Modern deep-learning systems are specialized to problem settings in which training occurs once and then never again, as opposed to continual-learning settings in which training occurs continually. If deep-learning systems are applied in a continual learning setting, then it is well known that they may fail to remember earlier examples. More fundamental, but less well known, is that they may also lose their ability to learn on new examples, a phenomenon called loss of plasticity. We provide direct demonstrations of loss of plasticity using the MNIST and ImageNet datasets repurposed for continual learning as sequences of tasks. In ImageNet, binary classification performance dropped from 89% accuracy on an early task down to 77%, about the level of a linear network, on the 2000th task. Loss of plasticity occurred with a wide range of deep network architectures, optimizers, activation functions, batch normalization, dropout, but was substantially eased by L2-regularization, particularly when combined with weight perturbation. Further, we introduce a new algorithm -- continual backpropagation -- which slightly modifies conventional backpropagation to reinitialize a small fraction of less-used units after each example and appears to maintain plasticity indefinitely.
While there has been a growing research interest in developing out-of-distribution (OOD) detection methods, there has been comparably little discussion around how these methods should be evaluated. Given their relevance for safe(r) AI, it is important to examine whether the basis for comparing OOD detection methods is consistent with practical needs. In this work, we take a closer look at the go-to metrics for evaluating OOD detection, and question the approach of exclusively reducing OOD detection to a binary classification task with little consideration for the detection threshold. We illustrate the limitations of current metrics (AUROC & its friends) and propose a new metric - Area Under the Threshold Curve (AUTC), which explicitly penalizes poor separation between ID and OOD samples. Scripts and data are available at https://github.com/glhr/beyond-auroc
One of the more significant obstacles in classification of skin cancer is the presence of artifacts. This paper investigates the effect of dark corner artifacts, which result from the use of dermoscopes, on the performance of a deep learning binary classification task. Previous research attempted to remove and inpaint dark corner artifacts, with the intention of creating an ideal condition for models. However, such research has been shown to be inconclusive due to lack of available datasets labelled with dark corner artifacts and detailed analysis and discussion. To address these issues, we label 10,250 skin lesion images from publicly available datasets and introduce a balanced dataset with an equal number of melanoma and non-melanoma cases. The training set comprises 6126 images without artifacts, and the testing set comprises 4124 images with dark corner artifacts. We conduct three experiments to provide new understanding on the effects of dark corner artifacts, including inpainted and synthetically generated examples, on a deep learning method. Our results suggest that introducing synthetic dark corner artifacts which have been superimposed onto the training set improved model performance, particularly in terms of the true negative rate. This indicates that deep learning learnt to ignore dark corner artifacts, rather than treating it as melanoma, when dark corner artifacts were introduced into the training set. Further, we propose a new approach to quantifying heatmaps indicating network focus using a root mean square measure of the brightness intensity in the different regions of the heatmaps. This paper provides a new guideline for skin lesions analysis with an emphasis on reproducibility.
Cyber attacks deceive machines into believing something that does not exist in the first place. However, there are some to which even humans fall prey. One such famous attack that attackers have used over the years to exploit the vulnerability of vision is known to be a Homoglyph attack. It employs a primary yet effective mechanism to create illegitimate domains that are hard to differentiate from legit ones. Moreover, as the difference is pretty indistinguishable for a user to notice, they cannot stop themselves from clicking on these homoglyph domain names. In many cases, that results in either information theft or malware attack on their systems. Existing approaches use simple, string-based comparison techniques applied in primary language-based tasks. Although they are impactful to some extent, they usually fail because they are not robust to different types of homoglyphs and are computationally not feasible because of their time requirement proportional to the string length. Similarly, neural network-based approaches are employed to determine real domain strings from fake ones. Nevertheless, the problem with both methods is that they require paired sequences of real and fake domain strings to work with, which is often not the case in the real world, as the attacker only sends the illegitimate or homoglyph domain to the vulnerable user. Therefore, existing approaches are not suitable for practical scenarios in the real world. In our work, we created GlyphNet, an image dataset that contains 4M domains, both real and homoglyphs. Additionally, we introduce a baseline method for a homoglyph attack detection system using an attention-based convolutional Neural Network. We show that our model can reach state-of-the-art accuracy in detecting homoglyph attacks with a 0.93 AUC on our dataset.
Unsupervised performance estimation, or evaluating how well models perform on unlabeled data is a difficult task. Recently, a method was proposed by Garg et al. [2022] which performs much better than previous methods. Their method relies on having a score function, satisfying certain properties, to map probability vectors outputted by the classifier to the reals, but it is an open problem which score function is best. We explore this problem by first showing that their method fundamentally relies on the ordering induced by this score function. Thus, under monotone transformations of score functions, their method yields the same estimate. Next, we show that in the binary classification setting, nearly all common score functions - the $L^\infty$ norm; the $L^2$ norm; negative entropy; and the $L^2$, $L^1$, and Jensen-Shannon distances to the uniform vector - all induce the same ordering over probability vectors. However, this does not hold for higher dimensional settings. We conduct numerous experiments on well-known NLP data sets and rigorously explore the performance of different score functions. We conclude that the $L^\infty$ norm is the most appropriate.
Algorithmic Fairness is an established field in machine learning that aims to reduce biases in data. Recent advances have proposed various methods to ensure fairness in a univariate environment, where the goal is to de-bias a single task. However, extending fairness to a multi-task setting, where more than one objective is optimised using a shared representation, remains underexplored. To bridge this gap, we develop a method that extends the definition of Strong Demographic Parity to multi-task learning using multi-marginal Wasserstein barycenters. Our approach provides a closed form solution for the optimal fair multi-task predictor including both regression and binary classification tasks. We develop a data-driven estimation procedure for the solution and run numerical experiments on both synthetic and real datasets. The empirical results highlight the practical value of our post-processing methodology in promoting fair decision-making.
In this work we propose a novel concept of a hierarchical confusion matrix, opening the door for popular confusion matrix based (flat) evaluation measures from binary classification problems, while considering the peculiarities of hierarchical classification problems. We develop the concept to a generalized form and prove its applicability to all types of hierarchical classification problems including directed acyclic graphs, multi path labelling, and non mandatory leaf node prediction. Finally, we use measures based on the novel confusion matrix to evaluate models within a benchmark for three real world hierarchical classification applications and compare the results to established evaluation measures. The results outline the reasonability of this approach and its usefulness to evaluate hierarchical classification problems. The implementation of hierarchical confusion matrix is available on GitHub.
Active learning is a paradigm that significantly enhances the performance of machine learning models when acquiring labeled data is expensive. While several benchmarks exist for evaluating active learning strategies, their findings exhibit some misalignment. This discrepancy motivates us to develop a transparent and reproducible benchmark for the community. Our efforts result in an open-sourced implementation (https://github.com/ariapoy/active-learning-benchmark) that is reliable and extensible for future research. By conducting thorough re-benchmarking experiments, we have not only rectified misconfigurations in existing benchmark but also shed light on the under-explored issue of model compatibility, which directly causes the observed discrepancy. Resolving the discrepancy reassures that the uncertainty sampling strategy of active learning remains an effective and preferred choice for most datasets. Our experience highlights the importance of dedicating research efforts towards re-benchmarking existing benchmarks to produce more credible results and gain deeper insights.
While there has been a surge of large language models for Norwegian in recent years, we lack any tool to evaluate their understanding of grammaticality. We present two new Norwegian datasets for this task. NoCoLA_class is a supervised binary classification task where the goal is to discriminate between acceptable and non-acceptable sentences. On the other hand, NoCoLA_zero is a purely diagnostic task for evaluating the grammatical judgement of a language model in a completely zero-shot manner, i.e. without any further training. In this paper, we describe both datasets in detail, show how to use them for different flavors of language models, and conduct a comparative study of the existing Norwegian language models.
