Linformer reduces transformer complexity to linear, improving efficiency.
problem High cost of training and deploying large transformer models for long sequences.
method Approximates self-attention with low-rank matrix, proposing Linformer with O(n) complexity. result Linformer performs similarly to standard transformers but is more memory- and time-efficient.
Active-memory mechanisms can replace self-attention in Transformers, but optimal results often require both.
problem Replacing self-attention with active-memory mechanisms in Transformers.
method Evaluation of various active-memory mechanisms in a Transformer model.
result Active-memory mechanisms can achieve comparable results to self-attention for language modeling, but optimal results are often achieved by combining both mechanisms.
Improves sequential recommendation with relation-aware self-attention.
problem Improving accuracy in sequential recommendation.
method Integrates Transformer's self-attention mechanism with a probabilistic model of recommendation context.
result Significant improvements over recent baseline models.
Self-attention prefers sparse functions of input sequences, reducing sample complexity.
problem Understanding the inductive biases of self-attention in modeling long-range dependencies.
method Theoretical analysis and synthetic experiments to probe sample complexity of learning sparse functions with Transformers.
result Bounded-norm Transformer networks can represent sparse functions of the input sequence with logarithmic sample complexity.
Proposes a faster Transformer decoding method by truncating target-side self-attention windows.
problem Efficiency in Transformer decoding with minimal BLEU score loss.
method N-gram assumption to truncate target-side self-attention windows.
result N-gram masked self-attention model maintains BLEU score for N values from 4 to 8. Improved hybrid acoustic model using interleaved self-attention and convolution.
problem Limited application of transformer in hybrid acoustic models.
method Proposed a model structure with interleaved self-attention and 1D convolution.
result Competitive recognition results on Librispeech dataset.
New model adds persistent memory to self-attention layers for improved performance.
problem Improving transformer performance by removing feed-forward layers.
method Augmenting self-attention layers with persistent memory vectors.
result The model outperforms standard transformers on language modeling benchmarks.
Transformers' self-attention mechanism is mapped to a generalized Potts model.
problem Uncertainty in what type of data distribution self-attention can efficiently learn.
method Decouple word positions and embeddings, then show self-attention learns a generalized Potts model.
result Training self-attention is equivalent to solving the inverse Potts problem.
New analysis shows how attention masks and LayerNorm prevent rank collapse in transformers.
problem Rank collapse in transformer models with increasing depth.
method General analysis of rank collapse under self-attention, considering attention masks and LayerNorm.
result Self-attention with LayerNorm can prevent rank collapse and maintain a rich set of equilibria.
Transformers can approximate any sequence-to-sequence function, surprising given their complexity.
problem Understanding the expressive power of Transformer models for sequence-to-sequence functions.
method Established that Transformers are universal approximators of continuous permutation equivariant sequence-to-sequence functions with compact support, and extended this to arbitrary functions using positional encodings.
result Transformers are universal approximators of arbitrary continuous sequence-to-sequence functions on a compact domain.
Paper investigates Lipschitz constants of self-attention modules in neural networks.
problem Lipschitz constants of self-attention modules in neural networks.
method Proved standard dot-product self-attention is not Lipschitz for unbounded input domain. Proposed L2 self-attention that is Lipschitz. Derived upper bound on L2 self-attention's Lipschitz constant.
result Proved standard self-attention is not Lipschitz for unbounded input domain and proposed an alternative L2 self-attention that is Lipschitz.
The p-Laplacian Transformer improves transformer models by assigning higher attention weights to tokens in close proximity.
problem The self-attention mechanism in transformers does not effectively distinguish attention weights between tokens in close and non-close proximity.
method Proposes a novel class of transformers, p-Laplacian Transformers, that use p-Laplacian regularization to assign higher attention weights to tokens in close proximity. result Empirically demonstrates that p-Laplacian Transformers outperform baseline transformers on various benchmark datasets.
Transformers become faster by linearizing self-attention.
problem Quadratic complexity of transformers makes them slow for long sequences.
method Expressed self-attention as a linear dot-product and used matrix product associativity to reduce complexity.
result Linear transformers are up to 4000x faster on long sequences.
Derives an empirical capacity model for self-attention neural networks.
problem Theoretical capacity of large transformer models is not fully utilized by current optimization algorithms.
method Analyzes memory capacity of transformers using synthetic training data and common training algorithms.
result Derives an empirical capacity model (ECM) for a generic transformer.
