Remembering Transformer for Continual Learning

Neural networks encounter the challenge of Catastrophic Forgetting (CF) in continual learning, where new task knowledge interferes with previously learned knowledge. We propose Remembering Transformer, inspired by the brain's Complementary Learning Systems (CLS), to tackle this issue. Remembering Transformer employs a mixture-of-adapters and a generative model-based routing mechanism to alleviate CF by dynamically routing task data to relevant adapters. Our approach demonstrated a new SOTA performance in various vision continual learning tasks and great parameter efficiency.

Mixing Histopathology Prototypes into Robust Slide-Level Representations for Cancer Subtyping

Whole-slide image analysis via the means of computational pathology often relies on processing tessellated gigapixel images with only slide-level labels available. Applying multiple instance learning-based methods or transformer models is computationally expensive as, for each image, all instances have to be processed simultaneously. The MLP-Mixer is an under-explored alternative model to common vision transformers, especially for large-scale datasets. Due to the lack of a self-attention mechanism, they have linear computational complexity to the number of input patches but achieve comparable performance on natural image datasets. We propose a combination of feature embedding and clustering to preprocess the full whole-slide image into a reduced prototype representation which can then serve as input to a suitable MLP-Mixer architecture. Our experiments on two public benchmarks and one inhouse malignant lymphoma dataset show comparable performance to current state-of-the-art methods, while achieving lower training costs in terms of computational time and memory load. Code is publicly available at https://github.com/butkej/ProtoMixer.

Statistically Significant Concept-based Explanation of Image Classifiers via Model Knockoffs

A concept-based classifier can explain the decision process of a deep learning model by human-understandable concepts in image classification problems. However, sometimes concept-based explanations may cause false positives, which misregards unrelated concepts as important for the prediction task. Our goal is to find the statistically significant concept for classification to prevent misinterpretation. In this study, we propose a method using a deep learning model to learn the image concept and then using the Knockoff samples to select the important concepts for prediction by controlling the False Discovery Rate (FDR) under a certain value. We evaluate the proposed method in our synthetic and real data experiments. Also, it shows that our method can control the FDR properly while selecting highly interpretable concepts to improve the trustworthiness of the model.

Heterogeneous Domain Adaptation with Positive and Unlabeled Data

Heterogeneous unsupervised domain adaptation (HUDA) is the most challenging domain adaptation setting where the feature space differs between source and target domains, and the target domain has only unlabeled data. Existing HUDA methods assume that both positive and negative examples are available in the source domain, which may not be satisfied in some real applications. This paper addresses a new challenging setting called positive and unlabeled heterogeneous domain adaptation (PU-HDA), a HUDA setting where the source domain only has positives. PU-HDA can also be viewed as an extension of PU learning where the positive and unlabeled examples are sampled from different domains. A naive combination of existing HUDA and PU learning methods is ineffective in PU-HDA due to the gap in label distribution between the source and target domains. To overcome this issue, we propose a novel method, positive-adversarial domain adaptation (PADA), which can predict likely positive examples from the unlabeled target data and simultaneously align the feature spaces to reduce the distribution divergence between the whole source data and the likely positive target data. PADA achieves this by a unified adversarial training framework for learning a classifier to predict positive examples and a feature transformer to transform the target feature space to that of the source. Specifically, they are both trained to fool a common discriminator that determines whether the likely positive examples are from the target or source domain. We experimentally show that PADA outperforms several baseline methods, such as the naive combination of HUDA and PU learning.

Instance-level Trojan Attacks on Visual Question Answering via Adversarial Learning in Neuron Activation Space

Malicious perturbations embedded in input data, known as Trojan attacks, can cause neural networks to misbehave. However, the impact of a Trojan attack is reduced during fine-tuning of the model, which involves transferring knowledge from a pretrained large-scale model like visual question answering (VQA) to the target model. To mitigate the effects of a Trojan attack, replacing and fine-tuning multiple layers of the pretrained model is possible. This research focuses on sample efficiency, stealthiness and variation, and robustness to model fine-tuning. To address these challenges, we propose an instance-level Trojan attack that generates diverse Trojans across input samples and modalities. Adversarial learning establishes a correlation between a specified perturbation layer and the misbehavior of the fine-tuned model. We conducted extensive experiments on the VQA-v2 dataset using a range of metrics. The results show that our proposed method can effectively adapt to a fine-tuned model with minimal samples. Specifically, we found that a model with a single fine-tuning layer can be compromised using a single shot of adversarial samples, while a model with more fine-tuning layers can be compromised using only a few shots.

