Arseni Ivanov
Doctoral Student
Deep Neural Network Acceleration
Information Theory
GPGPU
Compute Optimization
Cognition
Research
Research portal link
My research is focused on neural network acceleration for GPGPU. This means making neural networks execute efficiently by exploiting the hardware architecture of various graphics cards using interfaces such as Triton. I am interested in high-level optimizations that affect low-level runtime, such as reordering of compute, JIT compilation for GPU-info incorporation, or data reuse of execution between different kernels.
Publications
Quantization of Deep Neural Networks to facilitate self-correction of
weights on Phase Change Memory-based analog hardware
My paper on an alternative technique to handle weight drift on analog PCM hardware by binning the weights during training for the least amount of loss. Implemented in python using IBM's aihwkit.
Image classification for improving an IR-signal classification in a semi-supervised learning pipeline
My masters thesis paper for security survelliance company Verisure on using Image Classification to improve their labelling pipeline for PIR data. Project focus was extracting label and meta-data from images to aid in statistical analysis of signal diversity for PIR model training.
Selected Projects
Peer-to-Peer Augmented Reality Chess
Augmented reality chess written in C++ using OpenCV and Aruco library for plane estimation/camera motion, the frame is then ported as a texture into OpenGL for graphics rendering. The communication between the two players runs one each peers machine as a separate thread using a TCP network socket which decouples it from the graphics rendering.
Emotional Autoencoder
Emotional autoencoder prototype using open source data. Learn a shared embedding/vocabulary for human emotion in EEG signals, then train personal encoder-decoders for every person into this shared space. Can allow for emotional mapping, and at some point in the future with computer-to-brain-interfaces could allow emotional intelligence by default.
Forward-Forward Algorithm
Implementation of Hinton's backprop alternative from NeurIPS 2022 in Pytorch. Uses local updates to drive the network weights using contrastative learning.
About Me
My name is Arseni Ivanov, a 27-year-old Doctoral Student with an M.Sc in Computer Engineering from Lund University. My research focuses on GPGPU programming optimizations for neural networks, but I am also passionate about compute, machine learning, cognition and cognitive neuroscience.
Outside of my academic pursuits, I enjoy working on personal projects, such as applying my knowledge to practice and implementing small ML tools to help me in my daily life. My programming knowledge and industry experience have equipped me with the skills to plan and execute machine learning projects from start to finish. I am eager to continue expanding my expertise in cognitive topics surrounding machine learning such as neuroscience, information theory and cognition.
When I'm not working on projects, I enjoy cooking intricate meals, learning languages(currently French, Japanese, Portuguese) and dancing(currently Commercial Street, Performance).