Goldilocks activation functions improve neural network performance.
problem Improving neural network performance and understanding signal transformation.
method Introducing Goldilocks activation functions that locally deform input signals.
result Goldilocks networks outperform or match SELU and RELU on CIFAR-10 and CIFAR-100 datasets.
Convolutional DKMs improve kernel methods on MNIST, CIFAR-10, and CIFAR-100.
problem Improving kernel methods for image classification.
method Developed a novel inter-domain inducing point approximation and introduced various techniques to extend DKMs to convolutional networks.
result Achieved state-of-the-art performance on image classification benchmarks.
This paper examines how to calibrate ensemble members for better prediction accuracy.
problem Improper calibration of deep neural networks leads to unreliable probability estimates.
method Theoretical analysis and empirical evaluation on CIFAR-100 dataset.
result Well-calibrated ensemble members do not guarantee a well-calibrated ensemble prediction, but a well-calibrated ensemble prediction cannot exceed the average performance of its members.
CutMix training technique improves spatial locality in Vision Transformers.
problem Improving spatial locality in Vision Transformers trained from scratch.
method Comparison of Baseline and Modern training protocols on CIFAR-10, CIFAR-100, and Tiny-ImageNet.
result CutMix training component significantly reduces Mean Attention Distance (MAD) in early layers of Vision Transformers.
Unified view of contrastive and supervised learning improves model performance.
problem Improving model performance in classification, robustness, and detection.
method Hybrid discriminative-generative training of energy-based models.
result Improved performance on classification, robustness, out-of-distribution detection, and calibration.
Colored noise improves neural network robustness against adversarial attacks.
problem Vulnerability of neural networks to adversarial perturbations.
method Injection of colored noise into network weights and activations during adversarial training.
result Our approach outperforms previous methods in terms of adversarial accuracy on CIFAR-10 and CIFAR-100 datasets.
Improved sample efficiency with normalized RBF kernels in neural networks.
problem Learning more with less data in deep learning models.
method Two-phase method to train neural networks with normalized RBF kernels as output layer.
result Normalized RBF kernel networks achieve higher sample efficiency, compactness, and separability.
MALT improves adversarial attacks by targeting classes more efficiently.
problem Naive targeting of adversarial attacks based on classifier confidence.
method MALT - Mesoscopic Almost Linearity Targeting, based on medium-scale almost linearity assumptions.
result MALT wins over AutoAttack on CIFAR-100 and ImageNet datasets, five times faster.
SPAT improves adversarial robustness by preserving semantics in adversarial training.
problem Adversarial examples often have different semantics than original data, introducing unintended biases.
method Semantics-preserving adversarial training (SPAT) that encourages pixel perturbation shared among all classes.
result SPAT improves adversarial robustness and achieves state-of-the-art results in CIFAR-10 and CIFAR-100.
New method enhances adversarial robustness of deep learning models.
problem Improving the robustness of deep learning models against adversarial attacks.
method Optimal transport regularized divergences applied to distributionally robust optimization.
result Improved adversarial robustness on CIFAR-10 and CIFAR-100 datasets.
New bound on neural network generalization error using geometric complexity.
problem Understanding the generalization capabilities of deep neural networks.
method Derive a new upper bound on generalization error using margin-normalized geometric complexity.
result Empirical validation of the bound for ResNet-18 on CIFAR-10 and CIFAR-100 datasets.
This paper explores adversarial training limits and improves model robustness against norm-bounded perturbations.
problem Understanding and improving adversarial robustness of deep neural networks.
method Systematic study of adversarial training with various factors, including model size, activation functions, and unlabeled data.
result Training robust models that go beyond state-of-the-art results by combining larger models, Swish/SiLU activations, and model weight averaging.
We propose sequenced-replacement sampling (SRS) for training deep neural networks. The basic idea is to assign a fixed sequence index to each sample in the dataset. Once a mini-batch is randomly drawn in each training iteration, we refill the original dataset by successively adding samples according to their sequence i…
One of the ways to train deep neural networks effectively is to use residual connections. Residual connections can be classified as being either identity connections or bridge-connections with a reshaping convolution. Empirical observations on CIFAR-10 and CIFAR-100 datasets using a baseline Resnet model, with bridge-c…
Study improves top-k set prediction with low cardinality.
problem Improving top-k set prediction accuracy with low cardinality.
method Introduces new target loss function and surrogate losses.
result Demonstrates effectiveness of cardinality-aware algorithms.
