Md. Asif Uddin
Deep Learning Researcher — Vision, Medical Imaging, Vision-Language Models, Causal Inference, Bioinformatics & Gene Sequence Analysis
Correlation is cheap. Only an intervention costs anything.
— the author
Dhaka, Bangladesh
Prologue
I spend most of my time trying to understand AI and learn how to build useful AI systems. I work with different models, train and fine-tune them, modify existing architectures, and sometimes build new approaches when a problem needs it. I enjoy taking an idea, understanding how it works, and then trying to apply it to a real problem.
My undergraduate thesis, HierarchiRetina, is one example of this. It focuses on diabetic retinopathy grading from retinal photographs. The main idea is to make the model more interpretable by using both the original image and information about different retinal lesions. I have worked on the full process, from finding and preparing the data to training and evaluating the models. Getting the data was also a major part of the work. I visited hospitals, contacted researchers and institutions, and went through more than 37 published papers looking for suitable datasets. Most requests were declined, but the authors of FGADR responded and allowed us to use their dataset.
Alongside my thesis, I work with my advisors on different research problems. One project on open-world species recognition has already been completed and submitted for publication. Another project focuses on kidney disease. We are working with CT scan images to develop automated analysis and generate reports from the scans. These projects have helped me understand how AI can be adapted to very different problems and domains.
More recently, I have started working on a problem in computational biology. Instead of only asking which genes are related to Alzheimer’s disease, the work looks at what might happen to a microglial cell when a specific gene is knocked down. We then compare the model’s predictions with experimental CRISPR screens. I find this interesting because it brings together AI, biology, and experimental research.
Next is an MSc in Computer Science and Engineering at BRAC University, starting in Fall 2026. My proposal has already been accepted. I want to continue doing research while also building a career in AI and machine learning. I do not see research and industry as two separate paths. I want to keep learning through research while building AI systems that can solve real problems.
Outside computer science, I have broad interests in philosophy, law, public policy, science, and society. I also love cooking and trying new recipes. I enjoy learning about how ideas shape the way we understand the world, how technology affects people, and how we can use what we learn to build things that are genuinely useful.
Axioms
- HierarchiRetina, Stage IIIA number nobody can compare to their own is not a result.
- HierarchiRetina, Stage IThe test set is read once, and only after the threshold is locked.
- Cotton wool spots, v1When a model predicts nothing, the loss is usually right and the design is wrong.
- HierarchiRetina, evaluationName the limitation before the reviewer does.
- FGADR acquisitionAsk until someone answers. One reply is a dataset.
Instrumentarium
Fig. i
Languages
Python (primary). C (coursework level).
Everything below is written in the first of these.
Fig. ii
Deep learning
PyTorch, timm, segmentation_models_pytorch. Mixed-precision training with GradScaler tuning, EMA, test-time augmentation, k-fold CV with resumable per-fold checkpointing, class-imbalance handling (pos_weight, weighted samplers), threshold calibration on validation only, ordinal regression (CORN), mixture-of-experts, cross-attention fusion, attention gates, ASPP, GeM pooling, deep supervision, frozen-backbone transfer, Grad-CAM.
Training discipline more than library familiarity.
Fig. iii
Architectures used hands-on
ConvNeXt V2, Swin UNETR V2, DINOv2/v3, EfficientNet, U-Net, U-Net++, Attention U-Net, SwinHRUNetPP, HSMoE-AUNet, image-text encoders.
I build, train, modify and evaluate these rather than only reading about them.
Fig. iv
Medical imaging
Fundus photography, non-contrast CT, NIfTI and DICOM handling, FOV masking with morphological erosion, CLAHE on the green channel, retinal cropping, patch extraction and stitching, lesion-level connected-component evaluation, de-identification, radiologist report parsing.
Most of this is preprocessing nobody writes papers about.
Fig. v
Evaluation
QWK, ICC, AUC-ROC, AUC-PR, Dice, IoU, NMI, lesion-level and image-level recall, sensitivity/specificity tradeoff analysis, per-dataset breakdowns, held-out test discipline and leakage assertions.
The per-dataset breakdown is the part that matters.
Fig. vi
Causal inference and computational biology
Causal graphs and DAGs, d-separation, interventions, counterfactual reasoning, gene perturbation analysis, gene regulatory network inference, multi-omics integration and biomarker discovery.
Currently extending this direction toward single-cell foundation models, genomic sequence models and perturbation-based causal analysis.
Fig. vii
Data
pandas, NumPy, Albumentations, OpenCV, Matplotlib, Seaborn, scikit-learn, nbformat.
Ordinary tools, used daily.
Fig. viii
Backend
FastAPI, REST APIs, SQLAlchemy, JWT auth, Alembic, AWS S3, CloudFront.
From building CSMMS, in the appendix.
Fig. ix
MODEL FAMILIARITY
Beyond the architectures I have used hands-on, I regularly study modern vision, multimodal, language, medical imaging and computational biology models. This includes SigLIP 2, EVA-CLIP, SAM 3, nnU-Net, ViT, Qwen, Llama, DeepSeek, Kimi, MedGemma and genomic models such as AlphaGenome and Geneformer. I focus on understanding their architecture, training objective, scaling behavior, benchmarks, limitations and practical use rather than treating familiarity with a model as equivalent to hands-on experience. The model notes I keep are part of this ongoing study.
I study, compare and reproduce modern models to understand their architecture, training objectives, benchmarks, limitations and where they fit into my own research.
Fig. x
READING
I love reading books and research papers. My current reading is centered on machine learning, causality and computational biology. Key books include The Book of Why, Causality: Models, Reasoning, and Inference, Causal Inference in Statistics: A Primer, and Elements of Causal Inference: Foundations and Learning Algorithms. The reading is not separate from the research work. It is mainly about understanding what can be inferred from data, where the assumptions enter, and how interventions can provide stronger evidence than observation alone.
I read research papers and technical books to build the theoretical foundation behind the models and methods I use, especially in machine learning, causality and computational biology.
Fig. xi
Other
Git, Jupyter, LaTeX, lex/yacc, CUDA-level debugging. Trained on an NVIDIA RTX A6000.
One machine, a great deal of patience.
Not yet: SQL, propensity score methods, Mendelian randomization.
Compendium
HierarchiRetina
Interpretable diabetic retinopathy grading from retinal photographs
Rank Radii Transfer as Quantiles
Placing unseen species inside the taxonomy, not in flat clusters
Automated Renal Analysis
Automated Renal Measurement, Kidney Stone and Cyst Segmentation, and Radiology Report Generation from Thick-Slice Non-Contrast CT
HCGT-PG
What would happen if you knocked the gene down
Instrumenta propria
Fig. I
HSMoE-AUNet
A microaneurysm is a few dark pixels. A cotton wool spot is a soft pale patch. They do not respond to the same design.
Hierarchical sparse mixture-of-experts attention U-Net on a ConvNeXt-Small encoder, with a different expert set per lesion type. Custom experts: MultiScaleExpert, BrightSpotExpert, EdgeEnhancingExpert, SoftRegionExpert, TextureExpert.
Fig. II
SwinHRUNetPP
Vessel segmentation needs global context and hairline spatial precision at once, and most hybrids trade one away.
A Swin-based high-resolution U-Net++ hybrid, arrived at by rejection rather than by design.
Rejected along the way
- SwinV2-Large
- smp UNet++
- plain SwinUNet
- Lovász loss
- BoundaryLoss
- warm-restart scheduling
- label smoothing
- encoder freezing
Each of these made the result worse.
Fig. III
MultiScaleMANet
Microaneurysms vanish at low resolution and lose context at high resolution.
