Detects misclassifications and adversarial examples using neural network logits.
problem Unable to detect misclassifications and adversarial examples in neural networks.
method Introspection using pretrained neural network logits.
result Simple 3-layer neural network trained on logits detects misclassifications competitively.
Neural Photo Editor uses AI to edit photos accurately and coherently.
problem Achieving accurate reconstructions in photo editing.
method Introspective Adversarial Network (VAE-GAN hybrid) with weight-shared dilated convolutions and Orthogonal Regularization.
result Produces high-fidelity photo edits and reconstructions.
Introspection Learning asks policies introspective questions to improve learning.
problem Improving reinforcement learning efficiency and robustness.
method Synthesizes experience through direct policy introspection without interaction.
result Demonstrates improved training speed and robustness.
IntroVAC learns interpretable latent subspaces for better image quality.
problem Difficulties in interpreting latent spaces and limitations in image generation.
method Introspective Variational Classifier (IntroVAC) using additional labels and adversarial training.
result Improved image quality and meaningful latent directions for fine-grained manipulation.
Proposes a technique to interpret deep learning models by generating counterfactual inputs.
problem Understanding and explaining the decisions made by deep neural networks.
method Uses a generative model to edit input images and generate counterfactual scenarios for model interpretation.
result Demonstrates the effectiveness of the introspection approach on MNIST and CelebA datasets.
Introspects convolutional speech recognition models using Gradient-adjusted Neuron Activation Profiles.
problem Lack of interpretability in deep learning ASR models.
method Gradient-adjusted Neuron Activation Profiles (GradNAPs) for feature and representation visualization.
result Gains insight into how data is processed in convolutional ASR models.
ISP improves DNN uncertainty for better subgroup accuracy.
problem Improving accuracy-group robustness in deep neural networks.
method Introspective Self-play (ISP) adds an introspection task to improve bias-awareness and uncertainty.
result ISP improves the accuracy-group robustness trade-off of AL methods.
The paper improves CNN visualization for safer reinforcement learning.
problem Safety risks in reinforcement learning control systems.
method Extensions of CNN visualization algorithms for reinforcement learning.
result Visualization aids in understanding and improving reinforcement learning policies.
This paper explores neural networks for colorizing grayscale images.
problem Colorizing grayscale images using neural networks.
method Comparison of existing and novel generative models (CVAE, CWGAN-GP, AGE, IVAE) trained on CIFAR-10 images.
result CVAE with L1 reconstruction loss and IVAE achieve the highest Inception Score (IS).
RNN with attention detects system anomalies with high accuracy.
problem Lack of interpretability in deep learning models for system log anomaly detection.
method Recurrent Neural Network (RNN) with attention mechanisms.
result RNN with attention achieves high performance in anomaly detection.
IntroVAE synthesizes high-quality photos by self-evaluating and improving its outputs.
problem Generating high-quality photographic images with stable and realistic results.
method Introspective Variational Autoencoder (IntroVAE) that trains inference and generator models jointly, encouraging the inference model to distinguish between generated and real samples.
result Produces high-resolution photo-realistic images comparable to or better than state-of-the-art GANs.
RPO uses past and future state-action info for better policy optimization.
problem Sample inefficiency in on-policy reinforcement learning methods.
method Reflective Policy Optimization (RPO) integrates past and future state-action info for policy improvement.
result RPO improves policy performance and contracts the solution space, leading to faster convergence.
This work improves motion planning for quadcopters by learning and reasoning about controller performance.
problem Improving motion planning for quadcopters with safety margins and execution reliability.
method Introspective learning and reasoning to correct execution bias and improve collision checking.
result Substantial reduction in safety margins for motion actions, leading to safer execution.
DSS networks use scale-equivariant cross-correlations to improve image recognition.
problem Improving image recognition by exploiting scale invariance.
method Constructing scale-equivariant cross-correlations based on scale-spaces and semigroups.
result Demonstrated utility on Patch Camelyon and Cityscapes datasets.
Modular networks learn to decompose neural computation efficiently.
problem Efficiently scaling model capacity in deep learning.
method End-to-end learning of neural modules and their decomposition based on data.
result Superior performance in image recognition and language modeling tasks.
This paper explores intrinsic rewards to improve learning from multiple value functions.
problem How to adapt reinforcement learning systems to optimize learning from multiple value functions.
method Investigated and compared 14 different intrinsic reward mechanisms in a new bandit-like parallel-learning testbed.
result Intrinsic rewards based on the amount of learning can generate useful behavior, if each individual learner is introspective.
