Highlight

A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects

Surveys convolutional neural networks, covering their history, 1-D/2-D/multidimensional convolutions, key models, practical tips, and open issues.

Convolutional neural networks (CNNs) are among the most important deep-learning models, impacting computer vision, NLP, and more. Noting that prior reviews focus on applications rather than a general perspective and omit recent ideas, this survey offers a broader view spanning 1-D, 2-D, and multidimensional convolutions. It traces CNN history, overviews convolution types, introduces classic and advanced models and their key ideas, and derives rules of thumb for functions and hyperparameters via experiments, before reviewing applications and open future directions.

Based on: A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects · IEEE Transactions on Neural Networks and Learning Systems

Curated by Aramai Editorial

Read summary →
Highlight

Temporal Segment Networks: Towards Good Practices for Deep Action Recognition

Proposes temporal segment networks (TSN) with sparse sampling and video-level supervision for effective deep action recognition in video.

Deep convolutional networks dominate still-image recognition, but their advantage for video action recognition has been less clear. This paper seeks principles for designing effective ConvNet architectures that learn from limited data. Its main contribution is the temporal segment network (TSN), which models long-range temporal structure by combining sparse temporal sampling with video-level supervision to learn from whole videos. The authors also study good practices for training ConvNets on video, achieving state-of-the-art results on HMDB51 (69.4%) and UCF101 (94.2%).

Based on: Temporal Segment Networks: Towards Good Practices for Deep Action Recognition · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Using Deep Learning for Image-Based Plant Disease Detection

Trains a deep CNN on 54,306 leaf images to identify 14 crop species and 26 diseases, reaching 99.35% accuracy on held-out data.

Crop diseases threaten food security, yet rapid identification is difficult in regions lacking the necessary infrastructure. Rising smartphone use combined with deep-learning computer vision opens the way to smartphone-assisted diagnosis. Using a public dataset of 54,306 images of diseased and healthy leaves collected under controlled conditions, the authors train a deep CNN to recognize 14 crop species and 26 diseases. It reaches 99.35% accuracy on a held-out test set, demonstrating the feasibility of large-scale smartphone-assisted crop disease diagnosis.

Based on: Using Deep Learning for Image-Based Plant Disease Detection · Frontiers in Plant Science

Curated by Aramai Editorial

Read summary →
Highlight

The Limitations of Deep Learning in Adversarial Settings

Formalizes adversaries against DNNs and introduces algorithms that craft adversarial samples by exploiting the input-output mapping.

Deep neural networks excel at many tasks but training imperfections leave them vulnerable to adversarial samples crafted to cause misclassification. The authors formalize the space of adversaries against DNNs and introduce algorithms that craft such samples from the input-output mapping. In computer vision, their method yields samples humans classify correctly but the DNN misclassifies into chosen targets, with a 97% success rate while altering only 4.02% of features per sample. They also define a hardness measure and outline preliminary defenses.

Based on: The Limitations of Deep Learning in Adversarial Settings · European Symposium on Security and Privacy

Curated by Aramai Editorial

Read summary →
Highlight

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

Introduces the lottery ticket hypothesis: dense networks contain sparse subnetworks that can train in isolation to comparable accuracy.

Pruning can cut trained-network parameter counts by over 90% without hurting accuracy, but the resulting sparse architectures are typically hard to train from scratch. The authors show that standard pruning uncovers subnetworks whose initializations made them trainable, and state the lottery ticket hypothesis: dense, randomly-initialized networks contain subnetworks (winning tickets) that, trained in isolation, match the full network's accuracy in a similar number of iterations. They consistently find such tickets under 10-20% of the original size on MNIST and CIFAR10.

Based on: The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference

Proposes using matching to preprocess data before parametric analysis, making causal estimates more accurate and less model-dependent.

Published causal estimates are often selected from many unseen trial runs, and they can vary widely with choices of control variables, functional forms, and other modeling assumptions, raising doubts about their accuracy. Matching methods promise causal inference with fewer assumptions but are frequently misinterpreted. The authors explain how to avoid these misinterpretations and propose a unified approach: preprocess data with matching, then apply the parametric techniques one would have used anyway, yielding more accurate and less model-dependent inferences.

