Ahmad Rezaei

I am currently proceeding with my master's studies in RCSE major at Technische Universität Ilmenau. Concurrently I am working on a full-time research associate position on Explainable AI in the Automatic Optical Inspection domain.

Both my work and studies are ending in a few months (mentioned in my CV), and as a next major step, I am motivated to further extend my career in the AI and digital design domains.

I have been advised during my research by Dr.-Ing. Detlef Streitferdt and Prof. Ali Mahani.

Email  /  CV  /  Bio  /  Linkedin  /  Research Gate  /  Github

I am a young researcher with a HUGE interest in Explainable AI, Digital Design, Bioinformatics I am seeking experience and appropriate education in the field, and I desire to proceed in my career as a full-fledged researcher in the future. I am motivated to use these techniques and also combine technical knowledge with interpersonal skills to address challenges.

Utilization of explanations in a loop for training CNN models seems to increase the relevancy and localization of pertinent features in each decision that the model makes. We are currently studying methods to design loss functions and develop the right training approach for this purpose.

This software is built and tested on Windows and Linux Ubuntu 22LTS platforms. It enables the users to set various image settings and see the explanation after letting the Deep Learning model run on them.

Study on the selection and implementation of easily graspable explanation methods for end-users (operators at PCB production lines). The result was the implementation of local explanation methods with an approximate global performance metric for evaluation of model's validity.

Enhancing energy consumption and more lightweight implementation are our main focus. So far we have benefited both from software and hardware techniques in our project such as quantization, pruning, pipelining, and fixed-point precision system.

Discussing the research gaps in evaluation of explainable method and model pairs and developing a framework for evaluation of explainability in a quantitative time-friendly manner.

Paper: IEEE 47th Annual Computers, Software, and Applications Conference (COMPSAC), Torino, Italy
Paper DOI: https://doi.org/10.1109/COMPSAC57700.2023.00019

Proposal of an approximate performance metric for the automation of global evaluation on explainability. This work also studies the dependency of explainability on the datasets used during the training.

Paper: Modelling and Development of Intelligent Systems: 8th International Conference, MDIS 2022, Sibiu, Romania
Paper DOI: https://doi.org/10.1007/978-3-031-27034-5_4

A proposal for the use case of DNA base callers in active learning environments, which reaches the common prediction accuracy with less training data.

Paper: ArXiv (Submitted - under review)
Paper DOI: https://doi.org/10.48550/arXiv.2303.08915

A cutting edge logic locking method is propused to impel both the algorithmic and structure-based attacks.

Paper: IET Computers & Digital Techniques
Journal Paper DOI: 10.1049/cdt2.12022

In the first stage, the prominent You Only Look Once (YOLO) network of version 5 is used for pedestrian detection. The second stage benefits from the Dlib python library to extract 68 important facial points, and the final stage implements the sensor integration equation on the trained neural network to improve the crosswalk prediction accuracy.

CAN protocol is implemented on two Arduino modules to communicate with eachother, the master Arduino issues communication signals to receive the values for a light sensor from slave Arduino.

Initially the extent of data is reduced with changing the data types to less percise ones. Next and after developing a time-series RNN network,the relevant enegry consumption features are selected as inputs to this network. After training the model, we could reach a prediction state for the energy consumption in the production facility with only 7% error on average.

The research project involved an study on the state of machine learning in the Industry 4.0 with a focus on optical defect detection systems. Modification of YOLOv5 model by removing unused anchors for tiny objects leads to better performance as follows; The model is 14% and 60% faster performance in comparison to respectively YOLOv5 and state-of-the-art YOLOv4-MN2 model; The model is lighter than models in the literature, however it (with 10.59 Mbytes) does not reach minimum parameter count and is behind YOLOv4-MN2 (7.04 Mbytes); additionally, it sacrifises 0.92% mAP for the performance speed.

Abstract: Federated Learning and Distributed Learning were considered solutions to the problem of compromising private data in server-based Neural Network learning scenarios. However recent studies have shown, that these gradients can be inverted to reconstruct the input that produced them and investigated the influence of various attack- and hyperparameters on the quality of the reconstructions. In this paper we expand on this by analysing the influence of regularization techniques on the reconstructions and attack performance. We have found that Group Normalization and Batch Normalization have delayed gradient inversion attacks by a factor of 3 or more with variing degrees of influence on the training of the Neural Network. We also conclude, that high batch sizes seem to be benefitial for this cause. Finally, convolutional layers inside an architecture, that cannot be skipped have removed the ability to reconstruct details smaller than the kernel size, with only convolutional being completely resistent to all reconstruction attempts. We showed further, that the currently used metrics are not fit to assess the quality of reconstructions, especially depending on how much detail is needed to actually lose control of privacy on some data. However these results have been achieved on a particularly low resolution and easily classifiable dataset and have to be confirmed in higher dimensionalities and more cross-class variance.

After studying UK's governmental structure, we selected respective keywords an extracted the COVID related tweets (about 72,000) from various departments in the government. Next, we used data cleaning tecchnqiues and studied four different research questions about the relation of actual Covid status and the Tweets published by the government.

RBD of this network was implemented in SHARPE program and several factors were obtained

Various Schemes e.g. TMR, TMR simplex, and original circuit and also, apparent reliability were analysed based on exponential distribution

Obtaining fault coverage of the circuit with specified patterns and multiple fault centers.

Taking advantage of CORDIC algorithms to minimize the time consumed.

This computer was designed using Verilog VHDL language, after proceeding with synthesis procedure was implemented on Xillinx Spartan 6 FPGA using ISE design suite, and the outcome was scrutinized with ChipOscope.

Won first place by moving an object to the destination and simultaneously avoiding static forbidden zones and dynamic obstacles.

Baran Charity at SBUK

Baran Charity/ Sirjan, Iran

Hooshmand Internation Complex, Sirjan, Iran