Highlight

ShinyGO: a graphical gene-set enrichment tool for animals and plants

Introduces ShinyGO, a graphical web tool for gene-set enrichment analysis across many plant, animal, archaeal, and bacterial species.

ShinyGO is a graphical web application for gene-set enrichment analysis, helping researchers connect gene lists from genome-wide studies to pathways and functional categories like gene ontology. It draws on a large annotation database from Ensembl and STRING-db covering 59 plant, 256 animal, 115 archaeal, and 1678 bacterial species. Novel features include graphical visualization of enrichment results and gene characteristics, plus API access to KEGG and STRING for pathway diagrams and protein-protein interaction networks.

Based on: ShinyGO: a graphical gene-set enrichment tool for animals and plants · Bioinform.

Curated by Aramai Editorial

Read summary →
Highlight

CSPNet: A New Backbone that can Enhance Learning Capability of CNN

Proposes Cross Stage Partial Network (CSPNet), a CNN backbone that cuts inference computation by reducing duplicate gradient information.

CSPNet addresses the heavy inference computation of CNNs, which the authors attribute to duplicate gradient information during optimization. By integrating feature maps from the beginning and end of each network stage, the design respects gradient variability. Experiments show a 20% reduction in computation with equal or superior ImageNet accuracy, and stronger AP50 on MS COCO object detection. The module is easy to implement and works with ResNet, ResNeXt, and DenseNet.

Based on: CSPNet: A New Backbone that can Enhance Learning Capability of CNN · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

Curated by Aramai Editorial

Read summary →
Highlight

Group Normalization

Proposes Group Normalization, a batch-size-independent alternative to Batch Normalization that stays accurate with small batches.

Batch Normalization (BN) enabled deep network training, but normalizing along the batch dimension makes its error rise sharply as batch size shrinks, limiting its use for large models and memory-constrained tasks like detection, segmentation, and video. This paper proposes Group Normalization (GN), which divides channels into groups and computes the mean and variance within each group, making it independent of batch size. On ResNet-50 for ImageNet, GN has 10.6% lower error than BN at batch size 2, matches BN at typical sizes, and transfers well to COCO and Kinetics tasks.

Based on: Group Normalization · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection

Introduces Grounding DINO, an open-set object detector that fuses the DINO detector with grounded pre-training to detect objects from text prompts.

Grounding DINO is an open-set object detector that marries the Transformer-based DINO detector with grounded pre-training, letting it detect arbitrary objects specified by category names or referring expressions. To generalize to open-set concepts, it introduces language into a closed-set detector via a tight fusion design with a feature enhancer, language-guided query selection, and a cross-modality decoder. Evaluated on COCO, LVIS, ODinW, and RefCOCO/+/g, it reaches 52.5 AP on zero-shot COCO transfer and sets a new record of 26.1 mean AP on the ODinW zero-shot benchmark.

Based on: Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Taming Transformers for High-Resolution Image Synthesis

Combines CNN-learned image vocabularies with transformers to synthesize high-resolution, controllable images, including megapixel outputs.

Transformers model long-range interactions well but, unlike CNNs, lack a local inductive bias, making them expressive yet computationally infeasible for long sequences such as high-resolution images. This paper combines both: CNNs learn a context-rich vocabulary of image constituents, and transformers efficiently model how these constituents compose within high-resolution images. The approach supports conditional synthesis controlled by object classes or segmentations, and presents the first semantically-guided synthesis of megapixel images with transformers.

Based on: Taming Transformers for High-Resolution Image Synthesis · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

TCMSP: a database of systems pharmacology for drug discovery from herbal medicines

Presents TCMSP, a systems-pharmacology database of Chinese herbal medicines linking compounds, ADME properties, targets, and diseases for drug discovery.

TCMSP is a systems-pharmacology database and analysis platform built to support drug discovery from herbal medicines and to bridge modern and traditional medicine. It covers all 499 herbs in the Chinese pharmacopoeia with 29,384 ingredients, 3,311 targets, and 837 associated diseases, and provides twelve ADME-related properties such as oral bioavailability, drug-likeness, and blood-brain barrier data for screening. It automatically builds compound-target and target-disease networks to help reveal the mechanisms of action of Chinese herbs, and is freely available.

Based on: TCMSP: a database of systems pharmacology for drug discovery from herbal medicines · Journal of Cheminformatics

Curated by Aramai Editorial

Read summary →
Highlight

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

Presents Qwen2-VL, a vision-language model series with dynamic-resolution processing and multimodal position embeddings, scaled to 72B parameters.

Qwen2-VL is an upgraded vision-language model series that replaces fixed-resolution visual processing with a Naive Dynamic Resolution mechanism, encoding images of varying resolutions into different numbers of visual tokens for more efficient, human-like representations. It adds Multimodal Rotary Position Embedding (M-RoPE) to fuse positional information across text, images, and video, and processes images and video under a unified paradigm. Scaling the model to 2B, 8B, and 72B parameters, the 72B version rivals GPT-4o and Claude3.5-Sonnet across multimodal benchmarks.

Based on: Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Dueling Network Architectures for Deep Reinforcement Learning

Introduces the dueling network architecture that separately estimates state value and action advantage for model-free deep reinforcement learning.

This paper presents a new neural network architecture for model-free reinforcement learning. The dueling network uses two separate estimators, one for the state value function and one for the state-dependent action advantage function, factoring the value estimate to generalize learning across actions without changing the underlying RL algorithm. This design improves policy evaluation when many actions have similar values, and it enables the agent to outperform the previous state of the art on the Atari 2600 domain.

Based on: Dueling Network Architectures for Deep Reinforcement Learning · International Conference on Machine Learning

Curated by Aramai Editorial

Read summary →
Highlight

Bidirectional LSTM-CRF Models for Sequence Tagging

Applies LSTM, BI-LSTM, LSTM-CRF, and BI-LSTM-CRF models to sequence tagging, reaching state-of-the-art on POS, chunking, and NER.

This paper proposes a range of LSTM-based models for sequence tagging, including LSTM, bidirectional LSTM, LSTM with a CRF layer, and bidirectional LSTM with a CRF layer (BI-LSTM-CRF). It is the first work to apply a BI-LSTM-CRF model to NLP sequence tagging benchmarks. The model uses both past and future input features via the bidirectional LSTM and sentence-level tag information via the CRF layer, producing state-of-the-art or comparable accuracy on POS tagging, chunking, and NER while being robust and less reliant on word embeddings.

Based on: Bidirectional LSTM-CRF Models for Sequence Tagging · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

A review of feature selection techniques in bioinformatics

A review and taxonomy of feature selection techniques and their applications across common and emerging bioinformatics problems.

Feature selection has become a clear need in many bioinformatics applications, drawing both on established machine learning and data mining methods and on techniques developed specifically for biological data. This article introduces readers to the possibilities of feature selection, providing a basic taxonomy of the techniques and discussing their use, variety, and potential across a range of established and emerging bioinformatics applications.

Based on: A review of feature selection techniques in bioinformatics · Bioinform.

Curated by Aramai Editorial

Read summary →
Highlight

Prioritized Experience Replay

Proposes prioritized experience replay, which samples important transitions more often to make deep reinforcement learning more efficient.

Experience replay lets online reinforcement learning agents reuse past experiences, but uniform sampling replays transitions at their original frequency regardless of importance. This paper introduces a framework that prioritizes experience so that significant transitions are replayed more often, improving learning efficiency. Applied to Deep Q-Networks (DQN), prioritized replay sets a new state of the art, outperforming DQN with uniform replay on 41 of 49 Atari games.

Based on: Prioritized Experience Replay · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

HellaSwag: Can a Machine Really Finish Your Sentence?

Introduces HellaSwag, an adversarially-filtered commonsense NLI benchmark that is easy for humans but hard for state-of-the-art models.

HellaSwag is a challenge dataset for commonsense natural language inference, where a model must pick the most plausible continuation of an event description. It is built with Adversarial Filtering, in which discriminators iteratively select machine-generated wrong answers, scaling examples into a 'Goldilocks' zone that is trivial for humans (>95%) yet hard for top models (<48%). The work sheds light on pretrained models and argues that benchmarks should co-evolve adversarially with the state of the art.

Based on: HellaSwag: Can a Machine Really Finish Your Sentence? · Annual Meeting of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →