Ph.D. Student · Electrical & Computer Engineering · UMD
Mahmut Semih Akkoç
I study the fundamental limits of what can be discarded,
in data compression and inside large language models.
Background
I’m a Ph.D. student in Electrical and Computer Engineering at the University of Maryland, College Park, advised by Prof. Sennur Ulukus.
My research has two sides. On the compression side, I derive rate–distortion characterizations for task-aware compression: how many bits you need when the reconstruction is consumed by a downstream model rather than a person, and how far a lossy codec can steer the statistics of what it outputs. On the interpretability side, I analyze transformers as tractable dynamical systems, treating tokens as interacting particles to find where the representational work inside a network actually happens.
Before Maryland, I received my B.Sc. in Electrical and Electronics Engineering from Bilkent University, with an exchange semester at the National University of Singapore. At Bilkent I worked on semantic signal processing with Prof. Orhan Arıkan and on computational electromagnetics with Prof. Vakur B. Ertürk. I’ve also interned at Huawei (semantic communications), ASELSAN (RF & microwave systems), and BİTES (software engineering).
Education
Ph.D. in Electrical and Computer Engineering, University of Maryland, College Park
College Park, MD · Sep 2025 – Present
GPA: 3.93/4.00 · Advisor: Prof. Sennur Ulukus
B.Sc. in Electrical and Electronics Engineering, Bilkent University
Ankara, Türkiye · Sep 2021 – May 2025
GPA: 3.92/4.00
Exchange Student in Electrical and Computer Engineering, National University of Singapore
Singapore · Jan 2024 – May 2024
Relevant coursework: Quantum Communication and Cryptography, Quantum Mechanics I, Probability
Current Focus
My research establishes the fundamental limits of what can be discarded: deriving rate–distortion characterizations that govern task-aware compression, and analyzing transformers as tractable dynamical systems to locate where representational work actually happens.
Task-aware compression
Rate–distortion limits for compression whose output is consumed by a downstream model, including classifiers trained on a different distribution than the source.
Transformers as dynamical systems
Treating tokens as interacting particles to analyze attention and feed-forward blocks, and to find where the compressive work inside a network happens.
Steering output statistics
How far a lossy codec can steer the statistics of what it outputs while staying faithful to the source, with uses in watermarking, output auditing, and privacy.
Papers & Preprints
- AcceptedThe Rate–Distortion–Deception Tradeoff
Introduces deception as a third axis of the rate–distortion–perception framework: a compressed reconstruction stays within a target fidelity of the source while provably resembling a sample from a different distribution. Relevant to generative-model watermarking, output auditing, and privacy-preserving compression.
- PublishedPractical Hardware Demonstration of a Multi-Sensor Goal-Oriented Semantic Signal Processing and Communications Network
† Equal contribution
A hardware realization of a goal-oriented semantic communication network, showing that low-power commodity edge devices can sustain real-time semantic extraction and multi-sensor fusion.
- Under ReviewOn the Diverse Dynamical Behaviors Arising in Deep Linear Transformers
Recasts transformer inference as an interacting-particle system and proves that, in embedding dimension two, the depth-wise dynamics reduce exactly to a generalized Kuramoto model, giving a closed-form surrogate for stacked linear self-attention and explaining token clustering, oscillation, and bifurcation.
- SubmittedTask-aware Compression with Distribution Matching
Derives the optimal compression rate when the reconstruction is consumed by a mismatched downstream classifier, and recasts the design as a rate–distortion–perception problem with a KL constraint matching the classifier’s training distribution.
- SubmittedThe Role of Feed-Forward Layers in Transformer Dynamics
Extends the interacting-particle analysis of transformers beyond attention to the feed-forward block, isolating how pointwise nonlinearities reshape token trajectories and which layers carry the compressive work of the network.
Research and Industry
Research Experience
Graduate Research Assistant
University of Maryland, College Park · Information Theory & Machine Learning
Developing information-theoretic and dynamical-systems tools for compressing and interpreting transformers, from task-aware rate allocation to mechanistic analysis of attention and feed-forward blocks.
Undergraduate Researcher
Bilkent University · Semantic Signal Processing (Prof. Orhan Arıkan) · Computational Electromagnetics (Prof. Vakur B. Ertürk)
Built the embedded multi-sensor demonstrator behind a journal paper on goal-oriented semantic signal processing, and implemented a method-of-moments solver for plane-wave scattering by a conducting strip.
Industry Experience
Summer Intern
ASELSAN · RF & Microwave Systems
Designed a suspended ring resonator model in MATLAB and CST Studio that estimates dielectric constant and loss tangent, a low-cost alternative to commercial material characterization.
R&D Intern
Huawei · Semantic Communications
Prototyped a two-node semantic communication network on Raspberry Pi that cut transmission from a 10 Mbps video stream to about 10 bps of task-relevant tokens, and improved multi-sensor integration accuracy by 67.2%.
Summer Intern
BİTES Defence & Aerospace Technologies · Software Engineering
Built a multithreaded, socket-based file transfer system in C++, using factory and observer patterns to keep socket creation and worker notification decoupled.
Selected Projects
Li-Fi Protocol on an FPGA ↗
A complete Li-Fi communication stack in VHDL on a Xilinx Artix-7 (BASYS3): modular encoding and decoding of audio and text, LED transmission and LDR reception, with payloads in on-chip ROM.
Semantic Communication Prototype
A two-node network on Raspberry Pi that extracts visual semantics on-device, cutting transmission from a 10 Mbps video stream to about 10 bps of task-relevant tokens.
Dielectric Constant Prediction ↗
A suspended ring resonator model in MATLAB and CST Studio that estimates dielectric constant and loss tangent, generalized with lookup tables and polynomial fits so it transfers across materials.
Multithreaded File Transfer ↗
A socket-based file transfer system in C++ that handles arbitrary file types through binary I/O, using factory and observer patterns to keep sockets and workers decoupled.
Honors & Awards
Academic Excellence Award
Bilkent University, Dept. of Electrical and Electronics Engineering
Awarded to top-ranked students for sustained academic distinction across the degree program.
Research Excellence Award
Bilkent University, Dept. of Electrical and Electronics Engineering
Recognized for research contributions culminating in a first-authored journal publication as an undergraduate.
Best Project Award
Bilkent University, Dept. of Physics
Selected department-wide for excellence and originality in project design and execution.
Recent Updates
- Our paper The Rate–Distortion–Deception Tradeoff was accepted to the 62nd Allerton Conference on Communication, Control, and Computing.
- New preprint: On the Diverse Dynamical Behaviors Arising in Deep Linear Transformers, which shows that in two dimensions, stacked linear self-attention reduces exactly to a generalized Kuramoto model. Under review at Foundations of Computational Mathematics.
- Started my Ph.D. in Electrical and Computer Engineering at the University of Maryland, College Park, advised by Prof. Sennur Ulukus.
- Graduated from Bilkent University with a B.Sc. in Electrical and Electronics Engineering, and received the department’s Academic Excellence and Research Excellence Awards.
Get in Touch
Research and internships
I'm happy to talk about information theory, compression, and interpretability, and I'm currently looking for a Summer 2027 research internship in efficient AI, compression, or LLM interpretability. Email is the best way to reach me.