Merge branch 'main' of https://github.com/AppliedMachineLearning-Lab/AppliedMachineLearning-Lab.github.io

Author: hosamelsafty

_posts/2026-04-15-welcome-to-the-aml-lab-blog.md

@@ -5,6 +5,7 @@ date: 2026-04-15
 author: Tobias Deußer
 categories: [News]
 description: false
+tldr: "Welcome to the official blog of the Applied Machine Learning Lab at the University of Bonn: our space for research highlights, course announcements, lab news, and tutorials."
 ---
 
 Welcome to the official blog of the Applied Machine Learning Lab at the University of Bonn! This space is where we share updates about our research, courses, events, and insights from our work in machine learning.

_posts/2026-04-16-Towards-Automated-Recipe-Reconstruction.md

@@ -5,6 +5,7 @@ date: 2026-04-16
 author: Svetlana Schmidt
 categories: [Research]
 description: false
+tldr: "To automate the labor-intensive expansion of nutritional databases, we built a two-stage information retrieval system that matches food items by text and nutrient similarity. Adding SVM-based food category prediction (99% accuracy) boosted retrieval precision to 80%, and we outline an LLM-plus-optimization pipeline for simulating unmatched recipes."
 ---
 
 This is a short summary of our paper **"Towards Automated Recipe Reconstruction: Optimization of Dietary Data Collection using Information Retrieval, Large Language Models and Mathematical Optimization"** by *Svetlana Schmidt, Linda Klasen, Ute Nöthlings, and Rafet Sifa* published in the proceedings of the 2025 IEEE International Conference on Big Data. See [here](https://bonndoc.ulb.uni-bonn.de/xmlui/handle/20.500.11811/14070) for the full paper.

_posts/2026-04-29-leveraging-synthetically-generated-data-for-real-estate-document-classification.md

@@ -5,6 +5,7 @@ date: 2026-04-29
 author: Tobias Deußer
 categories: [Research]
 description: false
+tldr: "Sensitive documents make collecting real training data hard, so we generate realistic synthetic data instead: template-based documents filled by Faker and LLMs. A BERT classifier trained purely on synthetic data reaches 88% page-wise precision on a real-world test set it never saw during training."
 ---
 
 This is a short summary of our paper **"Leveraging Synthetically Generated Data for Real Estate Document Classification"** by *Tobias Deußer, Gregor Ramien, Nico Weber, Maximilian Meidinger, Max Hahnbück, Christian Bauckhage, and Rafet Sifa*, published in the proceedings of the 2025 IEEE International Conference on Big Data. See [here](https://ieeexplore.ieee.org/abstract/document/11400789) for the published version and [here](https://bonndoc.ulb.uni-bonn.de/xmlui/handle/20.500.11811/13972) for the open-access bonndoc version.

_posts/2026-05-22-When-RL-Wins-the-Benchmark-but-Loses-the-Patient.md

@@ -5,6 +5,7 @@ date: 2026-05-22
 author: Armin Berger
 categories: [Research]
 description: false
+tldr: "RL fine-tuning (GRPO) lifted our chest X-ray model's score 23% on one benchmark but dropped it 19% on another: it learned to predict dataset-specific labels, not to read X-rays. The same pattern appears at 50x the budget, so the problem is the recipe, not the resources. Carefully curated SFT generalized better across institutions."
 ---
 
 This is a TL;DR of our paper **"Benchmark Success, Clinical Failure: When Reinforcement Learning Optimizes for Benchmarks, Not Patients"** by *Armin Berger, Manuela Bergau, Helen Schneider, Saad Ahmad, Tom Anglim Lagones, Gianluca Brugnara, Martha Foltyn-Dumitru, Kai Schlamp, Philipp Vollmuth, and Rafet Sifa*, available on [arXiv](https://arxiv.org/abs/2512.23090) and to be published in the Proceedings of IJCNN 2026.

_posts/2026-05-28-From-Retinal-Pixels-to-Patients.md

@@ -5,6 +5,7 @@ date: 2026-05-28
 author: Muskaan Chopra
 categories: [Research]
 description: false
+tldr: "A survey of deep learning for diabetic retinopathy screening (2016–2025), spanning 50+ studies and 20+ datasets. The key takeaway: strong benchmark scores don't equal clinical readiness: the field needs reproducible code, external validation, calibration, and per-patient evaluation."
 ---
 
 This is a short summary of our paper **"From Retinal Pixels to Patients: Evolution of Deep Learning Research in Diabetic Retinopathy Screening"** by *Muskaan Chopra, Lorenz Sparrenberg, Armin Berger, Sarthak Khanna, Jan H. Terheyden, and Rafet Sifa*, published in the proceedings of the 2025 IEEE International Conference on Big Data. See [here](https://ieeexplore.ieee.org/document/11402603) for the published version and [here](https://arxiv.org/abs/2511.11065) for the open-access arxiv version.