Skeinformer accelerates self-attention for long sequences with linear complexity.
problem Efficiency of Transformer models in processing long sequences.
method Matrix sketching and column sampling to reduce quadratic complexity to linear.
result Skeinformer outperforms alternatives with smaller time/space footprint.
Unified theory explains two failure modes of deep transformers and provides initialisation guidelines.
problem Two failure modes (rank collapse and entropy collapse) of self-attention layers in deep transformers.
method Analytical theory of signal propagation through deep transformers, using the Random Energy Model analogy.
result Simple algorithm to compute trainability diagrams for correct initialisation hyper-parameters.
Kernel PCA explains self-attention mechanisms in deep learning models.
problem Understanding and explaining self-attention mechanisms in deep learning models.
method Deriving self-attention from kernel principal component analysis (kernel PCA).
result RPC-Attention, a robust attention mechanism, outperforms softmax attention in various tasks.
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.
OSA overcomes instability in skipless Transformers.
problem Instability in skipless Transformers using Softmax Self-Attention.
method OSA parametrizes attention matrix to be orthogonal via skew-symmetric matrix exponential.
result OSA allows for training non-causal Transformers without skip connections and normalisation layers.
Skyformer uses Gaussian kernel and Nyström method to speed up self-attention in transformers.
problem High computational cost of self-attention in transformers.
method Replaces softmax with Gaussian kernel and applies Nyström method for matrix approximation.
result Skyformer achieves comparable or better performance with fewer computation resources.
I-BERT extends Transformer's self-attention to arbitrary input lengths.
problem Transformer models struggle with inductive generalization to unseen input lengths.
method Replaces positional encodings with a recurrent layer.
result I-BERT achieves state-of-the-art results on algorithmic tasks.
SE(3)-Transformers maintain equivariance for 3D data under rotations and translations.
problem Ensuring stable and predictable performance in 3D data under transformations.
method Introducing a self-attention module that is equivariant under continuous 3D roto-translations.
result The SE(3)-Transformer outperforms non-equivariant and non-attention models on real-world datasets.
Particles representing tokens cluster in Transformers, influenced by initial tokens and matrix spectrum.
problem Understanding the geometry of learned representations in Transformers.
method Viewing Transformers as particle systems, applying dynamical systems and partial differential equations.
result Particles cluster towards limiting objects, confirming context-awareness and the emergence of leaders.
Hrrformer uses HRR to speed up self-attention for long sequences.
problem Infeasibility of using transformers with very long sequence lengths.
method Re-cast self-attention using Holographic Reduced Representations (HRR).
result Achieves near state-of-the-art accuracy with O(THlogH) time complexity and O(TH) space complexity. Transformers can interpolate finite input sequences exactly.
problem Interpolating finite input sequences of arbitrary lengths.
method Constructing a transformer with alternating feed-forward and self-attention layers, and low-rank parameter matrices.
result Exact interpolation of datasets of finite input sequences in R^d with corresponding output sequences of smaller or equal length.
Mimetic initialization improves Transformer training on small datasets.
problem Difficulty in training Transformers on small datasets.
method Initialize self-attention layers to look like pre-trained models.
result Vanilla Transformers trained with mimetic initialization achieve higher accuracy.
Researchers improve transformer networks' optimization and understanding.
problem Improving the understanding and optimization of transformer networks.
method Introducing a convex alternative to the self-attention mechanism and reformulating the training problem as a convex optimization problem.
result Revealed an implicit regularization mechanism that promotes sparsity across tokens.
A study on optimizing self-attention in tabular data using Optimal Transport.
problem Improving efficiency and accuracy of self-attention in tabular classification tasks.
method Developed an OT-based algorithm to generate class-specific dummy Gaussian distributions and train an MLP.
result Achieved comparable accuracy to Transformers with reduced computational cost and efficiency.
Transformers predict price movements from limit order books.
problem Predicting price movements from limit order books.
method Causal convolutional network with masked self-attention.
result Significantly outperforms existing architectures on FI-2010 dataset.
Streaming ASR with transformer achieves low WER.
problem Real-time ASR with speech recognition.
method Time-restricted self-attention and triggered attention mechanisms.
result 2.8% and 7.2% WER for LibriSpeech test data.
