Gür, Müslüm; Babacan, Yunus

In this study, a photoreceptor circuit composed solely of MOS (metal-oxide-semiconductor) transistors was proposed. Currently, researchers in neuromorphic engineering are focusing on mimicking biological nervous systems using silicon devices. The challenges involved in mimicking biological systems are as significant as their potential. These challenges include circuit complexity, low power consumption, computational difficulty, and the challenge of integrating passive components onto a chip. Biologically, millions of neurons consume only a few watts of energy to process information. Therefore, the main reasons challenging researchers in silicon circuit design are reducing power consumption and making it entirely composed of transistors for applicability within the chip. In this study, the modified photoreceptor circuit holds promise due to its low power consumption and entirely transistor-based design. The circuit design was created in LTspice. Permission for real production in TSMC 90-nm technology was granted for the simulation results. The metal-oxide-semiconductor field-effect transistors used in the design were operated below their threshold. The modified circuit is approximately equivalent to the potential change in a biological photoreceptor cell. The modified circuit has been shown to consume approximately 23 times less power compared to the reference circuits. Furthermore, the modified circuit was analyzed under different background light conditions and was found to operate stably. The proposed circuit, composed entirely of MOS transistors, enhances its compatibility with chip designs. This feature reduces its size, which in turn helps lower costs. Content of this journal is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

Gelen, Aykut Görkem

Convolutional neural networks with different architectures extract different features. Imagine a scenario where multiple models were trained on the same dataset, with the same hyperparameters, and under the same conditions. Because of their architecture, these models will learn to extract various features. As a result, even when all conditions are identical, different architectures will achieve varying levels of test accuracy. Test accuracy can be improved even more through ensembling models with varying levels of accuracy in different ways. The purpose of this research is to examine the encoder layers of models trained in such a scenario and to compare the effects of various architectures. In the study, the CIFAR-100 dataset was used to train five models based on ResNet18, GoogleNet, and MobileNetv3L. These models were ensembled in a multitude of ways. Class activation maps were computed for both analysis and ensembling. The research demonstrated that class activation maps could be used to learn how to ensemble models. As a result, model ensembling using class activation maps significantly improved test accuracy. Furthermore, one of the most important findings from the study is that including a model with low test accuracy in the ensemble can boost overall success.

Akar, Funda

Advancements in artificial intelligence (AI)-driven text-to-speech (TTS) technology have enabled transformative applications in accessibility and personalized learning. This study focuses on the development of a customizable Turkish TTS system aimed at enhancing lecture notes through personalized narration. Despite significant progress in widely spoken languages, the study addresses the lack of open-source Turkish TTS solutions by leveraging Narakeet TTS and fine-tuning the RVCv2 model. The methodology involves processing user-recorded voice data, segmenting audio, and training a voice conversion model to align synthesized TTS output with the speaker’s unique vocal characteristics. Metrics such as Mel-Frequency Cepstral Coefficients (MFCC), Dynamic Time Warping (DTW), Mel Cepstral Distortion (MCD), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligibility (STOI) were employed to evaluate system performance. In order to determine the similarity of the produced voice to the user’s voice, the first sound and the last sound were compared with the Emphasized Channel Attention, Propagation, and Aggregation Time Delay Neural Network (ECAPA-TDNN). Results demonstrate that the personalized model achieves high intelligibility and similarity to the original voice, with MCD values 0, PESQ scores 4.5, STOI scores 1, and ECAPA-TDNN similarity 0.62. This study underscores the potential of open-source TTS solutions in supporting less-resourced languages and highlights the importance of personalization in educational AI tools.

Orman, Kamil

Memristors are electronic devices, which are widely used in neuromorphic circuit design, because they exhibit some superior properties such as memory and nonlinear behavior. Previous nonvolatile memristor-based associative learning, which is a type of learning mechanism using circuits, has not taken into account the time interval between neutral stimulus and its effect on the circuit, after a long time-lapse. In this paper, a volatile memristor is used instead of a fully nonvolatile memristor in an effort to provide the sensitivity required for the time interval between applied pulses. Also, compared with previous studies, the response of the volatile memristor-based circuit is sensitive to additional input applied pulses. All simulations presented in this work were successfully carried out.

Babacan, Yunus

Previous studies have focused on the new mem-elements after the experimental realization of memristor which is the the fourth passive circuit element. This research presents a simple operational transconductance amplifier (OTA)-based memcapacitor and meminductor emulator circuit that does not comprise a mutator circuit. Each emulator was verified using the LTspice simulation program, and the simulation results were found to be compatible with the characteristic behaviors of the memcapacitor and meminductor.

Babacan, Yunus; Kacar, Firat

In this paper, new fully floating memristor emulator circuit that consumes ultra-low energy is presented. This proposed circuit is simple because of the fact that it doesn't contain any multiplication or various block circuits to obtain nonlinear characteristics of memristor. Transistors are operated in subthreshold region to obtain the non-linear behavior. Floating Current Source (FCS) is used in memristor emulator design to decrease the energy consumption of emulator. Associative learning is a type of the learning mechanisms and designed memristor are used in classical associative learning circuit successfully. All results are compatible with both memristor characteristics and learning mechanisms of circuit.