Avatar

Sadman Sakib

Graduate Student in Computer Sciences

University of Wisconsin-Madison

Hi! I am Sadman. I am completing my MS in Computer Science in Fall 2024 at the University of Wisconsin-Madison. I have a strong interest in developing deep learning methods to address modern challenges. My primary focus is on creating learning techniques for visual data (e.g., images, videos, 3D scenes, and synthesis), text data (e.g., corpora, tabular data, graphs, and closed-domain) and their various applications such as in healthcare, industrial automation, and intelligent assistants. In my recent projects, I have worked on multi-modal models, text analysis and generation, image analysis and generation, and state-space models.

During my MS, I also learned about Computer Systems, Networking, HPC and Data Systems. I interned at HPE in Summer 2023 and at Meta in Summer 2024, working on AI/HPC systems. Before coming to UW-Madison, I obtained my undergraduate major in Computer Science from Bangladesh University of Engineering and Technology and my undergraduate thesis was on Interpretable CNN .

In my free time, I can be found taking exercise, traveling places, reading non-fiction, cooking dishes and listening to music.

Interests

  • Vision and Language Models
  • Information Science

Education

  • Masters of Science in Computer Sciences (2022 - 2024)

    University of Wisconsin-Madison

  • Bachelor of Science in Computer Science (2017 - 2022)

    Bangladesh University of Engineering and Technology




Projects

 
 
 
 
 

Mitigating Hallucination in Vision-Language Models

Course Project

Oct 2024 - Dec 2024

Fine-tuned a vision-language model, BLIP2, with SFT and DPO-training to mitigate cross-modal hallucination. Evaluated BLIP2-base, BLIP2-tuned, and LLaVa-OneVision on POPE and MMHal-Bench benchmarks [Code]

 
 
 
 
 

Text Analysis and Generation with Language Models

Course Project

Sep 2024 - Dec 2024

Implemented LM components from scratch including Transformer, RNN, Beam Search, DPR, and Tree-of-Thought. Trained and evaluated the models on sentiment analysis, autoregressive generation, retrieval and reasoning. [Code]

 
 
 
 
 

Image Analysis and Generation with Vision Models

Course Project

Sep 2024 - Dec 2024

Implemented vision models from scratch including ViT, CLIP, FCOS detection, and denoising diffusion. Trained and evaluated models on MiniPlaces(78% top5), PASCAL VOC(0.61 mAP), Flicker-8k and AFHQ datasets [Code]

 
 
 
 
 

Unsupervised Domain Adaptation for Semantic Segmentation

Course Project

Oct 2024 - Dec 2024

Implemented and improved upon a test-time normalization, AdaBN, for unsupervised domain adaptation. Trained and evaluated a U-Net for semantic segmentation on Foggy Cityscapes (0.69 DICE, 12% inc. w/ TTN). [Code]

 
 
 
 
 

Hybrid Model Architecture for Language Tasks

Course Project

Oct 2024 - Dec 2024

Developed automated parallel-block hybrid architecture search combining Transformer and Mamba model blocks. Performed auto-regressive training and fine-tuning on MAD synthetic tasks, ELI5 and SQUAD datasets. [Code]

 
 
 
 
 

Memory-efficient Low-Rank Training for Large Foundation Models

Course Project

Mar 2024 - May 2024

Integrated Parallel LoRA with GPT-2, achieving equivalent performance with 24% less parameters. Combined LoRA with a Vision Transformer(ViT) and achieved 6x speedup in convergence of test loss [Code]

 
 
 
 
 

Explaining Grokking of Overparameterized Models

Course Project

Mar 2024 - May 2024

Demonstrated loss function and weight norm behavior in different grokking (delayed generalization) scenarios. Established that training data size and batch size are key factors for determining grokking delays [Code]




