Hypercolumns combine multi-layer CNN features for pixel-level object segmentation, boosting SDS from 49.7 to 60.0 mean APr.

Bharath Hariharan, Pablo Arbeláez, Ross Girshick et al.

Proposes a deep neural image captioning model combining CNN and LSTM, achieving BLEU-4 of 27.7, outperforming previous methods.

Oriol Vinyals, Alexander Toshev, Samy Bengio et al.

Proposes LRCN, combining CNN and LSTM for video recognition and captioning, outperforming single models with end-to-end training.

Jeff Donahue, Lisa Anne Hendricks, Marcus Rohrbach et al.

Proposes Fully Convolutional Networks (FCN) for pixel-level semantic segmentation, achieving 62.2% mean IU, surpassing previous SOTA with end-to-end training.

Jonathan Long, Evan Shelhamer, Trevor Darrell

Dictionary learning combined with `1-norm sparse coding improves image representation and recognition accuracy.

Julien Mairal, Francis Bach, Jean Ponce

Study on feature transferability in deep neural networks, revealing significant differences across layers.

Jason Yosinski, Jeff Clune, Yoshua Bengio et al.

A* sampling algorithm leverages Gumbel processes and A* search to efficiently draw exact samples from continuous distributions, outperforming adaptive rejection sampling.

Chris J. Maddison, Daniel Tarlow, Tom Minka

Neural Turing Machines extend neural networks with external memory, enabling inference of simple algorithms like copying and sorting.

Alex Graves, Greg Wayne, Ivo Danihelka

Graph-Sparse LDA uses ontology-guided sparsity, compressing an ASD topic from 119 diagnoses to 6 concepts without losing predictive performance.

Finale Doshi-Velez, Byron Wallace, Ryan Adams

Memory Networks integrate inference and long-term memory to enhance QA task performance.

Jason Weston, Sumit Chopra, Antoine Bordes

This paper analyzes the computational complexity of training neural networks, showing over-parameterized networks are easier to optimize and proposing polynomial activation-based algorithms for depth-2 and depth-3 networks.

Roi Livni, Shai Shalev-Shwartz, Ohad Shamir

Proposes a gradient reversal-based unsupervised domain adaptation method, significantly improving cross-domain image classification accuracy.

Yaroslav Ganin, Victor Lempitsky

Proposed Inception architecture employs multi-scale convolutions and dimension reduction, achieving 28.6% Top-5 error on ImageNet with 1/12 parameters of AlexNet.

Christian Szegedy, Wei Liu, Yangqing Jia et al.

Proposes a deep multi-layer LSTM-based end-to-end sequence-to-sequence model achieving BLEU 34.8 on WMT'14 English-French translation, outperforming phrase-based SMT.

Ilya Sutskever, Oriol Vinyals, Quoc V. Le

Using atomic frequency comb protocol in a 20-meter erbium-doped fiber, this work demonstrates 1532 nm photon storage with high fidelity and preserved entanglement.

Erhan Saglamyurek, Jeongwan Jin, Varun B. Verma et al.

Deep CNNs like AlexNet, VGG, ResNet trained on 14 million images achieved top-5 error rates below 7%, revolutionizing large-scale image recognition.

Olga Russakovsky, Jia Deng, Hao Su et al.

Introduces a neural machine translation model with joint alignment and translation via attention, achieving BLEU 28.45 on WMT’14 English-French.

Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio

Proposes an Extended Dynamic Mode Decomposition (EDMD) to approximate Koopman eigenvalues, eigenfunctions, and modes from data, with proven convergence to Galerkin methods.

Matthew O. Williams, Ioannis G. Kevrekidis, Clarence W. Rowley

Introduces SimLex-999, a gold standard for semantic similarity, covering multiple POS and abstract/concrete concepts, outperforming WordSim-353 and MEN.

Felix Hill, Roi Reichart, Anna Korhonen

Framework for statistical guarantees of EM and gradient EM, combining population and finite-sample analysis.

Sivaraman Balakrishnan, Martin J. Wainwright, Bin Yu