Efficiently accelerates attention calculation for Transformers with relative positional encoding.
problem Quadratic complexity of attention in long sequences.
method Kernelized attention with Fast Fourier Transform (FFT) for RPE.
result Achieves O(n log n) time complexity, mitigates training instability, and outperforms other models.
Randomized positional encodings boost transformer performance on longer sequences.
problem Transformers struggle with generalizing to sequences of arbitrary length.
method Introduced randomized positional encodings that simulate longer sequences and randomly select positions.
result Randomized positional encodings increase test accuracy by 12.0% on average for sequences of unseen length.
Paper introduces a new method for Transformers with linear complexity.
problem No efficient relative positional encoding for linear Transformer models.
method Stochastic Positional Encoding (SPE) that replaces classical RPE.
result SPE behaves like RPE and performs well on benchmarks.
ETC improves Transformer models for long and structured inputs.
problem Scaling input length and encoding structured inputs in Transformers.
method Introduces global-local attention, relative position encodings, and CPC pre-training.
result Achieves state-of-the-art results on four natural language datasets.
Data augmentations are important ingredients in the recipe for training robust neural networks, especially in computer vision. A fundamental question is whether neural network features encode data augmentation transformations. To answer this question, we introduce a systematic approach to investigate which layers of ne…
GTA improves transformer-based NVS models by encoding geometric structure.
problem Suboptimal positional encoding for 3D vision tasks.
method Geometry-aware attention mechanism encoding geometric structure of tokens.
result GTA improves learning efficiency and performance of NVS models.
Transformer adapts to graphs with adaptive attention and auto-regressive decoding.
problem Transformers struggle with graph data due to non-sequential nature.
method Proposes GRAT, a Transformer variant with adaptive attention and auto-regressive decoding.
result GRAT achieves state-of-the-art performance on molecule property predictions and generation tasks.
Transformer struggles with arithmetic length but improves with explicit structure encoding.
problem Transformers fail to generalize length in arithmetic tasks.
method Explicitly encoding structural symmetries via modified number formatting and custom positional encodings.
result Transformer can generalize up to 50-digit numbers without additional data.
New analysis shows PE in Transformers increases generalization gap and vulnerability.
problem Understanding the impact of PE on Transformer generalization and robustness.
method Generalization analysis and adversarial Rademacher bounds for a single-layer Transformer with trainable PE.
result PE systematically enlarges the generalization gap and makes models more vulnerable to attacks.
New method for Transformer models to encode position information without sequential bias.
problem Lack of flexible and learnable position encoding for Transformer models.
method Continuous dynamical model to learn position encoding.
result Consistent improvements over baselines in various NLP tasks.
TIMeSynC combines financial service interactions for intent prediction.
problem Aligning and learning from multi-domain, multi-resolution sequences for accurate intent prediction.
method An encoder-decoder transformer model addressing sequence alignment, temporal dynamics, and dynamic/static sequence combination.
result Significant improvement in intent prediction over existing methods.
Generative Adversarial Nets (GANs) and Variational Auto-Encoders (VAEs) provide impressive image generations from Gaussian white noise, but the underlying mathematics are not well understood. We compute deep convolutional network generators by inverting a fixed embedding operator. Therefore, they do not require to be o…
STRING improves 2D and 3D position encodings for better performance.
problem Efficient and accurate position encoding for 2D and 3D applications.
method STRING extends Rotary Position Encodings with a unifying theoretical framework, maintaining translation invariance and low computational cost.
result STRING shows substantial gains in open-vocabulary object detection and robotics.
New auto-encoder handles varying noise levels without retraining.
problem Auto-encoders degrade in noisy conditions.
method Formalized auto-encoders as transform learning, derived new architecture.
result Models generalize well to different noise levels.
WeatherFormer learns robust weather features from small datasets.
problem Modeling complex weather dynamics from limited data.
method Pretrained transformer encoder on large satellite dataset, with spatiotemporal encoding.
result State-of-the-art performance in county-level soybean yield prediction and influenza forecasting.
