Woojong Koh

#HunyuanOCR sets a new standard in end-to-end OCR. Built on Hunyuan’s native multimodal architecture, this lightweight 1B-parameter VLM is delivering state-of-the-art results across multiple industry benchmarks. What stands out is its ability to handle complex, multilingual document parsing with ease while staying efficient enough for real-world deployment. It performs strongly across a range of practical scenarios — from text spotting and open-field information extraction to video subtitle extraction and on-the-fly photo translation. A compact model with serious capability, and a solid step forward for production-grade OCR. Link : https://lnkd.in/gYwHvvAj

28

Video-As-Prompt: Unified Semantic Control for Video Generation Just came across one of the most exciting research papers in video generation lately — by Yuxuan Bian, Xin Chen, Zenan Li, Tiancheng Zhi, Shen Sang, Linjie Luo, and Qiang Xu. Achieves state-of-the-art results with 38.7% user preference rate, rivaling commercial models, and includes 100K+ paired video dataset. Instead of giving text prompts or rigid structure maps, what if you could guide an AI model just by showing it another video? That’s exactly what Video-As-Prompt does. It treats a reference video as a “semantic prompt.” Basically, it lets the model understand and transfer meaning from one video to another, without hard-coded rules or special training for every scenario. Why does this matter? Because semantic control is the missing piece for true creative direction in video AI. We’ve had strong generators, but fine-tuned, generalized control, the ability to say, “make it look like this, feel like that,” has always been the bottleneck. It opens new doors for film, gaming, education, and all kinds of content creation where meaning, not just motion, matters. Let me know which topics you would like to see new papers on, via comments or messages.

4

🔹 SpikeLink Update . Full Adapter Ecosystem Online SpikeLink now supports a complete cross-platform adapter layer for neuromorphic and spike-based systems. New adapters added: • Nengo (neuromorphic modeling) • Lava (Intel Loihi stack) • SpikeInterface (electrophysiology / spike sorting) Alongside existing support for: • Neo / EBRAINS • Brian2 • Tonic (event camera datasets) • PyNN • NEST All adapters enforce the same transport contract: • Exact spike/event count preservation • Causal ordering monotonicity • Explicit timing bounds • Deterministic encode → decode reconstruction • Monotonic, measurable degradation under noise Internally, SpikeLink maintains seconds as the canonical time base, with explicit conversion at the adapter boundary. v2 timing remains governed by: chunk_start_10us (10 μs resolution) Two declared modes: • Mode A — Timing-Exact (within quantization floor) • Mode B — Chunked (packet efficiency trade-off, distortion quantified) No hidden behavior. No silent timing drift. SpikeLink does not replace digital computation. It defines transport semantics for spike-native systems operating inside conventional digital infrastructure. Adapters are now modular via extras: pip install spikelink[nengo] pip install spikelink[lava] pip install spikelink[spikeinterface] pip install spikelink[full] Core unit suite: passing. Platform suites: aligned and under pytest integration. If you’re working on neuromorphic transport, Loihi workflows, PyNN simulations, or spike-sorted electrophysiology — the transport layer is now standardized. https://lnkd.in/giPq9pH5.

1

We all know Neural Networks are black box models So now what is this ? Visualizing Neural Networks - just woaooo At Design Korea 2024, Kim Seonghyun presented a real-time interactive visualization of a neural network trained on MNIST using PyTorch — visualized through Processing and rendered with OpenGL instancing. You can literally watch the data flow: → 784-pixel input → Layer-wise transformation → Flattened vector → Softmax output across 0–9 #AI #DeepLearning #Explainability #NeuralNetworks #AIxArt

