A pooling method that uses the spectrum of the feature maps to perform dimensionality reduction, helping to maintain more information than standard pooling.

A method to make the neural networks sparse by applying a special form of dropout and tuning the dropout rates using the variational inference framework.

ZSL is a method for recognizing unseen classes during the training phase. This is useful when there are no training examples for the category we want to recognize.

This model is specifically trained for scientific papers and takes into account citation information to better capture paper semantics.

SINDy is a method that uses sparse regression to identify the governing equations of a dynamical system only from measurement data.

These are transformer models that exploit the fact that attention matrices are often nearly low-rank (that is, attention can be modeled well as depending on only a few dimensions).

A weight normalization technique that stabilizes the training of discriminators (critics) in Generative Adversarial Networks (GANs).

Unlike traditional CNNs that require a fixed-size input, SPPnets can process images of any size and aspect ratio, providing a more flexible framework for object recognition tasks. It uses a spatial pyramid pooling layer to eliminate the fixed-size constraint by allowing pooling in arbitrary scales and locations.

A type of CNN that operates on spherical signals, making it suitable for tasks like 3D model recognition or climate forecasting.

These models introduce a new building block for CNNs that improves channel interdependencies at almost no computational cost. They adaptively recalibrate channel-wise feature responses by explicitly modeling interdependencies between channels.

A Bayesian variable selection approach where the posterior distributions of the coefficients of irrelevant variables tend to spike at zero, while the posteriors of the coefficients of relevant variables have a wider distribution (the slab).

A convolutional neural network that is used for LiDAR point cloud segmentation, specifically developed for road-object segmentation in autonomous driving.

These networks more closely mimic biological neural networks by incorporating the concept of time into their operating model. The network communicates by means of 'spikes', which are discrete events that take place at points in time, making them great for temporal data.

A variant of ResNet that introduces a split-attention mechanism, enhancing the representational capacity of the network.

A variant of principal component analysis (PCA) that involves a sparsity constraint on the loading vectors of the PCA.

SqueezeBERT is a transformer-based model designed for mobile devices. It optimizes the standard transformer architecture for speed and memory efficiency while maintaining competitive performance on NLP tasks.

Sparse autoencoders are a type of autoencoder designed to be sensitive to specific types of complex features in the input data, while ignoring others.

A specialized version of BERT for spoken language understanding. It's trained on transcriptions of spoken language and is used for tasks like intent recognition or named entity recognition in spoken language.

Techniques that make use of the spectrum (eigenvalues) of the similarity matrix of the data to perform dimensionality reduction before clustering in fewer dimensions.

This is a variant of Dropout that drops entire 1D feature maps instead of individual elements in convolutional neural networks.