Alireza Mansouri

Machine Learning & Optimization Researcher

I am a passionate Machine Learning and Software Engineer with a focus on optimization for machine learning and deep learning. My professional career includes designing scalable software solutions, developing efficient machine learning models, and contributing to open-source projects.

About Me

My research and professional background

I am Alireza Mansouri, a passionate Machine Learning and Software Engineer with a focus on optimization for machine learning and deep learning. My professional career includes designing scalable software solutions, developing efficient machine learning models, and contributing to open-source projects. I work at the intersection of research and application, and my targer is to create impactful solutions in AI.

My key interests include enhancing model performance through optimization and exploring hardware acceleration for neural networks. I'm particularly interested in the convergence of theoretical optimization algorithms and their practical applications in real-world machine learning systems.

Education

Academic background and qualifications

Sep 2019 - Feb 2024

Master of Science in Computer Science - Artificial Intelligence

Institute for Advanced Studies in Basic Sciences (IASBS)

Focused on optimization for deep learning models and neural network acceleration. Thesis explored conjugate gradient optimization algorithms for training deep neural networks to get faster convergence rates. It was a good try and I learned alot in this path.

Sep 2009 - Jan 2015

Bachelor of Science in Electronics

University of Guilan

Studied electronic systems design with a focus on digital systems and embedded programming. The project involved implementing a DC-DC Converter on FPGA platforms.

Research Interests

Areas of focus and expertise

Optimization for ML/DL

Software Architecture and Design

FPGA

Neural Network Acceleration

Distributed ML Systems

Optimization Algorithms

Projects

Research and development work

Optimization Algorithm for Deep Learning

Developed a hybrid optimization algorithm combining conjugate gradient methods with the Adam optimizer, trying to acheive faster convergence on benchmark datasets (MNIST, Cifar10).

def hybrid_optimizer(params, grads, state):
    # Combine conjugate gradient with Adam
    beta1, beta2 = 0.9, 0.999
    state['m'] = beta1 * state['m'] + (1-beta1) * grads
    state['v'] = beta2 * state['v'] + (1-beta2) * grads**2

    # Apply conjugate direction
    if state['t'] > 0:
        conj_dir = compute_conjugate_direction(grads, state)
        update = conj_dir * state['lr'] / (np.sqrt(state['v']) + 1e-8)
    else:
        update = state['m'] * state['lr'] / (np.sqrt(state['v']) + 1e-8)

    return params - update

GitHub Demo

Scalable ML Pipeline

Built a scalable machine learning pipeline using Docker and Kubernetes for efficient model training and deployment. The system automatically scales based on workload and optimizes resource utilization.

GitHub Live Demo

FPGA-based Neural Network Accelerator

Collaborated on an FPGA implementation to accelerate neural network inference, outperforming traditional CPU/GPU setups. The design focuses on optimizing matrix operations and reducing memory access latency.