Transfer learning improves ASR on limited resources.
problem End-to-end ASR training requires massive data and compute.
method Model adaptation using transfer learning.
result ASR models trained on limited resources achieve similar accuracy.
VSML unifies meta learning concepts and enables simple backpropagation.
problem Improving and unifying meta learning concepts for neural networks.
method Unified approach using variable shared meta learning and simple weight-sharing.
result Simple backpropagation can be implemented and meta learned without gradient calculation.
Method uses network biology to construct gene expression models for cancer.
problem Building models for cancer phenotypes using gene expression data.
method Unsupervised construction of computational graphs based on protein-protein networks.
result The method outperforms other models in cancer phenotype analysis.
Semantic embeddings improve safety-critical classifier performance.
problem Improving interpretability and error detection in safety-critical neural networks.
method Created embeddings from symbolic domain knowledge, used for misprediction interpretation and error detection, introduced semantic distance for confidence measurement.
result Semantic distance achieves near state-of-the-art performance in a traffic sign classifier, faster than other methods.
InterpNET generates natural language explanations for deep learning classifications.
problem Deep learning models lack explainability.
method Inspired by human visual system, InterpNET designs interpretable neural networks.
result InterpNET achieves a high METEOR score of 37.9 for generating explanations.
This review synthesizes uncertainty modeling in probabilistic image segmentation.
problem Relaxed Bayesian assumptions lead to missing uncertainty information in deep models.
method Standardizes theory, notation, and terminology for feature- and parameter-distribution modeling.
result Establishes a common framework for robust decision-making in segmentation tasks.
SoftAdapt dynamically adjusts loss weights for multi-part functions.
problem Slow convergence and poor weight selection for multi-part loss functions.
method SoftAdapt dynamically changes weights based on live performance statistics.
result Improved convergence and better weight selection for multi-part loss functions.
Generative Adversarial Trainer improves neural network robustness to adversarial perturbations.
problem Improving neural network robustness to adversarial perturbations.
method Alternately trains a classifier and generator network to generate and classify adversarial images.
result Significantly lowers the generalization error of the network.
Aims to make neural networks robust to small, imperceptible perturbations.
problem Neural networks can be fooled by small, imperceptible changes.
method Proposes a Generative Adversarial Networks (GAN) framework to learn robust classifiers.
result Our method outperforms adversarial training in robustness.
Capsule networks are vulnerable to adversarial attacks, similar to convolutional neural networks.
problem Vulnerability of capsule networks to adversarial attacks.
method Compared capsule networks to convolutional neural networks using various adversarial attacks.
result Capsule networks are vulnerable to adversarial attacks, similar to convolutional neural networks.
Compact Gaussian model approximates deep ensemble predictions.
problem Efficiently approximating deep ensemble models for image prediction.
method Sparse-structured multivariate Gaussian with Cholesky parameterization trained to match pre-trained ensemble outputs.
result Compact representation captures uncertainty and structured correlations explicitly.
Adversarial training adds dynamic perturbations to neural networks for robustness.
problem Accuracy trade-off and lack of diversity in adversarial examples.
method Dynamic adversarial perturbations in the parameter space of neural networks, updating perturbation biases during training.
result Adversarial training with negligible cost and reduced accuracy trade-off.
New method makes neural networks more resistant to adversarial attacks.
problem Neural networks are vulnerable to adversarial examples.
method Regulates adversarial gradients to increase robustness.
result Generated networks are near-immune to various adversarial attacks.
This paper investigates how network width and depth affect adversarially robust DNNs.
problem Understanding architectural configurations for adversarially robust DNNs.
method Comprehensive investigation on the impact of network width and depth on adversarial robustness.
result Optimal architectural configuration for adversarial robustness exists and can improve robustness.
Vizarel aims to help RL agents by visualizing their performance.
problem Existing visualization tools are not suitable for reinforcement learning.
method Designing a prototype system for interpreting RL agent performance.
result Prototype system identified features for visualizing RL agent performance.
New neural network units resist adversarial attacks effectively.
problem Adversarial attacks on machine learning models.
method Introduced MWD units, developed training techniques, and computed robustness.
result MWD networks are significantly more robust to adversarial attacks.
Study analyzes adversarial training dynamics without data distribution assumptions.
problem Understanding training dynamics of adversarial training without data distribution assumptions.
method Mean field theory approach to analyze adversarial training in random deep neural networks.
result Upper bounds of adversarial loss derived empirically and theoretically.
