New method interprets neural agent's actions to prevent unwanted outcomes.
problem Blackbox issue in RL where agents learn without foreseeing all outcomes.
method Action-conditional β-VAE for disentangled representation learning.
result Interpretable latent features enable modeling entire state space.
Optimizes neural networks with blackbox solvers using Time-cost Regularization.
problem Improving neural network performance by integrating efficient solvers for complex problems.
method Optimizes both the primary loss function and the performance of the blackbox solver using Time-cost Regularization. Introduces a hyper-blackbox concept to learn blackbox parameters.
result Significant improvement in neural network performance through optimization of blackbox solvers.
CoNES optimizes blackbox functions using convex optimization and information geometry.
problem Optimizing high-dimensional blackbox functions efficiently.
method Formulated as a convex program that adapts evolutionary strategies gradient estimates.
result Vastly outperforms conventional blackbox optimization methods on benchmarks and MuJoCo tasks.
ASEBO optimizes complex functions by learning optimal sensing directions.
problem Optimizing high-dimensional blackbox functions with expensive queries.
method Adapts to function geometry, learns sensing directions, uses active subspaces.
result More sample-efficient than state-of-the-art algorithms.
New method improves blackbox attack transferability by perturbing feature hierarchy.
problem Improving transferability of blackbox attacks across different models and datasets.
method Perturbs representations throughout feature hierarchy to mimic other classes.
result Achieves 10x increase in targeted success rate compared to other methods.
Proposes a query-efficient blackbox attack method.
problem Adversarial attacks on machine learning models, especially deep neural networks.
method QEBA: Query-Efficient Boundary-based blackbox Attack, using only final prediction labels.
result QEBA achieves 100% attack success rate with fewer queries and lower perturbation.
Counterfactual explanations are unreliable in many contexts.
problem Reliability issues with counterfactual explanations.
method Discussing three desirable properties: proximity, connectedness, and stability.
result Post-hoc counterfactual approaches often fail to satisfy these properties.
We optimize rank-based metrics using blackbox differentiation.
problem Challenges in directly optimizing rank-based metrics due to their non-differentiable and non-decomposable nature.
method Efficient, theoretically sound, and general method for differentiating rank-based metrics with mini-batch gradient descent.
result Competitive performance on standard image retrieval datasets and improved performance on object detectors.
Method uses neural networks to solve combinatorial problems.
problem Solving combinatorial problems on raw input data.
method Integrates blackbox combinatorial solvers into neural networks.
result Efficient backward pass through blackbox solvers implemented.
Novel method for scalable neural network-based blackbox optimization.
problem Scalability challenges in high-dimensional Bayesian Optimization.
method SNBO: Adds new samples using separate criteria for exploration and exploitation, adaptively controlling the sampling region.
result SNBO achieves better function values with 40-60% fewer function evaluations and reduced runtime.
The paper explores methods to explain decisions of deep learning models by faithfully reproducing their training data views.
problem Explaining decisions of complex deep learning models trained on large datasets.
method Data view extraction through hill-climbing and GAN-driven approaches, followed by creation of shadow models for explanation.
result Shadow models based on faithfully reproduced data views are effective for explaining decisions of blackbox deep learning models.
EGORSE optimizes high-dimensional problems using random and supervised embeddings.
problem Efficiently solving computationally expensive high-dimensional optimization problems.
method EGORSE combines random and supervised linear embeddings for adaptive optimization.
result EGORSE outperforms state-of-the-art methods in high-dimensional optimization.
InterpretML simplifies machine learning interpretability for users and researchers.
problem Making machine learning models understandable to non-experts.
method Unified Python package exposing interpretability algorithms and visualization.
result First implementation of Explainable Boosting Machine, a powerful, interpretable model.
Open-source Vizier optimizes complex systems for Google and beyond.
problem Optimizing large-scale systems with multiple objectives and constraints.
method Distributed, fault-tolerant, flexible API for blackbox optimization.
result OSS Vizier supports a wide range of optimization problems and is available as open-source.
PyXAB is a Python library for X-armed bandits and online optimization.
problem Efficiently solving X-armed bandit problems and online blackbox optimization.
method Implementation of 10+ X-armed bandit algorithms and synthetic objectives.
result Evaluation of different algorithms' performance on various synthetic objectives.