Three branches — full 512×512 through EfficientNet-B4, four 256×256 patches through a shared EfficientNet-B0, sixteen 128×128 patches through the same encoder — with spatial reassembly and bottleneck fusion. Alignment was proved on paper before implementation.
Fig. IV
RichAttentionUNet
A baseline worth trusting has to be one you can account for line by line.
Written from scratch with no pretrained weights: residual conv blocks with SE channel attention, an ASPP multi-scale bottleneck, attention gates in the decoder, deep supervision heads.
Fig. V
Custom losses
Off-the-shelf losses reward the wrong thing on sparse, bright, or boundary-defined lesions.
DualEdgeBoundaryLoss, OpticDiscSuppressionLoss, SoftBoundaryLoss, probability-weighted RegionCoherenceLoss, an FOV-masked composite loss, clDice.
Elementa
A course in deep learning, built to be read in order: transformers, vision models, and the vision-language systems built from them. Entries are propositions, not posts — each states a claim, names the propositions it depends on, and carries one figure. If the figure cannot be drawn, the proposition is not ready to be written.
- Book IFoundationstokenisation, attention, the transformer block
- Book IIVisionViT, DINOv2, self-supervised visual representation
- Book IIILarge Language Model (LLM)Transformers, attention mechanisms, prompt engineering, fine-tuning.
- Book IVVision-language (VLM)contrastive pretraining, fusion, VLM architectures
- Book VCausal inferenceDAGs, interventions, counterfactuals, perturbation
- Book VIBioinformaticssingle-cell data, gene regulatory networks, perturbation screens
- Book VIIPracticelosses, training dynamics, evaluation
Marginalia
Reviews of books and models, and notes that did not fit anywhere else.
Chronicle
Fall 2026 —
BRAC University, Dhaka
MSc, Computer Science and Engineering.
Thesis proposal accepted: HCGT-PG. Supervisor: Dr. Badhan Das.
Fall 2021 – Summer 2026
BRAC University, Dhaka
BSc, Computer Science and Engineering.
Thesis: HierarchiRetina, supervised by Rafeed Rahman, deposited in the BRAC University institutional repository. Coursework in machine learning, deep learning, compilers, theory of computation, DBMS, operating systems, algorithms and software engineering.
2020
Uttara High School and College, Dhaka
Higher Secondary Certificate.
Science.
2018
Ullapara Merchant Pilot High School, Sirajganj
Secondary School Certificate.
Science.
Courseworkmachine learning, deep learning, compilers, theory of computation, DBMS, operating systems, algorithms, software engineering.
Appendix
Engineering work, set smaller than the research chapters.
CSMMS — Campus Student Management and Marketplace System
FastAPI backend. Item CRUD with keyword, category, price-range and sort filtering plus pagination. Image upload to S3 with Pillow compression, delivered through CloudFront. Service priority levels and live-computed impact metrics — weighted student ratings, students served, completion rate. SQLAlchemy models, JWT auth, Alembic migrations.
Leveraging CNN and Random Forest for Accurate Food Expiry Prediction
An Artificial Intelligence-driven solution designed to predict the expiry dates of food items like fruits, vegetables, and packaged goods to minimize global food waste. The system uses Convolutional Neural Networks (CNNs) to analyze image data for freshness, combined with Random Forest and Multi-Layer Perceptron (MLP) models to process environmental metadata such as temperature, humidity, and storage duration. Built using the Kaggle dataset 'Fruits Fresh and Rotten for Classification' alongside self-generated features like "days since purchase" to improve generalization.
Compiler symbol table
lex and yacc/bison. Hash table using sum-of-characters mod bucket count, scope stack via parent pointers, and a scope_started flag with mid-rule actions so function parameters land inside function scope while if and for blocks get their own. Made portable across macOS and Linux.
Correspondence
To —
md.asif.uddin@g.bracu.ac.bdgithub.com/asifuddin01linkedin.com/in/md-asif-uddin01Dhaka, Bangladesh