Mahdi Beitollahi

This API creates short-form videos with AI from YouTube videos. It uses OpenAI to create 60 seconds videos from YouTube links with one click that can be shared on TikTok, Instagram, YouTube shorts, or more. The motivation behind this project was to familiarize myself with microservice architecture design and implementation. I learned how to create an scalable system by working with databases, messaging protocols, docker containers, Linux infrastructures, etc.

In this paper, we explore the composability of LoRA modules, examining if combining these pre-trained modules enhances generalization to unseen downstream tasks. Our experimental results on both vision and language models reveal that in few-shot settings, where only a limited number of samples are available for the downstream task, both uniform and learned composition methods result in better transfer accuracy; outperforming full fine-tuning and training a LoRA from scratch.

This paper introduces FedPFT (Federated Learning with Parametric Feature Transfer), a methodology that harnesses the transferability of foundation models to enhance both accuracy and communication efficiency in one-shot FL. The approach involves transferring per-client parametric models (specifically, Gaussian mixtures) of features extracted from foundation models. Subsequently, each parametric model is employed to generate synthetic features for training a classifier head.

Can diffusion models function as data augmentation tools to address domain generalization (DG) from a data-centric perspective, rather than relying on the loss functions? Our findings reveal that trivial cross domain data augmentation (CDGA) along with the vanilla ERM using readily available diffusion models outperforms state-of-the-art (SOTA) training algorithms. This paper delves into the exploration of why and how this generative augmentation can outperform complicated DG algorithms.

Centralized servers in Federated Learning (FL) present challenges including communication bottlenecks and susceptibility to a single point of failure. Additionally, user devices inherently exhibit model and data heterogeneity. To address these issues, we propose a Decentralized Federated Mutual Learning (DFML) framework that is serverless, supports nonrestrictive heterogeneous models, and avoids reliance on public data.

Recently, layer normalization (LN) has been shown to be surprisingly effective in federated learning (FL) with non-i.i.d. data. However, exactly why and how it works remains mysterious. To understand layer normalization better in FL, we identify the key contributing mechanism of normalization methods in FL, called feature normalization (FN), which applies normalization to the latent feature representation before the classifier head. Although LN and FN do not improve expressive power, they control feature collapse and local overfitting to heavily skewed datasets, and thus accelerates global training.

In this survey, we discuss each of the challenges of deploying federated learning over wireless networks and their respective state-of-the-art proposed solutions in an in-depth manner. By illustrating the tradeoff between each of the solutions, we discuss the underlying effect of the wireless network on the performance of FL.

In this paper, we introduce a novel Dynamic Sparsification for Federated Learning (DSFL) approach that enables users to compress their local models based on their communication capacity at each iteration by using two novel sparsification methods: layer-wise similarity sparsification (LSS) and extended top-K sparsification. LSS enables DSFL to utilize the global redundant information in users’ models by using the Centralized Kernel Alignment (CKA) similarity for sparsification.

In this paper, we propose the Federated Learning with Autoencoder Compression (FLAC) approach that utilizes the redundant information and error-correcting capability of Federated Learning (FL) to compress user devices’ models for uplink transmission. FLAC trains an autoencoder to encode and decode users’ models at the server in the Training State, and then, sends the autoencoder to user devices for compressing local models for future iterations during the Compression State. We theoretically prove that FLAC converges for FL systems with strongly convex ML models and non-i.i.d. data distribution.

This project aims to develop, analyze, and evaluate methods that can accurately estimate the crowd count from a single image-based and generate the density map of images. To this end, we propose novel ideas in three main parts of preprocessing, model architectures, and loss function of our deep learning pipeline. More specifically, we utilize transfer learning methods by using pre-trained depth and image models to develop depth-guided attention models and VGG-based U-Net architecture to address the limited number of samples in the dataset and increase the accuracy.

In this project, we utilized Spatio-temporal Representation Learning for EEG-based Brain-Computer Interfaces. We propose a novel method that can capture spatio-temporal representations of raw EEG in commercial settings using autoencoders. More specifically, we modify the machine learning (ML) pipeline by adding a feature learning preprocessing method that can capture cross-subject informationIn this project, we utilize Pytorch to build an end-to-end classification pipeline for Motor Imagery (MI) tasks using cross-subject data.