Puzzle Mix optimizes mixup by leveraging saliency and local statistics.
problem Overfitting and adversarial vulnerability of deep networks.
method Puzzle Mix alternates between multi-label and saliency-based objectives for optimal mixing masks.
result Puzzle Mix outperforms other mixup methods in generalization and adversarial robustness.
Generative models improve adversarial robustness by adding synthetic data.
problem Improving robustness in machine learning models trained on limited data.
method Using synthetic data generated from a large dataset to augment the original training set.
result Generative models can significantly reduce the robust-accuracy gap compared to models trained with additional real data.
ThriftyNet uses a single convolutional layer recursively to maximize parameter usage.
problem Maximizing the use of parameters in deep convolutional neural networks.
method A single convolutional layer is used recursively, with normalization, non-linearities, downsampling, and shortcuts to maintain model expressivity.
result ThriftyNet achieves competitive performance with significantly fewer parameters.
TrustNet robustly learns noise patterns from trusted data to improve weakly-supervised classification.
problem Robustness to label noise in weakly-supervised learning.
method TrustNet learns noise patterns from trusted data, then trains a robust classifier using these patterns.
result TrustNet outperforms state-of-the-art methods in robustness to various noise patterns.
A new method uses optimal transport for semi-supervised classification.
problem Semi-supervised learning with limited labeled data.
method Optimal transport formulation with Sinkhorn iteration for label assignment.
result Improved performance on CIFAR-10, CIFAR-100, and SVHN datasets compared to FixMatch.
A simple technique improves continual learning by 50% on image datasets.
problem Challenges in training neural networks on a stream of shifting data.
method Experience Replay (ER) with five tricks to mitigate its shortcomings.
result ER, enhanced with tricks, achieves significant accuracy gains.
This paper investigates how forgetting affects neural network representations and stabilizes deeper layers.
problem Catastrophic forgetting in machine learning models trained on sequential tasks.
method Representational analysis techniques and empirical studies on CIFAR-10 and CIFAR-100 datasets.
result Deeper layers are disproportionately the source of forgetting, and methods to mitigate forgetting stabilize these layers.
Softmax is found ineffective for NL block, leading to improved performance.
problem Inefficiency of softmax in NL block for global context modeling.
method Empirical analysis and replacement of softmax with scaling factor.
result Improved performance on various datasets with reduced computational cost.
Proposes robust model through Wasserstein geodesic interpolation of training data.
problem Improving model robustness through data augmentation.
method Augment data by finding worst-case Wasserstein barycenter on geodesic path.
result Improves robustness on CIFAR-10 up to 7.7% and on CIFAR-100 up to 16.8%.
Residual networks (ResNets) have recently achieved state-of-the-art on challenging computer vision tasks. We introduce Resnet in Resnet (RiR): a deep dual-stream architecture that generalizes ResNets and standard CNNs and is easily implemented with no computational overhead. RiR consistently improves performance over R…
A new pruning method reduces neural network computation without retraining.
problem Efficiently reduce neural network computation while maintaining accuracy.
method Structured directional pruning via perturbation orthogonal projection.
result Achieves state-of-the-art pruned accuracy without retraining.
New method learns convolution-like structures from scratch.
problem Learning convolution-like structures from data.
method Minimum description length principle and β-LASSO algorithm. result Learned architectures achieve state-of-the-art accuracies.
T2FSNN improves deep SNNs by reducing spikes and latency.
problem Inefficiency in spiking neural networks due to lack of scalable training algorithms.
method Introduces time-to-first-spike coding with kernel-based dynamic threshold and dendrite, combined with gradient-based optimization and early firing methods.
result Reduces inference latency and spike count by 22% and less than 1% compared to burst coding.
A new method improves training speed and robustness for deep learning models.
problem Training deep learning models with heavy-tailed gradient noise.
method Per-sample gradient clipping combined with SGD.
result PS-Clip-SGD achieves optimal convergence rates and high-probability guarantees.
AANets balance stability and plasticity in CIL.
problem Stability-plasticity dilemma in class-incremental learning.
method Adaptive Aggregation Networks (AANets) with stable and plastic residual blocks.
result AANets improve performance on CIL benchmarks.
Study linear transformations' effects on data augmentation for improved estimation.
problem Improving performance in image and text classification tasks.
method Examined a family of linear transformations in over-parametrized linear regression settings.
result Transformations that preserve labels or mix data can improve estimation.
Improved VAE for heavy-tailed data using Student's t-distributions.
problem Over-regularization in VAEs with Gaussian priors.
method Proposed t3VAE framework with Student's t-distributions for prior, encoder, and decoder. result Significantly outperforms other models on heavy-tailed datasets.