_posts/2026-06-24-Towards-Reliable-Machine-Translation.md

@@ -0,0 +1,23 @@
+---
+layout: post
+title: "Towards Reliable Machine Translation: Scaling LLMs for Critical Error Detection and Safety"
+date: 2026-06-24
+author: Muskaan Chopra
+categories: [Research]
+description: false
+tldr: "Instruction-tuned and fine-tuned LLMs can substantially improve critical error detection in machine translation, making them useful safeguards for safer multilingual information access."
+---
+
+This is a short summary of our paper **"Towards Reliable Machine Translation: Scaling LLMs for Critical Error Detection and Safety"** by *Muskaan Chopra, Lorenz Sparrenberg, and Rafet Sifa*, published in the proceedings of ECIR 2026. See [here](https://link.springer.com/chapter/10.1007/978-3-032-21324-2_21) for the published version and [here](https://publica.fraunhofer.de/entities/publication/16269df5-543e-4b26-8b1f-2218fe17e692) for the open-access Fraunhofer Publica version.
+
+Machine translation plays a major role in multilingual information access. It supports public communication, cross-lingual search, professional workflows, and everyday digital assistance. But when translations contain critical meaning errors, the consequences can be serious: wrong medical advice, distorted legal meaning, incorrect financial information, or biased public communication.
+
+This paper studies whether instruction-tuned LLMs can act as safeguards by detecting such critical translation errors.
+
+We frame Critical Error Detection as a binary task: given an English source sentence and a German translation, the model must decide whether the translation contains a meaning-critical error or preserves the intended meaning. We compare encoder-only baselines such as BERT, ModernBERT, mmBERT, and XLM-R with decoder LLMs including GPT-4o, GPT-4o-mini, LLaMA, and GPT-OSS models.
+
+![Figure 1: Towards Reliable Machine Translation](/assets/blog/2026-06-24-Towards-Reliable-Machine-Translation.png)
+
+The experiments cover zero-shot prompting, few-shot prompting, tuned prompts, committee voting, and fine-tuning. The results show that scaling and instruction alignment improve sensitivity to critical errors, and that fine-tuned decoder models can outperform strong encoder-only baselines across WMT21, WMT22, and SynCED-EnDe 2025.
+
+The broader point is that CED should not be viewed only as a technical benchmark. It is a practical safety layer for multilingual information systems, especially when translations are used in contexts where factual accuracy, fairness, and trust matter.

_posts/2026-06-25-History-Rhymes-Macro-Contextual-Retrieval.md

@@ -0,0 +1,21 @@
+---
+layout: post
+title: "History Rhymes: Macro-Contextual Retrieval for Robust Financial Forecasting"
+date: 2026-06-25
+author: Muskaan Chopra
+categories: [Research]
+description: false
+tldr: "Forecasting becomes more robust when the model stops treating today as isolated and instead asks which past macroeconomic regimes today resembles."
+---
+
+This is a short summary of our paper **"History Rhymes: Macro-Contextual Retrieval for Robust Financial Forecasting"** by *Sarthak Khanna, Armin Berger, Muskaan Chopra, David Berghaus, and Rafet Sifa*, published in the proceedings of the 2025 IEEE International Conference on Big Data. See [here](https://xplorestaging.ieee.org/document/11400811) for the published version and [here](https://publica.fraunhofer.de/entities/publication/51ca8ad1-eaf3-464a-99a6-8ff9f07743c2) for the open-access Fraunhofer Publica version.
+
+Financial forecasting models often look convincing during validation and then fail when the market regime changes. This is not surprising: markets are non-stationary. Inflation, unemployment, interest-rate spreads, GDP growth, sector narratives, and investor behavior all shift over time.
+
+The idea behind this paper is simple: financial history does not repeat exactly, but it often rhymes.
+
+We introduce a macro-contextual retrieval framework that grounds each prediction in historically similar macroeconomic situations. Instead of relying only on today’s price indicators or news sentiment, the model retrieves past periods with similar combinations of financial text and macro indicators such as CPI, unemployment, yield spread, and GDP growth. These retrieved historical analogues then provide context for the forecast.
+
+The system is trained on S&P 500 data from 2007-2023 and evaluated out-of-distribution on AAPL and XOM in 2024. The important part is that retrieval is causal: for each prediction, the model can only retrieve earlier historical periods, not future information.
+
+The main result is that macro-conditioned retrieval is more robust under distribution shift than static numeric, text-only, or naive multimodal baselines. It achieves the only positive out-of-sample trading outcomes on both AAPL and XOM, while also producing interpretable evidence chains through retrieved historical neighbors.

assets/blog/2026-06-24-Towards-Reliable-Machine-Translation.png

@@ -28,7 +28,7 @@ If you want to reach out to us, please write to:
     {% endif %}
   </div>
   <div class="post-excerpt">
-    {{ latest.excerpt | strip_html | truncatewords: 60 }}
+    {% if latest.tldr %}{{ latest.tldr }}{% else %}{{ latest.excerpt | strip_html | truncatewords: 60 }}{% endif %}
   </div>
   <a href="{{ latest.url | relative_url }}" class="read-more-link">Read more &rarr;</a>
 </article>

index.md

@@ -16,7 +16,7 @@ <h2 class="post-card-title"><a href="{{ post.url | relative_url }}">{{ post.titl
       {% endif %}
     </div>
     <div class="post-excerpt">
-      {{ post.excerpt | strip_html | truncatewords: 60 }}
+      {% if post.tldr %}{{ post.tldr }}{% else %}{{ post.excerpt | strip_html | truncatewords: 60 }}{% endif %}
     </div>
     <a href="{{ post.url | relative_url }}" class="read-more-link">Read more &rarr;</a>
   </article>