ControlSHAP stabilizes Shapley value approximations using control variates.
problem High computational cost of exact Shapley values in blackbox models.
method ControlSHAP uses Monte Carlo control variates to stabilize Shapley value approximations.
result Significant reduction in Monte Carlo variability of Shapley estimates.
Improved stability and generalization for blackbox learned optimizers.
problem Stability and generalization issues in blackbox learned optimizers.
method Investigation using dynamical systems, modifications to optimizer architecture and meta-training procedure.
result Improved stability and generalization of learned optimizers.
No free lunch theorems show all algorithms perform equally under uniform distribution.
problem Analyzing scenarios involving non-uniform distributions and comparing algorithms.
method No Free Lunch theorems applied to analyze and compare algorithms without distribution assumptions.
result Anti-cross-validation performs as well as cross-validation under non-uniform distributions.
End-to-end trainable graph matching using improved combinatorial solvers.
problem Graph matching in deep learning.
method Combining deep learning with optimized combinatorial solvers.
result Advances state-of-the-art on deep graph matching benchmarks.
New sampling method reduces variance in correlated high-dimensional distributions.
problem Reducing variance in Monte Carlo estimators for correlated high-dimensional distributions.
method DPPMC (Determinantal Point Processes Monte Carlo) method for structured sampling.
result DPPMCs improve state-of-the-art in various optimization and machine learning problems.
Despite advances in scalable models, the inference tools used for Gaussian processes (GPs) have yet to fully capitalize on developments in computing hardware. We present an efficient and general approach to GP inference based on Blackbox Matrix-Matrix multiplication (BBMM). BBMM inference uses a modified batched versio…
Adversarial perturbations fool deepfake detectors with high accuracy.
problem Improving deepfake detection accuracy against adversarial attacks.
method Used adversarial perturbations and two defenses: Lipschitz regularization and Deep Image Prior (DIP).
result Deepfake detectors achieved 27% accuracy on perturbed images, compared to 95% on unperturbed.
Private learning of Gaussian Mixture Models without boundedness assumptions.
problem Private estimation of parameters of Gaussian Mixture Models with unbounded components.
method Reduction to non-private problem, blackbox privatization, Moitra and Valiant's algorithm.
result First sample complexity upper bound and polynomial time algorithm for privately learning GMMs.
New algorithm RBO improves RL in noisy environments.
problem Derivative-free optimization sensitivity to noisy rewards and dynamics.
method Robust Blackbox Optimization (RBO) using gradient flows and robust regression.
result RBO can recover gradients to high accuracy even with up to 23% corrupted measurements.
CALLISTO generates tests and assesses ML data quality using prediction entropy.
problem Validating ML systems for accuracy and data quality.
method Entropy-based test generation and data quality assessment framework.
result CALLISTO detects up to 20x more errors than traditional methods.
IETNet identifies important channels for MVTS classification.
problem Multivariate time series classification with blackbox deep networks.
method End-to-end network combining temporal feature extraction, variable selection, and interaction.
result IETNet improves model accuracy and reduces overfitting by identifying and removing non-predictive variables.
Framework ensures alignment between humans and machines in LLMs.
problem Human-machine misalignment in LLMs scoring mechanisms.
method Lightweight calibration framework for blackbox models.
result Provably guarantees alignment between humans and machines.
Unified framework for realizable and agnostic learning.
problem Lack of a unified theory for realizable and agnostic learnability.
method Three-line blackbox reduction.
result Unified understanding across various learning settings.
Imputation-Powered Inference improves subpopulation efficiency in missing data settings.
problem Complex missing data patterns challenge standard inference methods.
method Imputation-Powered Inference (IPI) combines blackbox imputation with bias correction.
result IPI provides valid and efficient M-estimation under MCAR blockwise missingness.
SureMap estimates model performance across subpopulations efficiently.
problem Estimating model performance across subpopulations with scarce data.
method Simultaneous Gaussian mean estimation with external data.
result High accuracy in both multi-task and single-task disaggregated evaluations.
Distributional reinforcement learning (distributional RL) has seen empirical success in complex Markov Decision Processes (MDPs) in the setting of nonlinear function approximation. However, there are many different ways in which one can leverage the distributional approach to reinforcement learning. In this paper, we p…
Deep neural network learns optimal trading controls for high-frequency finance.
problem Optimal trading on high-frequency data with market impact and limited data.
method Deep neural network, Monte-Carlo initialization, transfer learning, explainable controls.
result Neural network learns optimal controls for trader preferences.
