Highlight

Optimal approximations by piecewise smooth functions and associated variational problems

Introduces and studies three variational problems for optimally approximating images by piecewise smooth functions, motivated by computer vision.

This paper introduces and studies three new variational problems motivated by applications to computer vision. A fundamental problem is to appropriately decompose the domain of an image function g(x,y)—the light intensity striking a plane domain when a three-dimensional world is viewed from a point. Because light reflected from distinct objects strikes different regions, the image is generally discontinuous along the boundaries, or edges, between them. The variational problems formalize optimal approximation of such images by piecewise smooth functions over these regions.

Based on: Optimal approximations by piecewise smooth functions and associated variational problems · Semantic Scholar

Curated by Aramai Editorial

Read summary →
Highlight

Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

Introduces the Exponential Linear Unit (ELU), an activation with negative values that speeds deep network training and improves accuracy over ReLUs.

The paper introduces the exponential linear unit (ELU), an activation that speeds up deep network learning and raises accuracy. Like ReLUs, ELUs avoid vanishing gradients via the identity for positive inputs, but their negative values push mean activations toward zero, reducing bias shift and bringing gradients closer to the natural gradient. ELUs saturate to a negative value for small inputs, giving noise-robust deactivation. Experiments show faster learning and better generalization than ReLUs and LReLUs, with top CIFAR results and under 10% single-crop ImageNet error.

Based on: Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs) · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

ChatGPT for good? On opportunities and challenges of large language models for education

A position paper examining opportunities and challenges of large language models for education from student and teacher perspectives.

This position paper examines the opportunities and challenges of large language models in education, from student and teacher perspectives. It highlights how such models can create educational content, improve student engagement and interaction, and personalize learning. It argues teachers and learners must build competencies to grasp the technology's limitations and brittleness, aided by strategies for critical thinking and fact checking. Challenges such as output bias, human oversight, and misuse are discussed, with recommendations for responsible, ethical use.

Based on: ChatGPT for good? On opportunities and challenges of large language models for education · Learning and Individual Differences

Curated by Aramai Editorial

Read summary →
Highlight

ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks

Proposes Efficient Channel Attention (ECA), a lightweight module using 1D convolution without dimensionality reduction to boost CNN accuracy cheaply.

Channel attention improves deep CNNs, but most modules add substantial complexity. This paper proposes Efficient Channel Attention (ECA), which uses very few parameters yet yields clear gains. Analyzing SENet, the authors show avoiding dimensionality reduction matters and that local cross-channel interaction preserves performance while cutting complexity; it is implemented via 1D convolution with an adaptively chosen kernel size. On ResNets and MobileNetV2, ECA improves classification, detection and segmentation with negligible overhead.

Based on: ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Seaborn: Statistical Data Visualization

Presents seaborn, a Python library offering a high-level, dataset-oriented interface to matplotlib for statistical data visualization.

Seaborn is a Python library for statistical graphics, offering a high-level interface to matplotlib and integrating with pandas data structures. Its declarative, dataset-oriented API lets users translate questions about data into graphics: given a dataset and plot specification, seaborn maps data values to visual attributes like color, size, and style, computes statistical transformations internally, and adds informative labels and legends. It also produces multi-panel figures for comparisons, supporting rapid exploratory analysis and polished, publication-quality output.

Based on: Seaborn: Statistical Data Visualization · Journal of Open Source Software

Curated by Aramai Editorial

Read summary →
Highlight

Dense Passage Retrieval for Open-Domain Question Answering

Introduces Dense Passage Retrieval, a dual-encoder that learns dense embeddings to outperform BM25 for open-domain question answering passage retrieval.

Open-domain question answering depends on efficient passage retrieval, traditionally handled by sparse models like TF-IDF or BM25. This work shows retrieval can be done using dense representations alone, with embeddings learned from a small number of question-passage pairs via a simple dual-encoder framework. Across several open-domain QA datasets, the dense retriever beats a strong Lucene-BM25 system by 9-19% absolute in top-20 retrieval accuracy, and helps the end-to-end QA system set new state-of-the-art results on multiple benchmarks.

Based on: Dense Passage Retrieval for Open-Domain Question Answering · Conference on Empirical Methods in Natural Language Processing

Curated by Aramai Editorial

Read summary →
Highlight

YOLOX: Exceeding YOLO Series in 2021

Introduces YOLOX, an anchor-free YOLO detector with a decoupled head and SimOTA label assignment achieving state-of-the-art speed-accuracy.

