks. From the point of view of data analysis, one of the key problems here is high input dimensionality of the data, as the spectrum is usually recorded in 1–2 thousand channels. The two main approaches used for dimensionality reduction are feature selection and feature extraction. In this paper, three feature extraction methods are compared: channel aggregation, principal component analysis, and discrete wavelet transformation. It is demonstrated that for neural network solution of the inverse problem of determination of salt composition, the best results are provided by channel aggregation.

Neural networks based on metric recognition methods have a strictly determined architecture. Number of neurons, connections, as well as weights and thresholds values are calculated analytically, based on the initial conditions of tasks: number of recognizable classes, number of samples, metric expressions used. This paper discusses the possibility of transforming these networks in order to apply classical learning algorithms to them without using analytical expressions that calculate weight values. In the received network, training is carried out by recognizing images in pairs. This approach simplifies the learning process and easily allows to expand the neural network by adding new images to the recognition task. The advantages of these networks, including such as: (1) network architecture simplicity and transparency; (2) training simplicity and reliability; (3) the possibility of using a large number of images in the recognition problem using a neural network; (4) a consistent increase in the number of recognizable classes without changing the previous values of weights and thresholds.

We develop adaptive algorithms for interpolation of multidimensional signals based on a local autoregressive model of signals. Their adaptability is the result of calculation of the model’s parameters for the reading of each signal using estimates of a local autocorrelation function of the compressed signal. Under developing these interpolation algorithms, we do not include excess hierarchical grids of readings. That is the reason why our interpolators are well adapted for use in the framework of a compression method based on hierarchical grid interpolation. Computer simulations on real images confirm that the proposed interpolators allow us to increase a degree of hierarchical compression.