Recent developments have established the vulnerability of deep reinforcement learning to policy manipulation attacks via intentionally perturbed inputs, known as adversarial examples. In this work, we propose a technique for mitigation of such attacks based on addition of noise to the parameter space of deep reinforcem…
New method tracks shifts in infinite-armed bandits without prior knowledge.
problem Tracking shifts in non-stationary infinite-armed bandits.
method Blackbox conversion of finite-armed MAB to infinite-armed non-stationary, randomized elimination.
result First parameter-free optimal regret bounds for all reservoir regularity regimes.
As convolutional neural networks (CNNs) enable state-of-the-art computer vision applications, their high energy consumption has emerged as a key impediment to their deployment on embedded and mobile devices. Towards efficient image classification under hardware constraints, prior work has proposed adaptive CNNs, i.e., …
Safe Bayesian optimization tackles safety constraints in control engineering.
problem Handling safety constraints in parameter tuning of control systems.
method Lipschitz-only Safe Bayesian Optimization (LoSBO) and LoS-GP-UCB.
result SafeBO algorithms can violate safety constraints due to unreliable uncertainty bounds.
The ability to interpret machine learning models has become increasingly important now that machine learning is used to inform consequential decisions. We propose an approach called model extraction for interpreting complex, blackbox models. Our approach approximates the complex model using a much more interpretable mo…
Prior Networks improve adversarial attack detection by estimating uncertainty.
problem Detecting adversarial attacks in AI systems, especially those designed to evade detection methods.
method Prior Networks estimate predictive uncertainty, outperforming Monte-Carlo dropout ensembles.
result Prior Networks significantly outperform baseline approaches in detecting adversarial attacks.
Workshop reviews techniques to understand neural NLP models.
problem Understanding the inner workings of neural networks in natural language processing.
method Systematic manipulation of inputs, decoding intermediate representations, modifying architectures, and testing on simplified languages.
result Various techniques can improve explainability of neural network models.
We present an information-theoretic framework for sequential adaptive compressed sensing, Info-Greedy Sensing, where measurements are chosen to maximize the extracted information conditioned on the previous measurements. We show that the widely used bisection approach is Info-Greedy for a family of k-sparse signals b…
GPEX interprets deep neural networks without strict assumptions.
problem Interpreting deep neural networks without strict assumptions.
method Derive an evidence lower-bound that encourages GP's posterior to match ANN's output.
result GPs can closely match ANN's outputs without strict assumptions.
Joint state and parameter estimation is a core problem for dynamic Bayesian networks. Although modern probabilistic inference toolkits make it relatively easy to specify large and practically relevant probabilistic models, the silver bullet---an efficient and general online inference algorithm for such problems---remai…
New adversarial attack method based on deep feature distributions.
problem Adversarial attacks on CNN classifiers using output layer information.
method Modeling and exploiting class-wise and layer-wise deep feature distributions.
result Achieves state-of-the-art transfer-based attack results for undefended ImageNet models.
ES optimization improved by structured control variates.
problem Improving accuracy of Evolution Strategies in RL.
method RL-specific variance reduction through structured control variates.
result Structured control variates outperform general variance reduction methods.
Bayesian optimization (BO) aims to minimize a given blackbox function using a model that is updated whenever new evidence about the function becomes available. Here, we address the problem of BO under partially right-censored response data, where in some evaluations we only obtain a lower bound on the function value. T…
BO method improved by density-ratio estimation for better efficiency and scalability.
problem Limitations in Bayesian optimization due to analytical tractability of predictive models.
method Reformulated Bayesian optimization by casting expected improvement as a binary classification problem.
result Improved efficiency and scalability of Bayesian optimization.
AutoAIViz visualizes AutoAI's model generation process, improving user trust.
problem Limited transparency in AutoAI systems leads to lack of user understanding and trust.
method Developed and evaluated AutoAIViz, an experimental system that visualizes AutoAI's model generation process.
result AutoAIViz helps users complete data science tasks and increases their understanding, improving trust in AutoAI systems.
Automates design of lightweight neural networks for image classification.
problem Designing efficient neural networks for edge devices with limited computational resources.
method Uses the Mesh Adaptive Direct Search (MADS) algorithm to optimize network architecture.
result Achieves comparable performance to standard methods with fewer design trials.