Official platform of Tsinghua Big Data Research Center, sharing the team's latest research, teaching updates, and big data news.
Google DeepMind’s AlphaGenome, featured on Nature’s cover, reads up to one million DNA bases at once, predicts the functional impact of any mutation across gene expression, splicing, chromatin and protein binding, and outperforms prior models by more than double on key benchmarks.
This article analyzes Sakana AI's three recent papers that challenge traditional Transformer long‑sequence handling by removing positional embeddings, reconstructing position awareness, and adding a fast‑weight external memory, showing how each approach improves ultra‑long text understanding.
The article explains that setting the mean to zero recenters data around the origin, making optimization algorithms converge faster, while scaling variance to one equalizes feature scales so no single feature dominates, illustrated with examples and visualizations of how standardization improves machine‑learning models.
Despite achieving over 95% accuracy on ImageNet, popular models like ResNet, VGG, and EfficientNet can be easily misled by carefully crafted adversarial examples using FGSM, revealing deep learning’s inherent vulnerability and prompting the need for robust defense strategies.
AutoLink introduces an autonomous, iterative schema‑linking approach for Text‑to‑SQL that treats schema discovery as a progressive, agent‑driven exploration, dramatically improving recall while cutting token costs, and outperforms existing database‑level and element‑level methods on large benchmarks such as Spider 2.0‑Lite and BIRD.
The article examines the Test‑Time Training to Discover (TTT‑Discover) approach, which applies reinforcement learning during inference to let large language models continuously improve on single test problems, and reports strong results across mathematics, GPU kernel optimization, algorithm design, and biology.
This article recounts a Tsinghua University PhD student's journey through a multidisciplinary big‑data training program, detailing the acquisition of AI and data‑science skills, the creation of novel algorithms like MicroFlowSAM and ImageRAG, and their successful application to chemical engineering research and industry projects.
Even as Agentic RAG evolves, the quality of the underlying retrieval embedding model remains crucial, and this article compares three training losses—pairwise cosine embedding, triplet margin, and InfoNCE—detailing their inputs, formulas, and practical trade‑offs.
PropMolFlow, a new flow‑matching model introduced by researchers from the University of Florida and NYU, generates property‑guided molecules up to ten times faster than prior SOTA methods while preserving chemical validity and achieving superior performance on benchmarks such as QM9.
The Tsinghua University Big Data Capability Project applied GRU‑based deep learning, pulse‑event encoding, and advanced feature engineering to transform discrete refinery energy data into continuous sequences, achieving prediction accuracies of 84.2%, 82.7% and 81.6% for fuel gas, medium‑pressure and low‑pressure steam respectively.
