Highlight

Knowledge Distillation: A Survey

A comprehensive survey of knowledge distillation: knowledge types, training schemes, teacher-student architectures, algorithms, and applications.

Deep neural networks succeed by scaling to large data and billions of parameters, but their size and computational cost hinder deployment on resource-limited devices like phones and embedded systems. Knowledge distillation is a representative model compression technique that trains a small student model to learn from a large teacher. This survey reviews it across knowledge categories, training schemes, teacher-student architectures, distillation algorithms, performance comparisons, and applications, and discusses open challenges and future directions.

Based on: Knowledge Distillation: A Survey · International Journal of Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Neural Architectures for Named Entity Recognition

Presents neural network architectures for named entity recognition, published at the 2016 NAACL conference in San Diego.

The available source record provides a bibliographic presentation note rather than a technical summary, indicating that this work was presented at the 2016 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), held in San Diego, California, on June 12-17, 2016. As its title states, the paper introduces neural network architectures for the task of named entity recognition.

Based on: Neural Architectures for Named Entity Recognition · North American Chapter of the Association for Computational Linguistics

Curated by Aramai Editorial

Read summary →
Highlight

Fully-Convolutional Siamese Networks for Object Tracking

Introduces a fully-convolutional Siamese network, trained offline for object tracking, that runs beyond real-time with state-of-the-art accuracy.

Arbitrary object tracking is traditionally handled by learning an appearance model online from the video itself, which limits the richness of the model. Adapting deep networks online via stochastic gradient descent restores expressiveness but badly hurts speed. This paper trains a fully-convolutional Siamese network end-to-end on the ILSVRC15 video object detection dataset, then uses it in a basic tracker. Despite its simplicity, the tracker runs beyond real-time frame-rates and achieves state-of-the-art performance across multiple benchmarks.

Based on: Fully-Convolutional Siamese Networks for Object Tracking · ECCV Workshops

Curated by Aramai Editorial

Read summary →
Highlight

Learning Deconvolution Network for Semantic Segmentation

Proposes a deep deconvolution network with unpooling layers, applied proposal-wise, for detailed multi-scale semantic segmentation.

This work proposes a semantic segmentation algorithm based on a deep deconvolution network built on top of the convolutional layers of VGG-16. The deconvolution network, composed of deconvolution and unpooling layers, predicts pixelwise class labels and segmentation masks. Applied to each object proposal and combined, it captures detailed structures and handles objects at multiple scales, addressing limitations of fully convolutional methods. On PASCAL VOC 2012, an ensemble with an FCN reaches 72.5% accuracy, the best among methods trained without Microsoft COCO.

Based on: Learning Deconvolution Network for Semantic Segmentation · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Variational Graph Auto-Encoders

Introduces the variational graph auto-encoder (VGAE), a VAE-based framework for unsupervised representation learning on graph-structured data.

The variational graph auto-encoder (VGAE) adapts the variational auto-encoder to graph-structured data, using latent variables to learn interpretable representations of undirected graphs. The authors instantiate it with a graph convolutional network encoder and a simple inner product decoder. On link prediction in citation networks, VGAE achieves competitive results. Unlike many prior unsupervised graph models, it naturally incorporates node features, which significantly boosts predictive performance across benchmark datasets.

Based on: Variational Graph Auto-Encoders · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Fourier Neural Operator for Parametric Partial Differential Equations

Introduces a neural operator that parameterizes the integral kernel in Fourier space to learn solution maps for families of parametric PDEs.

Neural operators generalize neural networks from mappings between Euclidean spaces to mappings between function spaces, learning an entire family of PDEs rather than a single instance. This work parameterizes the integral kernel directly in Fourier space, yielding an expressive and efficient architecture. Tested on Burgers' equation, Darcy flow, and Navier-Stokes (including turbulence), the Fourier neural operator reaches state-of-the-art accuracy and runs up to three orders of magnitude faster than traditional PDE solvers.

Based on: Fourier Neural Operator for Parametric Partial Differential Equations · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

SimCSE: Simple Contrastive Learning of Sentence Embeddings

Presents SimCSE, a simple contrastive learning framework for sentence embeddings, with unsupervised dropout-based and supervised NLI-based variants.