Covariance Matrix Adaptation Evolutionary Strategy with Worst-Case Ranking Approximation for Min--Max Optimization and its Application to Berthing Control Tasks

In this study, we consider a continuous min--max optimization problem $\min_{x \in \mathbb{X} \max_{y \in \mathbb{Y}}}f(x,y)$ whose objective function is a black-box. We propose a novel approach to minimize the worst-case objective function $F(x) = \max_{y} f(x,y)$ directly using a covariance matrix adaptation evolution strategy (CMA-ES) in which the rankings of solution candidates are approximated by our proposed worst-case ranking approximation (WRA) mechanism. We develop two variants of WRA combined with CMA-ES and approximate gradient ascent as numerical solvers for the inner maximization problem. Numerical experiments show that our proposed approach outperforms several existing approaches when the objective function is a smooth strongly convex--concave function and the interaction between $x$ and $y$ is strong. We investigate the advantages of the proposed approach for problems where the objective function is not limited to smooth strongly convex--concave functions. The effectiveness of the proposed approach is demonstrated in the robust berthing control problem with uncertainty.ngly convex--concave functions. The effectiveness of the proposed approach is demonstrated in the robust berthing control problem with uncertainty.

Few-Shot Image-to-Semantics Translation for Policy Transfer in Reinforcement Learning

We investigate policy transfer using image-to-semantics translation to mitigate learning difficulties in vision-based robotics control agents. This problem assumes two environments: a simulator environment with semantics, that is, low-dimensional and essential information, as the state space, and a real-world environment with images as the state space. By learning mapping from images to semantics, we can transfer a policy, pre-trained in the simulator, to the real world, thereby eliminating real-world on-policy agent interactions to learn, which are costly and risky. In addition, using image-to-semantics mapping is advantageous in terms of the computational efficiency to train the policy and the interpretability of the obtained policy over other types of sim-to-real transfer strategies. To tackle the main difficulty in learning image-to-semantics mapping, namely the human annotation cost for producing a training dataset, we propose two techniques: pair augmentation with the transition function in the simulator environment and active learning. We observed a reduction in the annotation cost without a decline in the performance of the transfer, and the proposed approach outperformed the existing approach without annotation.

Adaptive Scenario Subset Selection for Worst-Case Optimization and its Application to Well Placement Optimization

In this study, we consider simulation-based worst-case optimization problems with continuous design variables and a finite scenario set. To reduce the number of simulations required and increase the number of restarts for better local optimum solutions, we propose a new approach referred to as adaptive scenario subset selection (AS3). The proposed approach subsamples a scenario subset as a support to construct the worst-case function in a given neighborhood, and we introduce such a scenario subset. Moreover, we develop a new optimization algorithm by combining AS3 and the covariance matrix adaptation evolution strategy (CMA-ES), denoted AS3-CMA-ES. At each algorithmic iteration, a subset of support scenarios is selected, and CMA-ES attempts to optimize the worst-case objective computed only through a subset of the scenarios. The proposed algorithm reduces the number of simulations required by executing simulations on only a scenario subset, rather than on all scenarios. In numerical experiments, we verified that AS3-CMA-ES is more efficient in terms of the number of simulations than the brute-force approach and a surrogate-assisted approach lq-CMA-ES when the ratio of the number of support scenarios to the total number of scenarios is relatively small. In addition, the usefulness of AS3-CMA-ES was evaluated for well placement optimization for carbon dioxide capture and storage (CCS). In comparison with the brute-force approach and lq-CMA-ES, AS3-CMA-ES was able to find better solutions because of more frequent restarts.

Max-Min Off-Policy Actor-Critic Method Focusing on Worst-Case Robustness to Model Misspecification

In the field of reinforcement learning, because of the high cost and risk of policy training in the real world, policies are trained in a simulation environment and transferred to the corresponding real-world environment. However, the simulation environment does not perfectly mimic the real-world environment, lead to model misspecification. Multiple studies report significant deterioration of policy performance in a real-world environment. In this study, we focus on scenarios involving a simulation environment with uncertainty parameters and the set of their possible values, called the uncertainty parameter set. The aim is to optimize the worst-case performance on the uncertainty parameter set to guarantee the performance in the corresponding real-world environment. To obtain a policy for the optimization, we propose an off-policy actor-critic approach called the Max-Min Twin Delayed Deep Deterministic Policy Gradient algorithm (M2TD3), which solves a max-min optimization problem using a simultaneous gradient ascent descent approach. Experiments in multi-joint dynamics with contact (MuJoCo) environments show that the proposed method exhibited a worst-case performance superior to several baseline approaches.

Convergence rate of the (1+1)-evolution strategy on locally strongly convex functions with lipschitz continuous gradient and their monotonic transformations

Evolution strategy (ES) is one of promising classes of algorithms for black-box continuous optimization. Despite its broad successes in applications, theoretical analysis on the speed of its convergence is limited on convex quadratic functions and their monotonic transformation. In this study, an upper bound and a lower bound of the rate of linear convergence of the (1+1)-ES on locally $L$-strongly convex functions with $U$-Lipschitz continuous gradient are derived as $\exp\left(-\Omega_{d\to\infty}\left(\frac{L}{d\cdot U}\right)\right)$ and $\exp\left(-\frac1d\right)$, respectively. Notably, any prior knowledge on the mathematical properties of the objective function such as Lipschitz constant is not given to the algorithm, whereas the existing analyses of derivative-free optimization algorithms require them.