Encoder-decoder based sequence-to-sequence models have demonstrated state-of-the-art results in end-to-end automatic speech recognition (ASR). Recently, the transformer architecture, which uses self-attention to model temporal context information, has been shown to achieve significantly lower word error rates (WERs) co…
GATES improves neural architecture search by modeling operations as information transformation.
problem Improving predictor-based neural architecture search efficiency.
method GATES models operations as information transformation, covering both node and edge cell search spaces.
result GATES boosts sample efficiency and improves predictor performance.
CT improves neural network performance on cell complex data.
problem Improving predictive performance of neural networks on complex data.
method Introducing the Cellular Transformer (CT) that generalizes graph-based transformers to cell complexes.
result CT achieves state-of-the-art performance on cell complex datasets without complex enhancements.
Transformer learns graph structure better with subgraph info.
problem Transformer struggles with structural similarity in graph learning.
method Structure-Aware Transformer with subgraph attention.
result Improves graph prediction benchmarks significantly.
Transformers learn causal structure through gradient descent on self-attention mechanisms.
problem Understanding how transformers learn causal structure during training.
method In-context learning task and simplified two-layer transformer model.
result Gradient descent on a simplified transformer learns to encode latent causal graphs.
Bi-linear feature learning models, like the gated autoencoder, were proposed as a way to model relationships between frames in a video. By minimizing reconstruction error of one frame, given the previous frame, these models learn "mapping units" that encode the transformations inherent in a sequence, and thereby learn …
Financial fraud detection in digital banking requires reasoning over multiple heterogeneous event streams.
problem Financial fraud detection in digital banking requires reasoning over multiple heterogeneous event streams.
method Multi-Stream Fraud Transformer (MSFT) architecture that encodes each event stream with independent Transformer encoders and fuses their representations through configurable mechanisms.
result Sequence models significantly outperform gradient-boosted trees operating on aggregated features.
Self-supervised VAEs improve data compression and generation.
problem Efficient data compression and generation.
method Introducing self-supervised Variational Auto-Encoders with deterministic and discrete variational posteriors.
result Self-supervised VAEs simplify the objective function and improve data reconstruction.
New techniques improve channel prediction in noisy wireless systems.
problem Predicting channels in wireless communication systems from noisy observations.
method Adapted sequence-to-sequence models and transformers with reverse positional encoding and reversed encoder outputs.
result Improved robustness and relationship capture in channel prediction models.
We incorporate Tensor-Product Representations within the Transformer in order to better support the explicit representation of relation structure. Our Tensor-Product Transformer (TP-Transformer) sets a new state of the art on the recently-introduced Mathematics Dataset containing 56 categories of free-form math word-pr…
Sparse coding is a common approach to learning local features for object recognition. Recently, there has been an increasing interest in learning features from spatio-temporal, binocular, or other multi-observation data, where the goal is to encode the relationship between images rather than the content of a single ima…
Bayesian quadrature improves integration efficiency with invariant priors.
problem Efficient numerical integration with known structure.
method Invariance priors for bijective transformations in input domain.
result Superior performance in synthetic and real-world applications.
OLinear forecasts time series more efficiently by transforming data orthogonally.
problem Efficiently forecasting time series with entangled dependencies.
method OLinear uses OrthoTrans to transform data orthogonally, then applies NormLin for linear layer.
result OLinear achieves state-of-the-art performance with high efficiency.
New analysis shows RPE-based Transformers can't approximate all functions.
problem Understanding the limitations of RPE-based Transformers in approximating continuous functions.
method Mathematical analysis and development of a novel attention module (URPE) to overcome limitations.
result RPE-based Transformers can't approximate all continuous sequence-to-sequence functions, even with depth and width.
HFformer outperforms LSTM in high-frequency trading with multiple signals.
problem Improving high-frequency trading performance using deep learning models.
method Introducing HFformer, a hybrid Transformer model for time series forecasting.
result HFformer achieves higher cumulative PnL than LSTM in backtesting.
Sparse Transformers degrade semantic information first, with early layers encoding more.
problem Understanding how sparse Transformers affect learned representations and semantic information.
method Probed Transformers with progressively pruned weights to observe changes in semantic information and model behavior.
result Complex semantic information is first to degrade in sparse Transformers, with early layers encoding more.