Courseworks

Course Title Instructor Year Keywords Projects
Computer Vision Yin Li 2024 CNN, Image Classification, Object Detection, Semantic Segmentation, Image Generation, Human Pose Detection, Video Understanding, 3D Scene Understanding, Medical Image Unsupervised Domain Adaptation for Semantic Segmentation
Natural Language Processing Junjie Hu 2024 Language Modeling, RNN, Transformers, Pre-training, Sequence Labeling, Document Modeling, RAG, Knowledge Graph, Dialogue System Hybrid Model Architecture for Language Tasks
Foundation Models Fred Sala 2024 Transformers, SSM, In-context Learning, Chain-of-Thought, Specialization, Alignment, Data and Benchmarks, Multimodal, Scaling Laws Mitigating Hallucination in Vision-Language Models
Advanced Big Data Systems Shivaram Venkataraman 2024 Big Data Stacks, Scheduling, Resource Management, ML Systems, Batch and Stream Analytics, Graph Processing Memory Efficient Low-Rank Systems for Large Foundation Models
Theoretical Foundations of Large-Scale ML Dimitris Papailiopoulos 2024 Generalization, Stochastic Methods, Transformers, System Tradeoffs, Distributed Optimization, Federated Learning, Model Compression Grokking - What/When/Why?
Cloud-native Database Xiangyao Yu 2023 Storage Disaggregation, Analytical and Transaction Processing, Serverless, Auto-Scaling, GPU DB, Memory Disaggregation, RDMA, NIC-assisted DB Evaluation of Wasm for Computation Pushdown in Cloud Database
Introduction to Big Data Systems Tyler Caraza-Harter 2023 Distributed Filesystem, Distributed Database, Distributed ML, Partitioning, Fault Tolerance, Availability, Streaming Distributed Data Processing with Hadoop, Spark and Cassandra
High-Performance Computing Dan Negrut 2023 GPU Architecture, GPU Memory, GPU Scheduling, ILP, Compiler Optimization, CUDA, Multi-core Programming, Supercomputer Parallel Computing with CUDA, OpenMP and MPI
SmartNIC Systems Ming Liu 2023 FPGA NIC, SoC NIC, Application Offloading, Caching, Job Scheduling, Load Balancing Flexible Block Storage Offload for Datacenters
Advanced Computer Networks Ming Liu 2023 Datacenter Architecture, Flow Scheduling, Routing, Load Balancing, Congestion Control, SDN, Endhost Network DPDK Evaluation, Datacenter Congestion Control, Disaggregated Storage Profiling
Operating Systems Remzi Arpaci-Dusseau 2022 Process Management, Process Scheduling, Virtual Memory, Threads, Locking, File Systems, Logging, Storage XV6, Parallel Sort, Distributed File System
Introduction to Computer Architecture Swamit Tannu 2022 Processor, Pipelining, Superscalar, MIPS ISA, SIMD, Cache, Memory Hierarchy 5-stage Pipelined Processor for MIPS-like ISA



Experience

Production Engineering Intern

Meta
May, 2024 - Aug, 2024. Menlo Park, CA, USA.

Worked on InfiniBand fabric for AI/HPC clusters

Platform Engineering Intern

Hewlett Packard Enterprise
May, 2023 - Aug, 2023. Bloomington, MN, USA.

Contributed to Linux kernel drivers for HPE Slingshot networking devices. Designed and implemented kernel-module API across multiple drivers.

Graduate Researcher

University of Wisconsin-Madison
Jan, 2024 - Sep, 2024. Madison, WI, USA.

Implemented a distributed hash join algorithm with linear-probing indexing utilizing remote shared memory. Evaluated the implementation on TPC-H benchmark and achieved performance close to Arrow’s native hashjoin.

Undergraduate Researcher

Bangladesh University of Engineering and Technology
Mar, 2021 - May, 2022. Dhaka, Bangladesh.

Designed an array of interpretable CNN models to analyze sequence motifs of DNA bendability. Developed efficient algorithms for processing large-scale data.

Lead Developer

University of Illinois at Urbana-Champaign
Jan, 2020 - Feb, 2022. Remote.

Collaborated with Prof. Matthew Caesar and UIUC students to develop a virtual IoT platform for deployment and monitoring of IoT network. Participated in both development and leadership roles.

Undergraduate Researcher

Feb, 2020 - Oct, 2021. Dhaka, Bangladesh.

Designed network architecture for authentication and application mobility solutions in federated edge computing. Developed 3GPP-compliant mobility and roaming protocols for 4G/5G networks.




Papers

Samin Rahman Khan, Sadman Sakib, M. Sohel Rahman, Md. Abul Hassan Samee DeepBend: An Interpretable Model of DNA Bendability (2023). In Cell iScience.
Asad Ali, Tushin Mallick, Sadman Sakib, Md. Shohrab Hossain and Ying-Dar Lin Provisioning Fog Services to 3GPP Subscribers: Authentication and Application Mobility (2022). In IEEE International Conference on Communications 2022.
Asad Ali, Samin Rahman Khan, Sadman Sakib, Md. Shohrab Hossain and Ying-Dar Lin Federated 3GPP Mobile Edge Computing Systems: A Transparent Proxy for 3rd-Party Authentication with Application Mobility Support (2022). In IEEE Access.



Awards

  • Head TA, Jan 2024 - Dec 2024
  • Dean's List Scholarship (BUET), 2018-2022
  • University Merit Scholarship (BUET), 2017-2022
  • Winner in National Undergraduate Mathematics Olympiad, 2018
  • Honorable Mention in Notre Dame College, 2016
  • National Winner in Bangladesh Mathematical Olympiad, 2014-16



Contact