Improves deep neural networks' robustness against adversarial attacks.
problem Deep neural networks' vulnerability to adversarial perturbations.
method Joint training with a discriminator to filter adversarial noise.
result Preliminary experiments show improved adversarial robustness.
This paper analyzes the limits of neural networks under adversarial attacks.
problem Understanding how neural networks perform under adversarial noise.
method Defined the adversarial VC-dimension and combined it with the VC-dimension of halfspace classifiers.
result Concluded the adversarial VC-dimension of neural networks with sign activation functions.
New neural network units resist adversarial attacks effectively.
problem Adversarial attacks on neural networks that misclassify inputs.
method Introduced RBFI units with non-linear structure.
result RBFI units maintain high accuracy in adversarial attacks.
BPN defends against adversarial attacks by generating beneficial perturbations.
problem Adversarial attacks cause deep neural networks to misclassify clean inputs.
method BPN generates beneficial perturbations during training to neutralize future adversarial attacks.
result BPN is robust to adversarial examples and more efficient than classical adversarial training.
NTK neural networks are robust to adversarial attacks in nonparametric regression.
problem Adversarial robustness of neural networks in nonparametric regression.
method Gradient flow with early stopping for NTK neural networks, proving robustness in Sobolev spaces.
result NTK neural networks achieve optimal adversarial robustness rates in Sobolev spaces.
New method improves neural network robustness without adversarial training.
problem Adversarial robustness of neural networks under flat loss surface.
method Visualizing decision surfaces in input space to assess robustness.
result Decision surface geometry in input space correlates with adversarial robustness.
Comment on recent paper suggesting adversarially trained BNNs are more robust, finds no strong evidence.
problem Evaluating the robustness of adversarially trained Bayesian Neural Networks (BNNs) against adversarial attacks.
method Developed a new type of adversarial attack to incorporate the stochastic nature of BNNs, used to evaluate robustness.
result No strong evidence of higher robustness of adversarially trained BNNs.
This work simplifies adversarial attacks using neural networks, reducing computation and improving training convergence.
problem Efficiently generating and training against ideal adversarial attacks with minimal computational overhead.
method Representing ideal adversarial attacks as smooth piece-wise functions and approximating them with neural networks. Using a mathematical game between an attack network and a defense network for adversarial training.
result Obtained convergence rates of adversarial loss in terms of sample size n for adversarial training. Adaptive networks improve model robustness through conditional normalization.
problem Limited robustness of adversarial-trained networks due to network capacity and training samples.
method Proposes a conditional normalization module to adapt networks during adversarial training.
result Adaptive networks outperform both clean validation accuracy and robustness compared to non-adaptive counterparts.
A novel method to train networks from each other's adversarial examples to resist black-box attacks.
problem Machine learning models can be fooled by adversarial examples, especially in black-box attacks.
method Simultaneous adversarial training combining two networks to learn from each other's adversarial examples.
result The method improves the networks' resilience to black-box attacks.
The paper examines how neural network topology affects adversarial robustness.
problem Understanding how neural network topology influences adversarial robustness.
method Investigated the graph of input traversing all layers of a neural network, comparing clean and adversarial inputs.
result Under-optimized edges in neural network graphs are a source of adversarial vulnerability and can be used to detect adversarial inputs.
New method improves neural network interpretability against adversarial attacks.
problem Adversarial attacks can hide from neural network interpretability methods.
method Develops an interpretability-aware defensive scheme promoting robust interpretation.
result Achieves both robust classification and robust interpretation.
Adversarial training achieves optimal test error for shallow networks.
problem Achieving optimal adversarial test error for general data distributions.
method Applying new Rademacher complexity bounds and properties of optimal adversarial predictors.
result Adversarial training can achieve optimal adversarial test error for general data distributions.
Adversaries reprogram text classification models without changing the original network.
problem Reprogramming neural networks trained on discrete input spaces like text classification.
method Context-based vocabulary remapping model for white-box and black-box settings.
result Successfully repurposed various text-classification models for new tasks.
Biologically inspired neural networks resist adversarial attacks.
problem Resisting adversarial attacks on deep neural networks.
method Developed a scheme based on biophysical principles of neural circuits to train deep networks to be robust to adversarial attacks.
result Generated deep neural networks achieve state-of-the-art performance on adversarial examples without exposure during training.