SimCSE is a simple contrastive learning framework for sentence embeddings. Its unsupervised variant predicts an input sentence from itself using only dropout as noise; dropout acts as minimal augmentation, and removing it causes representation collapse. The supervised variant adds NLI pairs, using entailment as positives and contradiction as hard negatives. On semantic textual similarity, unsupervised and supervised BERT-base reach 76.3% and 81.6% Spearman correlation—gains of 4.2% and 2.2% over prior best—and contrastive learning makes the embedding space more uniform.

Based on: SimCSE: Simple Contrastive Learning of Sentence Embeddings · Conference on Empirical Methods in Natural Language Processing

Curated by Aramai Editorial

Read summary →
Highlight

Large language models encode clinical knowledge

Introduces MultiMedQA and Med-PaLM, evaluating LLMs on medical question answering with a human framework and instruction prompt tuning.

MultiMedQA is a benchmark combining six medical question-answering datasets and a new online-search dataset, HealthSearchQA, paired with a human evaluation framework covering factuality, comprehension, reasoning, harm, and bias. Evaluating the 540B PaLM and its instruction-tuned Flan-PaLM, the latter reaches state-of-the-art accuracy on all multiple-choice sets, including 67.6% on MedQA—over 17% above prior work. Instruction prompt tuning then yields Med-PaLM, which performs encouragingly but remains inferior to clinicians, underscoring the need for better evaluation.

Based on: Large language models encode clinical knowledge · Nature

Curated by Aramai Editorial

Read summary →
Highlight

Congestion avoidance and control

Diagnoses 1980s Internet congestion collapse and adds seven TCP algorithms—including slow-start and fast retransmit—rooted in conservation of packets.

In 1986 the Internet suffered a series of congestion collapses, with throughput between two nearby sites dropping from 32 Kbps to 40 bps. Investigating whether 4.3BSD TCP was misbehaving, the author added seven algorithms to 4BSD TCP, including round-trip-time variance estimation, exponential retransmit backoff, slow-start, dynamic window sizing, Karn's clamped backoff, and fast retransmit. Most stem from one principle: a stable connection should conserve packets, injecting a new one only as an old one exits, so congestion control means fixing violations of that conservation.

Based on: Congestion avoidance and control · Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication

Curated by Aramai Editorial

Read summary →
Highlight

High-Dimensional Continuous Control Using Generalized Advantage Estimation

Introduces Generalized Advantage Estimation to cut policy-gradient variance, with trust-region optimization for stable high-dimensional continuous control.

Policy gradient methods directly optimize cumulative reward with neural network approximators, but need many samples and give unstable, nonstationary updates. This work cuts the variance of policy gradient estimates, at some bias, via value functions and an exponentially-weighted advantage estimator analogous to TD(lambda)—generalized advantage estimation—and stabilizes both policy and value networks with trust region optimization. The model-free method learns 3D locomotion, including biped and quadruped running gaits and biped stand-up, mapping kinematics to joint torques.

Based on: High-Dimensional Continuous Control Using Generalized Advantage Estimation · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation

Introduces DeepSDF, a learned continuous signed distance function representing a class of 3D shapes for quality representation, interpolation, and completion.

DeepSDF is a learned continuous Signed Distance Function representation for a class of 3D shapes, enabling high-quality representation, interpolation, and completion from partial, noisy input. Like a classical SDF, it encodes a surface as a continuous volumetric field where a point's magnitude gives distance to the surface and the sign marks inside versus outside—the surface being the zero-level-set. Unlike classical SDFs for one shape, DeepSDF represents an entire class, reaching state-of-the-art representation and completion while cutting model size by an order of magnitude.

Based on: DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

End to End Learning for Self-Driving Cars

Trains a CNN to map raw front-camera pixels directly to steering commands, learning to drive end-to-end from human steering as the only signal.

The authors train a convolutional neural network to map raw pixels from a single front-facing camera directly to steering commands. With minimal human training data, the end-to-end system learns to drive on roads with or without lane markings, on highways, and in areas with unclear guidance like parking lots. Using only the human steering angle as its training signal, it automatically learns internal representations such as road features, without explicit lane or path modeling, and runs at 30 fps.

Based on: End to End Learning for Self-Driving Cars · arXiv.org

Curated by Aramai Editorial

Read summary →