YOLOX upgrades the YOLO object-detection series by switching to an anchor-free design and adding a decoupled head and the SimOTA label assignment strategy. These changes yield state-of-the-art results across model sizes: YOLO-Nano reaches 25.3% AP on COCO, YOLOv3 is boosted to 47.3% AP, and YOLOX-L attains 50.0% AP at 68.9 FPS on a Tesla V100, exceeding YOLOv5-L by 1.8% AP. A single YOLOX-L model won first place in the CVPR 2021 Streaming Perception Challenge, and deploy versions support ONNX, TensorRT, NCNN, and OpenVINO.

Based on: YOLOX: Exceeding YOLO Series in 2021 · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Large Scale GAN Training for High Fidelity Natural Image Synthesis

Trains GANs (BigGAN) at the largest scale yet, using a truncation trick to set new state of the art in class-conditional ImageNet image synthesis.

Generating high-resolution, diverse samples from complex datasets like ImageNet remains difficult. The authors train GANs at the largest scale yet attempted and study the resulting instabilities. Applying orthogonal regularization to the generator enables a 'truncation trick' that trades off sample fidelity against variety by shrinking input variance. These BigGAN models set a new state of the art in class-conditional synthesis, reaching an Inception Score of 166.5 and FID of 7.4 at 128x128 on ImageNet, up from 52.52 IS and 18.6 FID.

Based on: Large Scale GAN Training for High Fidelity Natural Image Synthesis · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Database resources of the National Center for Biotechnology Information

Reviews the NCBI's online databases and tools, including GenBank, PubMed, Entrez, and BLAST, with new and updated resources.

The National Center for Biotechnology Information (NCBI) provides online resources for biological data, including the GenBank nucleic acid sequence database and the PubMed literature database. Its Entrez system offers search and retrieval across 39 distinct databases, with E-utilities as its programming interface and custom BLAST for specialized data sets. New resources include PubMed Data Management, RefSeq Functional Elements, genome data download, variation services, and Magic-BLAST, alongside updates to the genome data viewer, Gene, and PubChem.

Based on: Database resources of the National Center for Biotechnology Information · Nucleic Acids Res.

Curated by Aramai Editorial

Read summary →
Highlight

Neural Architecture Search with Reinforcement Learning

Uses an RNN trained with reinforcement learning to automatically generate neural network architectures that rival human-designed models.

This paper uses a recurrent network to generate neural network descriptions and trains this RNN with reinforcement learning to maximize the expected validation accuracy of generated architectures. On CIFAR-10, the method designs a novel architecture from scratch that rivals the best human-invented designs, achieving a 3.65% test error—0.09% better and 1.05x faster than a comparable prior model. On Penn Treebank, it composes a new recurrent cell that beats the LSTM, reaching 62.4 perplexity, and transfers to character language modeling with 1.214 perplexity.

Based on: Neural Architecture Search with Reinforcement Learning · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

bold: The Barcode of Life Data System (http://www.barcodinglife.org)

Presents BOLD, a web-based informatics workbench for acquiring, storing, analyzing, and publishing DNA barcode records.

The Barcode of Life Data System (BOLD) is an informatics workbench that supports the acquisition, storage, analysis, and publication of DNA barcode records. By assembling molecular, morphological, and distributional data, it bridges a traditional bioinformatics gap and is freely available to any DNA barcoding researcher. Its specialized services help assemble records meeting standards for BARCODE designation in global sequence databases. Web-based delivery and a flexible data security model position it to support projects involving broad research alliances.

Based on: bold: The Barcode of Life Data System (http://www.barcodinglife.org) · Molecular Ecology Notes

Curated by Aramai Editorial

Read summary →
Highlight

Empirical Analysis of Predictive Algorithms for Collaborative Filtering

Describes and compares collaborative filtering algorithms—correlation, vector similarity, and Bayesian methods—for recommender prediction.

This paper describes and compares several collaborative filtering algorithms for predicting items a user might like, including correlation coefficients, vector-based similarity, and Bayesian methods. Accuracy is measured with two metric classes: average absolute deviation over individual predictions, and the utility of a ranked recommendation list. Across three domains, four protocols, and both metrics, Bayesian networks with decision-tree nodes and correlation methods outperform Bayesian clustering and vector similarity; the best depends on the dataset and application.

Based on: Empirical Analysis of Predictive Algorithms for Collaborative Filtering · Conference on Uncertainty in Artificial Intelligence

Curated by Aramai Editorial

Read summary →