Nikolai Zhukov

Towards Understanding Geometry of Time (2)

interpretability

code

visualisation

How Gemma 2 2B represents time: PCA finds a clean two-axis temporal structure at layer 20 that individual GemmaScope SAE features dilute across many latents.

Finding Joy in Gemma2 2B (1)

interpretability

code

visualisation

Based on latest paper “Emotion Concepts and their Function in a Large Language Model”.

Gemma Niti: Fully offline recovery assistant for patients after ischemic stroke

code

Runs open-weight Gemma 4 E2B model locally on Raspberry Pi 5.

Understanding Per Layer Embeddings Method

interpretability

text

Per-layer residual embedding injection used in Gemma4 models.

NanoGPT for Poetry

code

Nano Generative Pre-trained Transformer (GPT) on W.Shakespeare and A. Pushkin datasets.

Pose Detection for Table Tennis

code

tt

video

Skeletal pose landmark detection with MediaPipe and OpenCV.

Pet’s Facial Expression Detection App

code

that classify a pet’s emotions by a photo. Upload an image, get the predicted emotion + confidence. Trained EfficientNetB5 CNN on Kaggle’s pet’s facial expression dataset.

Reinforcement Learning for Chess

visualisation

code

Zero-like reinforcement learning model, Monte Carlo Tree Search for hexapawn chess, similar to the DeepMind’s AlphaZero approach. And visualisation of tree search.

Understanding MeshGraphNet: Spatio-Temporal Graph Node Prediction on Simulation Meshes

text

The goal was to understand how Graph Neural Networks could be used for learning physical dynamics on simulation meshes.The original motivation was industrial: simulation of…

Hello World 👀!

news

💻🏓🎨 🔦❤️🎶

PyTorch Metric Learning

code

Metric learning method with classifier (PyTorch implementation).

Words Representation in Vector Space

code

text

Word2vec explained.

Signature Detection in Documents: Heuristics -> CNNs -> YOLO

code

With YOLOv3 - “You Only Look Once” model.

Transcendental learning

visualisation

Different ways of humans’s learning.

Shapley Values Explaining Machine Learning Models

interpretability

code

Explaining black box machine learning models with Shapley Values, SHAP, …

Explainable Machine Learning in Praxis

interpretability

visualisation

Explainable Machine Learning in praxis.

Mirrorless Safety Windshield System

text

video

Our team AutoGenZ won 2nd place of Valeo Innovation Challenge 2015. With prototype called Mirrorless Safety Windshield System - computer vision system that replaces…