This research tackles image classification with noise, proposing committees of CNNs.
problem Image classification with concurrent feature and label noise.
method Committees of Convolutional Neural Networks (CNNs) for MNIST, CIFAR-10, and CIFAR-100 datasets.
result Committees outperform single models in noisy conditions, especially on difficult datasets.
Regularising for invariance to data augmentation improves machine learning models.
problem Improving generalization in machine learning models through data augmentation.
method Explicit regularisation to encourage invariance at the level of individual model predictions.
result Explicit regularisation improves generalization and equalizes performance differences between objectives.
Paper introduces new loss functions for multi-class abstention learning.
problem Learning with multi-class classification and the ability to abstain.
method Developed new families of surrogate losses for abstention.
result Proved strong consistency guarantees for new surrogate losses.
Minimalistic unsupervised learning with sparse manifold transform achieves SOTA performance.
problem Achieving state-of-the-art unsupervised learning performance without complex engineering.
method Sparse manifold transform, leveraging sparse coding, manifold learning, and slow feature analysis.
result 99.3% KNN top-1 accuracy on MNIST, 81.1% on CIFAR-10, and 53.2% on CIFAR-100.
Meta-Semi learns to optimize SSL with minimal hyper-parameter tuning.
problem Limited labeled data in SSL makes it impractical to tune many hyper-parameters.
method Meta-learning approach that dynamically reweights unlabeled data loss.
result Meta-Semi achieves competitive performance on various SSL tasks.
Guided Evolution improves NAS efficiency and accuracy.
problem NAS methods converge to local minima and are complex.
method G-EA: guided evolutionary approach with initialization evaluation and continuous knowledge extraction.
result G-EA achieves state-of-the-art results in CIFAR-10, CIFAR-100, and ImageNet16-120.
The study shows removing fully connected output layers improves efficiency without sacrificing performance.
problem Large number of parameters in fully connected layers for high-category datasets.
method Examined architectures replacing fully connected output layers with fixed layers and compared performance.
result Fixed classifiers offer no additional benefit over removing the output layer and its parameters.
Artificial neural networks typically have a fixed, non-linear activation function at each neuron. We have designed a novel form of piecewise linear activation function that is learned independently for each neuron using gradient descent. With this adaptive activation function, we are able to improve upon deep neural ne…
Subset pretraining speeds up neural network training.
problem Efficiently training neural networks with partial gradients.
method Use small subsets to approximate the loss minimum of the full training set.
result Subset minima can approximate the full training set's loss minimum efficiently.
New findings challenge the use of flatness measures in neural networks.
problem The validity of flatness measures in assessing generalization in neural networks.
method Analysis of Hessian-based flatness norms and their relation to generalization.
result Solutions with large weights and low loss are often sharper than expected, contradicting flatness measures.
Improved robustness of machine learning models with controlled Lipschitz constants.
problem Vulnerability of state-of-the-art models to adversarial attacks.
method Proposes a CLL loss that calibrates the margin and Lipschitz constant penalties, improving robustness certificates.
result Consistently outperforms other losses on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets.
New distance metric for neural architecture search reduces search space complexity.
problem Reducing the complexity of neural architecture search.
method Fisher task distance for measuring task similarity and online neural architecture search.
result Reduced search space complexity for task-specific architectures.
A new mutual information optimization method using self-supervised binary contrastive learning.
problem Improving self-supervised contrastive learning for better model performance.
method Proposes a novel loss function for contrastive learning that optimizes mutual information in positive and negative pairs.
result The proposed method outperforms state-of-the-art self-supervised contrastive frameworks on various benchmark datasets.
BatchEnsemble reduces ensemble costs by 3X in training and testing.
problem High costs for training and testing ensembles of neural networks.
method Defines each weight matrix as a Hadamard product of a shared matrix and a rank-one matrix per member.
result Achieves 3X speedup and 3X memory reduction in test time for ensembles of size 4.
Paper proposes CCVR to improve classifier calibration in federated learning with non-IID data.
problem Improving classifier calibration in federated learning with non-IID data.
method Proposes CCVR, a simple algorithm that adjusts the classifier using virtual representations.
result CCVR achieves state-of-the-art performance on popular federated learning benchmarks.
This paper improves model robustness against adversarial attacks using optimal transport.
problem Adversarial attacks can mislead deep learning models with imperceptible perturbations.
method Exploits optimal transport theory to align adversarial and original image distributions.
result SAT (Sinkhorn Adversarial Training) leads to more robust models compared to state-of-the-art.