Improved Transformer language models using dynamic evaluation.
problem Language model perplexity and accuracy improvements.
method Combining Transformers with dynamic evaluation techniques.
result Significant improvement in language model performance (e.g., 0.99 to 0.94 bits/char).
Transformer models outperform recurrent ones in modeling hierarchical data.
problem Modeling hierarchical structure in data.
method Introducing Multiresolution Transformer Networks leveraging self-attention.
result Multiresolution Transformer Networks significantly outperform state-of-the-art models on query suggestion datasets.
Efficient sparse attention reduces self-attention complexity and improves model performance.
problem Quadratic compute and memory requirements of self-attention for long sequences.
method Content-based sparse attention with dynamic routing module.
result Routing Transformer achieves state-of-the-art performance on various benchmarks.
Deep vanilla transformers trained without shortcuts achieve similar performance to standard models.
problem Training deep vanilla transformers without shortcuts and normalizations.
method Parameter initializations, bias matrices, and location-dependent rescaling.
result Deep vanilla transformers can train at similar speeds and performance to standard models.
Paper tackles time series forecasting weaknesses in Transformers.
problem Transformer's locality-agnostic and memory-bottleneck issues in time series forecasting.
method Proposes convolutional self-attention and LogSparse Transformer to address locality and memory issues.
result Improves forecasting accuracy for fine-granularity, long-term dependency time series.
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.
A framework for transformer attention layers derived from SVR.
problem Developing principled attention mechanisms for transformers.
method Mapping self-attention to SVR, deriving new attention types.
result Improved transformer performance and efficiency.
Develops first robustness verification for complex Transformers.
problem Certify prediction behavior of Transformers with complex self-attention layers.
method Resolves challenges of cross-nonlinearity and cross-position dependency in Transformers.
result Certified robustness bounds are significantly tighter than Interval Bound Propagation.
A new transformer model accelerates training with optimization techniques.
problem Training deep neural networks efficiently and effectively.
method Interprets transformer layers as optimization steps, applying Nesterov acceleration.
result The new model outperforms existing models on benchmark datasets.
ConViT combines CNN and ViT strengths, improving image classification.
problem Combining the strengths of CNNs and ViTs while avoiding their limitations.
method Introducing GPSA, a form of positional self-attention with a soft convolutional inductive bias.
result ConViT outperforms DeiT on ImageNet while offering improved sample efficiency.
Centroid Transformers reduce memory and computation by summarizing inputs into centroids.
problem Efficiently summarize inputs with reduced memory and computation.
method Generalizes self-attention to map N inputs to M centroids (M ≤ N), reducing complexity.
result Centroid Transformers reduce memory and computation while preserving key information.
Improved Transformer performance by addressing 'explaining away' effect.
problem Transformer's self-attention mechanism can explain away important input features.
method Proposed a doubly-normalized attention scheme to avoid 'explaining away' effect.
result Improved performance on benchmarks with the new attention scheme.
Transformers cluster meaningless words around leaders for sentiment analysis.
problem Capturing context in sentiment analysis using transformers.
method Characterized transformers with hardmax self-attention and normalization, showing asymptotic convergence to clustered equilibrium.
result Transformers can effectively capture context by clustering meaningless words around leader words.
Gated attention improves performance by using a hierarchical mixture of experts.
problem Improving performance of self-attention mechanisms in Transformers.
method Rigorously show that gated attention can be modeled as a hierarchical mixture of experts, providing a theoretical justification for its benefits.
result Gated attention is more sample-efficient than multi-head self-attention, requiring fewer data points to achieve the same estimation error.
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.
Transformers learn to perform logistic regression in-context.
problem Understanding how transformers learn to perform specific tasks in-context.
method Constructed multi-layer transformers that perform in-context logistic regression through normalized gradient descent.
result Transformers can be trained to perform in-context logistic regression effectively.
Softmax attention approximates complex functions and subsumes many known universal approximators.
problem Universal approximation of continuous sequence-to-sequence functions.
method Interpolation-based analysis of attention's internal mechanism, showing its ability to approximate ReLU functions.
result Softmax attention is a universal approximator for continuous sequence-to-sequence functions.
Transformers learn to generalize unseen tasks by composing self-attention layers.
problem How Transformers generalize to unseen, out-of-distribution tasks.
method Examined synthetic examples and pretrained LLMs, focusing on induction heads and latent subspace.
result Transformers can learn hidden rules for unseen tasks by composing self-attention layers, achieving OOD generalization.