Georgii Mikriukov

AI Expert, PhD (Dr.-Ing.)
Berlin, Germany

Download CV
Georgii Mikriukov

About Me

PhD AI Expert with 7+ years of experience in academic and industrial R&D. I work on ML/DL methods for computer vision, NLP, and multimodal systems, focusing on trustworthy AI, including safety validation, model robustness, and interpretability. I have contributed to safety-critical applications (autonomous driving, remote sensing) and delivered both research outputs and production-ready solutions. Experienced in supervising students and PhD candidates, leading research projects, and driving collaborative grant proposals.

Open Source Projects

Local Concept Embeddings
Deep Unsupervised Contrastive Hashing
Cross-Modal Hashing with Noise Robustness
Concept-Based Adversarial Attack Analysis

Local Concept Embeddings

LoCEs (Local Concept Embeddings) provide a way to analyze how DNNs represent object concepts in complex, real-world scenes. Unlike traditional global approaches, LoCEs generate sample-specific embeddings that capture both the target object and its surrounding context within a single, compact representation.

This context-aware analysis helps uncover how models encode, separate, and confuse visual concepts across diverse scenarios. LoCEs reveal meaningful patterns in feature space, supporting model inspection, debugging, and evaluation.

Use cases include:

  • Concept Understanding – Examine how models distinguish objects and their contexts.
  • Sub-Concept Discovery – Identify unlabeled variations within categories (e.g., near vs. distant car).
  • Concept Confusion Detection – Detect overlaps in representations of similar categories (e.g., bus vs. truck).
  • Outlier Detection – Find unusual or challenging examples in the data.
  • Information Retrieval – Search for samples using LoCE-based similarity.
  • Model Comparison – Compare internal feature spaces across architectures or training methods.
LoCEs Example
LoCE optimization and generalization.
LoCEs Concept Separation
Concept confusion across models revealed by LoCEs.
View on GitHub

Deep Unsupervised Contrastive Hashing

DUCH (Deep Unsupervised Contrastive Hashing) is an unsupervised cross-modal retrieval method designed for efficient search and retrieval of semantically related images and text in large-scale datasets. Unlike traditional supervised retrieval methods that rely on extensive labeled data, DUCH learns discriminative feature representations in an unsupervised manner, making it scalable and adaptable for various multi-modal applications.

Key Features:
  • Cross-Modal Hashing: Generates compact binary hash codes for fast and memory-efficient retrieval across different modalities (e.g., image-to-text, text-to-image).
  • Unsupervised Contrastive Learning: Leverages contrastive objectives to preserve semantic relationships, enabling effective retrieval without labeled training data.
  • Adversarial Learning: Enforces alignment between image and text representations using an adversarial loss function, improving retrieval consistency.
  • Multi-Objective Optimization: Combines multiple loss functions to enhance retrieval performance.
Applications:
  • Multi-Modal Search: Enables image-to-text and text-to-image retrieval in diverse domains, including visual recognition, media analysis, and scientific research.
  • Scalable Indexing: Compact hash codes enable efficient large-scale data retrieval with minimal storage and computational cost.
  • Semantic Representation Learning: Captures complex relationships between different modalities, aiding in knowledge discovery and AI-driven content understanding.
DUCH Architecture
DUCH Architecture.
View on GitLab

Cross-Modal Hashing with Noise Robustness

CHNR (Cross-Modal Hashing with Noise Robustness) is an unsupervised technique designed for retrieving images based on textual descriptions, even in the presence of noisy image-text correspondences. It extends DUCH (Deep Unsupervised Contrastive Hashing) by introducing a noise detection module that mitigates errors in training data.

  • Unsupervised Learning – Learns robust cross-modal representations without labeled data, making it adaptable to large-scale datasets.
  • Robust Feature Extraction – Extracts deep representations of both images and texts, ensuring effective cross-modal learning.
  • Noise Detection Module – Identifies and reduces the impact of incorrectly paired image-text data.
  • Efficient Hashing Mechanism – Generates binary hash codes for scalable retrieval with minimal storage requirements.
CHNR Architecture
CHNR Architecture.
View on GitLab

Concept-Based Adversarial Attack Analysis

Concept-Based Adversarial Analysis investigates how adversarial attacks manipulate DNNs at the concept level. This study reveals how adversarial attacks distort, introduce, or remove concepts (i.e., latent features) in a model’s feature space.

Through concept discovery techniques, adversarial perturbations are decomposed into components, and their effects are analyzed across multiple DNN architectures and attack types. It is revealed that adversarial attacks systematically distort concept representations, with adversarial perturbations being linearly decomposable into a small set of shared latent vectors. Analysis demonstrates that attack components exploit target-specific directions.