Based on: Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference · Political Analysis

Curated by Aramai Editorial

Read summary →
Highlight

HMDB: A large video database for human motion recognition

Introduces HMDB, the largest action video database of its time, with 51 categories and about 7,000 manually annotated clips.

Recognition and search in video is an emerging frontier, yet human action datasets lag behind large image datasets, typically offering only about ten categories with near-ceiling performance. To address this, the authors collected the largest action video database to date, with 51 action categories and roughly 7,000 manually annotated clips drawn from sources ranging from digitized movies to YouTube. They use it to evaluate two representative action recognition systems and study robustness under camera motion, viewpoint, video quality, and occlusion.

Based on: HMDB: A large video database for human motion recognition · Vision

Curated by Aramai Editorial

Read summary →
Highlight

Generating Sequences With Recurrent Neural Networks

Shows how LSTM recurrent networks generate complex sequences with long-range structure by predicting one data point at a time.

Long Short-term Memory recurrent neural networks can generate complex sequences with long-range structure simply by predicting one data point at a time. The approach is demonstrated on discrete text data and real-valued online handwriting. It is then extended to handwriting synthesis by letting the network condition its predictions on a text sequence, producing highly realistic cursive handwriting across a wide variety of styles.

Based on: Generating Sequences With Recurrent Neural Networks · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Improved Regularization of Convolutional Neural Networks with Cutout

Introduces cutout, a simple regularizer that randomly masks square regions of CNN inputs during training to reduce overfitting.

Convolutional neural networks learn powerful representations but need large capacity, leaving them prone to overfitting and reliant on regularization to generalize. This paper introduces cutout, a simple technique that randomly masks out square regions of the input during training to improve robustness and overall performance. It is trivial to implement and complements existing data augmentation and other regularizers. Applied to state-of-the-art architectures, cutout yields new best test errors of 2.56% on CIFAR-10, 15.20% on CIFAR-100, and 1.30% on SVHN.

Based on: Improved Regularization of Convolutional Neural Networks with Cutout · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering

Balances the VQA dataset with complementary image pairs (VQA v2.0) to counter language priors and force models to use visual information.

In visual question answering, language and world structure are often easier signals than images, so models can ignore visual input yet appear capable. To counter these language priors, the authors balance the popular VQA dataset by pairing each question with two similar images that yield different answers. The resulting VQA v2.0 dataset is more balanced and roughly twice as large, and state-of-the-art models perform significantly worse on it, showing they exploit language priors. It also enables an interpretable model that explains answers with a counter-example image.

Based on: Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Multimodal Machine Learning: A Survey and Taxonomy

Surveys multimodal machine learning and proposes a taxonomy organized around representation, translation, alignment, fusion, and co-learning.

Human experience is inherently multimodal, spanning sight, sound, touch, smell, and taste, and progress in AI requires interpreting such signals together. Multimodal machine learning builds models that process and relate information across multiple modalities. Rather than focusing on specific applications, this paper surveys recent advances and organizes them in a common taxonomy. Moving beyond the usual early- versus late-fusion split, it identifies five core challenges: representation, translation, alignment, fusion, and co-learning.

Based on: Multimodal Machine Learning: A Survey and Taxonomy · IEEE Transactions on Pattern Analysis and Machine Intelligence

Curated by Aramai Editorial

Read summary →
Highlight

DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation

Personalizes text-to-image diffusion models from a few subject images by binding a unique identifier to the subject for novel-context synthesis.

Large text-to-image models generate high-quality images from prompts but cannot reproduce specific subjects or place them in new contexts. DreamBooth personalizes a pretrained diffusion model by fine-tuning on just a few images of a subject, binding a unique identifier to it via a new class-specific prior preservation loss that leverages the model's semantic prior. The identifier can then synthesize novel photorealistic images of the subject across diverse scenes, poses, views, and lighting, supported by a new dataset and evaluation protocol for subject-driven generation.

Based on: DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →