Sharif Digital Repository

The approximation power of general feedforward neural networks with piecewise linear activation functions is investigated. First, lower bounds on the size of a network are established in terms of the approximation error and network depth and width. These bounds improve upon state- of-the-art bounds for certain classes of functions, such as strongly convex functions. Second, an upper bound is established on the difference of two neural networks with identical weights but different activation functions. © The Author(s) 2018  

In this article, two adaptive model predictive controllers (AMPC) are applied to regulate the blood glucose in type 1 diabetic patients. The first controller is constructed based on a linear model, while the second one is designed by using a nonlinear Hammerstein model. The adaptive version of these control schemes is considered to make them more robust against model mismatches and external disturbances. The least squares method with forgetting factor is used to update the model parameters. For simulation study, two well-known mathematical models namely, Puckett and Hovorka which describe the dynamical behavior of patient's body have been selected. The performances and robustness of the... 

This paper introduces the concept of mapping of Artificially Intelligent (AI) computational systems. The concept of homunculus from human neurophysiology is extended to AI systems. It is assumed that an AI system behaves similarly to a mini-column or ganglion in the natural animal brain that comprises a layer of afferent (input) neurons, a number of interconnecting processing cells, and a layer of efferent (output) neurons or organs. The objective of the present study was to identify the correlation between the stimulus to each afferent neuron and the corresponding response from each efferent organ when the intelligent system is subjected to certain stimuli. To clarify the general concept, a... 

This paper concerns the design of a neural state observer for nonlinear dynamic systems with noisy measurement channels and in the presence of small model errors. The proposed observer consists of three feedforward neural parts, two of which are MLP universal approximators, which are being trained off-line and the last one being a Linearly Parameterized Neural Network (LPNN), which is being updated on-line. The off-line trained parts are able to generate state estimations instantly and almost accurately, if there are not catastrophic errors in the mathematical model used. The contribution of the on-line adapting part is to compensate the remainder estimation error due to uncertain parameters... 

This study describes an adaptive sliding mode technique for attitude and position control of a rigid body insect-like flapping wing model in the presence of uncertainties. For this purpose, a six-degrees-of-freedom nonlinear and time-varying dynamic model of a typical hummingbird is considered for simulation studies. Based on the quasi-steady assumptions, three major aerodynamic loads including delayed stall, rotational lift and added mass are presented and analyzed, respectively. Using the averaging theory, a time-varying system is then transformed into the time-invariant system to design the adaptive controller. The controller is designed so that the closed-loop system will follow any... 

Blur identification is one important part of image restoration process. Linear motion blur is one of the most common degradation functions that corrupts images. Since 1976, many researchers tried to estimate motion blur parameters and this problem is solved in noise free images but in noisy images improvement can be done when image SNR is low. In this paper we have proposed a method to estimate motion blur parameters such as direction and length using Radon transform and Feed-Forward back propagation neural network for noisy images. To design the desired neural network, we used Weierstrass approximation theorem and Steifel reference Sets. The experimental results showed algorithm precision... 

In this paper, a wideband full-360o phase shifter with 6 bits of accuracy has been designed and simulated with minimal root mean square (RMS) phase error. The proposed phase shifter deployed a feed forward path including a frequency-to-voltage converter (FVC) to minimize the mismatch in quadrature generation to eventually reduce the RMS phase error for S-band (2-4GHz) applications. The designed phase shifter in 180nm CMOS technology achieves an RMS phase error in the range of 0.607-1.18? with -50dBm input signal over 2-4GHz frequency band. With lower input signal of -75dBm, the RMS phase error is 0.621-1.34? for 2-4GHz input frequency. The proposed phase shifter shows an RMS amplitude error...