Concept changes under adversarial attacks
Concept distortions under adversarial attacks.
Similarity of directions of concepts discovered in adversarial attacks targeting 'taxi' class.
Similarity of directions of concepts discovered in adversarial attacks targeting 'taxi' class.
Read paper

Publications

2026

Explaining, Verifying, and Aligning Semantic Hierarchies in Vision-Language Model Embeddings

arXiv preprint, 2026.

Gesina Schwalbe, Mert Keser, Moritz Bayerkuhnlein, Edgar Heinert, Annika Mütze, Marvin Keller, Sparsh Tiwari, Georgii Mikriukov, Diedrich Wolter, Jae Hee Lee, Matthias Rottmann.

PDF GitHub
Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

arXiv preprint, 2026.

Georgii Mikriukov, Grégoire Montavon, Marina M.-C. Höhne.

PDF GitHub

2025

On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs

Explainable Artificial Intelligence, xAI 2025.

Gesina Schwalbe, Georgii Mikriukov, Stavros Gerolymatos, Edgar Heinert, Annika Mütze, Mert Keser, Alois Knoll, Matthias Rottmann.

PDF GitHub
Local Concept Embeddings for Analysis of Concept Distributions in Vision DNN Feature Spaces

International Journal of Computer Vision (IJCV), May, 2025.

Georgii Mikriukov, Gesina Schwalbe, Korinna Bade.

PDF GitHub

2024

Concept-Based Explanations in Computer Vision: Where Are We Going?

Explainable Computer Vision @ European Conference on Computer Vision, ECCV 2024.

Jae Hee Lee, Georgii Mikriukov, Gesina Schwalbe, Stefan Wermter, Diedrich Wolter.

PDF
Locally Testing Model Detections for Semantic Global Concepts

Explainable Artificial Intelligence, xAI 2024.

Franz Motzkus, Georgii Mikriukov, Christian Hellert, Ute Schmid.

PDF
Unveiling the Anatomy of Adversarial Attacks: Concept-Based XAI Dissection of CNNs

Explainable Artificial Intelligence, xAI 2024.

Georgii Mikriukov, Gesina Schwalbe, Franz Motzkus, Korinna Bade.

PDF

2023

Revealing Similar Semantics Inside CNNs: An Interpretable Concept-Based Comparison of Feature Spaces

Machine Learning and Principles and Practice of Knowledge Discovery in Databases, ECML PKDD 2023

Georgii Mikriukov, Gesina Schwalbe, Christian Hellert, Korinna Bade.

PDF
Evaluating the Stability of Semantic Concept Representations in CNNs

Explainable Artificial Intelligence, xAI 2023.

Georgii Mikriukov, Gesina Schwalbe, Christian Hellert, Korinna Bade.

Best industry paper award: Awards – The World Conference on eXplainable Artificial Intelligence

PDF

2022

An Unsupervised Cross-Modal Hashing Method Robust to Noisy Training Image-Text Correspondences

IEEE International Conference on Image Processing, ICIP 2022.

Georgii Mikriukov, Mahdyar Ravanbakhsh, Begüm Demir.

PDF GitLab
Unsupervised Contrastive Hashing for Cross-Modal Retrieval in Remote Sensing

IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022.

Georgii Mikriukov, Mahdyar Ravanbakhsh, Begüm Demir.

PDF GitLab

Patents and Inventions

Method for Finding the Cause of Detection Failures of an Artificial Neural Network

EP 4421682 A1

Georgii Mikriukov, Christian Hellert, Erwin Kraft, Gesina Schwalbe.

View Patent
Vehicle, device, computer program, method for checking a machine learning-based model for an error, machine learning-based model and use thereof

DE 10 2023 212 859 A1

Georgii Mikriukov, Christian Hellert.

View Patent
Vehicle, device, computer program and method for embedding and evaluating visual concepts within the latent feature space of object detectors aimed at improving the safety and reliability of autonomous systems

DE 10 2024 200 029 A1

Erwin Kraft, Gesina Schwalbe, Christian Hellert, Georgii Mikriukov.

View Patent
Vehicle, device, computer program, machine learning-based model, method for training a machine learning-based model, and method for adjusting a sample for training a machine learning-based model

DE 10 2023 212 519 A1

Georgii Mikriukov, Christian Hellert, Gesina Schwalbe.

View Patent

Contact

mikriukov.georgii@gmail.com

linkedin.com/in/georgii-mikriukov

github.com/comrados