The complexity of a learning task is increased by transformations in the input space that preserve class identity. Visual object recognition for example is affected by changes in viewpoint, scale, illumination or planar transformations. While drastically altering the visual appearance, these changes are orthogonal to r…
Disentangled representation learning finds compact, independent and easy-to-interpret factors of the data. Learning such has been shown to require an inductive bias, which we explicitly encode in a generative model of images. Specifically, we propose a model with two latent spaces: one that represents spatial transform…
Study presents a method to induce a generalized neural network from joint group invariant functions.
problem Encoding rule of neural network internal data representation.
method Systematic method using joint group invariant function on data-parameter domain.
result Induces a generalized neural network and its inverse operator (ridgelet transform).
Researchers prove Transformers are Turing-complete and analyze their components.
problem Understanding the computational power and limitations of Transformers.
method Analyzed Turing-completeness of vanilla and modified Transformers, and necessity of components.
result Transformers with positional masking and positional encodings are Turing-complete.
This paper introduces a new task to better understand Transformers in quantitative contexts.
problem Understanding Transformers in high-stakes quantitative and scientific applications.
method Introduces a novel contextual counting task and analyzes it with causal and non-causal Transformer architectures.
result Causal attention is better suited for the contextual counting task, and no positional embeddings lead to the best accuracy.
The key idea of current deep learning methods for dense prediction is to apply a model on a regular patch centered on each pixel to make pixel-wise predictions. These methods are limited in the sense that the patches are determined by network architecture instead of learned from data. In this work, we propose the dense…
Deep learning transforms time series into images for anomaly detection in industrial assets.
problem Detecting anomalies in time series data from industrial assets.
method Transforming time series data into image-like representations and using them as inputs for deep learning models.
result Some encodings provide competitive results for anomaly detection in industrial asset monitoring.
Improved graph embedding through refined linear transformation and community recovery.
problem Identifying meaningful latent communities in graph data.
method Refined graph encoder embedding via linear transformation, self-training, and latent community recovery.
result Improved vertex embedding and better decision boundaries for vertex classification.
SplitWise enhances stepwise regression by adaptively encoding numeric predictors into binary features.
problem Capturing nonlinear relationships in regression models without sacrificing interpretability.
method Adaptive encoding of numeric predictors into binary features using shallow decision trees, assessed by AIC or BIC.
result Consistently produces more parsimonious and generalizable models than traditional techniques.
Transformer model improves source code summarization.
problem Generating readable summaries of source code.
method Transformer model with self-attention mechanism for code representation.
result Transformer model outperforms state-of-the-art techniques.
We define a twistor-like transform of the equations of eleven-dimensional supergravity. More precisely these equations are encoded by the CR-structure on the twistor space P^{2*15+11|8*2+16}. In addition equations of the linearized eleven-dimensional supergravity adapted to the 3-form potential can be transformed into …
A transformer model improves spell correction with hierarchical attention.
problem Improving spell correction accuracy and speed.
method Multi encoder-single decoder transformer architecture with hierarchical attention.
result Significant improvement in CER, WER, and SER error rates.
Data augmentation (DA) is fundamental against overfitting in large convolutional neural networks, especially with a limited training dataset. In images, DA is usually based on heuristic transformations, like geometric or color transformations. Instead of using predefined transformations, our work learns data augmentati…
TiDE uses MLP for fast, simple long-term time-series forecasting.
problem Long-term time-series forecasting challenges.
method Time-series Dense Encoder (TiDE) based on MLP.
result TiDE matches or outperforms Transformer models while being 5-10x faster.
Improves contrastive learning invariance with novel training objectives and feature averaging.
problem Contrastive learning's implicit invariance is insufficient for robust performance.
method Introduces a novel training objective and feature averaging approach to enforce invariance.
result Improved performance and robustness to transformations on downstream tasks.
In this work, we study abstractive text summarization by exploring different models such as LSTM-encoder-decoder with attention, pointer-generator networks, coverage mechanisms, and transformers. Upon extensive and careful hyperparameter tuning we compare the proposed architectures against each other for the abstractiv…
Paper studies Transformer learning theory for Euclidean and Riemannian domains.
problem Understanding and optimizing Transformer networks for regression tasks.
method Constructive approximation framework using softmax partition of unity and attention mechanism.
result Transformer can achieve uniform ε-approximation error with minimal parameters.