6

🚀I’m very excited to share two of our recent works on offline reinforcement learning and its application in sequential recommendation, accepted at SIGIR 2025 and KDD 2025! 1. Offline Trajectory Optimization for Offline Reinforcement Learning
 (accepted at #KDD2025) Authors: Ziqi Zhao, Zhaochun Ren, Liu Yang, Yunsen Liang, Fajie Yuan, Pengjie Ren, Zhumin Chen, Jun Ma, Xin Xin
 Code: https://lnkd.in/esKZJrBc
 Preprint: https://lnkd.in/eMJ7T4D6 In this work, we propose OTTO, a plug-in framework that improves offline RL by generating long-horizon trajectories using an ensemble of Transformers. We further introduce an uncertainty-aware evaluator to correct low-confidence simulations, leading to more reliable and effective data augmentation. OTTO significantly boosts policy performance across standard and challenging RL benchmarks. 2. Improving Sequential Recommenders through Counterfactual Augmentation of System Exposure
 (accepted at #SIGIR2025) Authors: Ziqi Zhao, Zhaochun Ren, Jiyuan Yang, Zuming Yan, Zihan Wang, Liu Yang, Pengjie Ren, Zhumin Chen, Maarten de Rijke, Xin Xin
 Code: https://lnkd.in/eQK-KDzp
 Preprint: https://lnkd.in/ekpVBR-s We investigate sequential recommendation from a system exposure perspective, and propose CaseRec, which uses a decision transformer-based offline RL model to model user exposure behaviors. By generating counterfactual exposure sequences via data augmentation and a user simulator, CaseRec explores unseen user interests and effectively reduces exposure bias.

54

🛠 Tech Stack & Skills: ~Streamlit – Interactive UI development ~LangChain – Retrieval-Augmented Generation (RAG) pipeline design ~FAISS – Vector database for semantic search ~Google Gemini Pro 2.5 – Context-aware LLM integration ~SentenceTransformers – Text embeddings generation ~PyPDF2 – PDF text extraction ~Python – End-to-end system development ~LLM Prompt Engineering – Optimizing responses for domain-specific queries ~Document Parsing & Chunking – Preparing structured context for retrieval ~Semantic Search Optimization – Improving relevance of retrieved chunks 🔗 GitHub Repo: https://lnkd.in/gPQKzHp4

2

New short course: DSPy: Build and Optimize Agentic Apps DSPy is a powerful open-source framework for automatically tuning prompts for GenAI applications. In this course, you'll learn to use DSPy, together with MLflow. This is built in partnership with Databricks and taught by Chen Qian, co-lead of the DSPy framework. Many AI builders spend hours hand-tuning prompts. When given a set of evals, DSPy automates this process. It’s especially useful for optimizing prompts, including few-shot prompts, in complex agentic AI workflows. Further, if you switch an application to a newer LLM, performance can degrade if your prompts were optimized to the previous model. DSPy automatically optimizes the entire system for the new LLM as well, using just a few evaluation examples. This course teaches DSPy works, and best practices for using it. You’ll write programs using DSPy’s signature-based programming model, debug them with MLflow tracing -- to gain visibility into how different parts of a pipeline, as well as how the overall system, are performing -- and automatically improve their accuracy with DSPy Optimizer. Please sign up here: https://lnkd.in/gdjae8AX

4,671

190 Comments

# Modul 5 : Data platform Analytics Engineering with Bruin (Local / DuckDB) I built a complete, production-style analytics pipeline locally using Bruin + DuckDB Key learnings & skills: • I designed a multi-layer data pipeline: ingestion → staging → reports • I Ingested raw NYC Taxi data (Parquet + CSV seed) into DuckDB • I built incremental models using time_interval • I defined dependencies correctly between assets • Implemented data quality checks (not null, unique, non-negative) • Debugged real-world issues: - YAML parsing errors - Missing refs - DuckDB file locks (solved via single-worker execution) • I Validated outputs using SQL: - row counts - date ranges - reconciliation between staging and reporting layers • I Queried and inspected tables via information_schema This hands-on work really highlighted the difference between writing SQL and engineering reliable data pipelines. here the link: https://lnkd.in/epRpYsjF What’s next – Part 5: Deploy to Cloud • Deploy the Bruin pipeline to a cloud data warehouse • Replace DuckDB with a cloud backend (e.g. BigQuery) • Configure cloud connections & environments • Run the same pipeline at scale • Prepare the pipeline for production usage • Understand cost, performance, and orchestration considerations #JustBeDataEngineer #AnalyticsEngineering #Bruin #DuckDB #DataPipelines #Zoomcamp_2026

3

1 Comment