Skip to content
View nsdmlk's full-sized avatar
⚜️
Cooked
⚜️
Cooked

Highlights

  • Pro

Block or report nsdmlk

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
nsdmlk/README.md
Kanterville Ghost

☾   MACHINE LEARNING   ·   RESEARCH   ·   PYTHON   ☾





between mathematics & machines

CS student at Beijing Institute of Technology. I research how models learn when data is scarce, focusing on small-sample learning, conformal prediction, and reinforcement learning.

I build open-source libraries that bridge the gap between theoretical guarantees and practical implementations — from robust gradient boosting to calibrated neural networks.




Selected Works


Ⅰ. Small Data Machine Learning (Paper) Fundamental Challenges and Targeted Solutions

The theoretical foundation of Small Data. Formalizes the "Trilemma of Small Data" (variance, bias, and uncertainty under validation depletion) and evaluates the exact scaling boundaries where standard models fail.


Ⅱ. SmallGBM Robust gradient boosting for tabular data (n < 1000)

Replaces fragile mean-based second-order Taylor approximations with robust median leaf estimation. Built with a native C core (256-bin histograms) for 2-3x faster training.

  • Outperforms XGBoost and LightGBM on 22 small datasets
  • Limits AUC degradation to 3.7x less than XGBoost under 20% label corruption

Downloads Best of BITs


Ⅲ. SmallMLP Adaptive neural networks with calibrated conformal intervals (n < 500)

Eliminates manual architecture search by deriving layer width and dropout directly from dataset geometry. Produces prediction intervals with finite-sample coverage guarantees.

  • Best mean rank among 10 models on 18 datasets (zero-tuning)
  • Produces 19% narrower intervals than split conformal at equal coverage

Downloads Best of BITs


Ⅳ. RL-from-scratch Value-based reinforcement learning

Implementing algorithms from first principles to understand the mechanics behind agents, rather than treating them as black boxes.

DQN | Double DQN | Dueling DQN




Arsenal


Tech Stack



Footer
☾   BEIJING · CHINA · 2026   ☾

Pinned Loading

  1. SmallGBM SmallGBM Public

    SmallGBM is a gradient boosting library designed for small datasets (n < 1000). It combines robust leaf weight estimation with stochastic split selection to outperform XGBoost and LightGBM — with l…

    Python 16

  2. SmallMLP SmallMLP Public

    SmallMLP is a library for small, nonlinear datasets ( n < 500 ). It replaces manual architecture search with adaptive hyperparameters derived from dataset parameters ( n , d , K ), and produces cal…

    Python 6

  3. sRFTM sRFTM Public

    A small transformer that converts informal math expressions into LaTeX.

    Jupyter Notebook 1

  4. RL-from-scratch RL-from-scratch Public

    Reinforcement learning algorithms implemented from scratch in PyTorch. No Stable-Baselines3, no RLlib — every algorithm written line by line.

    Python 1

  5. LLM-Core LLM-Core Public

    LLM core built from scratch. Transformer from zero, agent mode, tools, memory. Learning by building.

    Python 1

  6. SmallData_paper SmallData_paper Public

    Small Data Machine Learning: Fundamental Challenges and Targeted Solutions

    Jupyter Notebook 2