We present Strokes2Surface, an offline geometry reconstruction pipeline that recovers well-connected curve networks from imprecise 4D sketches to bridge concept design and digital modeling stages in architectural design. The input to our pipeline consists of 3D strokes' polyline vertices and their timestamps as the 4th dimension, along with additional metadata recorded throughout sketching. Inspired by architectural sketching practices, our pipeline combines a classifier and two clustering models to achieve its goal. First, with a set of extracted hand-engineered features from the sketch, the classifier recognizes the type of individual strokes between those depicting boundaries (Shape strokes) and those depicting enclosed areas (Scribble strokes). Next, the two clustering models parse strokes of each type into distinct groups, each representing an individual edge or face of the intended architectural object. Curve networks are then formed through topology recovery of consolidated Shape clusters and surfaced using Scribble clusters guiding the cycle discovery. Our evaluation is threefold: We confirm the usability of the Strokes2Surface pipeline in architectural design use cases via a user study, we validate our choice of features via statistical analysis and ablation studies on our collected dataset, and we compare our outputs against a range of reconstructions computed using alternative methods.
Algorithmic recourse provides explanations that help users overturn an unfavorable decision by a machine learning system. But so far very little attention has been paid to whether providing recourse is beneficial or not. We introduce an abstract learning-theoretic framework that compares the risks (i.e., expected losses) for classification with and without algorithmic recourse. This allows us to answer the question of when providing recourse is beneficial or harmful at the population level. Surprisingly, we find that there are many plausible scenarios in which providing recourse turns out to be harmful, because it pushes users to regions of higher class uncertainty and therefore leads to more mistakes. We further study whether the party deploying the classifier has an incentive to strategize in anticipation of having to provide recourse, and we find that sometimes they do, to the detriment of their users. Providing algorithmic recourse may therefore also be harmful at the systemic level. We confirm our theoretical findings in experiments on simulated and real-world data. All in all, we conclude that the current concept of algorithmic recourse is not reliably beneficial, and therefore requires rethinking.
Entity Matching (EM) involves identifying different data representations referring to the same entity from multiple data sources and is typically formulated as a binary classification problem. It is a challenging problem in data integration due to the heterogeneity of data representations. State-of-the-art solutions have adopted NLP techniques based on pre-trained language models (PrLMs) via the fine-tuning paradigm, however, sequential fine-tuning of overparameterized PrLMs can lead to catastrophic forgetting, especially in low-resource scenarios. In this study, we propose a parameter-efficient paradigm for fine-tuning PrLMs based on adapters, small neural networks encapsulated between layers of a PrLM, by optimizing only the adapter and classifier weights while the PrLMs parameters are frozen. Adapter-based methods have been successfully applied to multilingual speech problems achieving promising results, however, the effectiveness of these methods when applied to EM is not yet well understood, particularly for generalized EM with heterogeneous data. Furthermore, we explore using (i) pre-trained adapters and (ii) invertible adapters to capture token-level language representations and demonstrate their benefits for transfer learning on the generalized EM benchmark. Our results show that our solution achieves comparable or superior performance to full-scale PrLM fine-tuning and prompt-tuning baselines while utilizing a significantly smaller computational footprint $\approx 13\%$ of the PrLM parameters.
Detecting factual errors in summaries has been an important and challenging subject in summarization research. Inspired by the emergent ability of large language models (LLMs), we explore evaluating factual consistency of summaries by directly prompting LLMs. We present a comprehensive empirical study to assess the ability of LLMs as factual consistency evaluators, which consists of (1) analyzing different LLMs such as the GPT model series and Flan-T5; (2) investigating a variety of prompting methods including vanilla prompting, chain-of-thought prompting, and a sentence-by-sentence prompting method to tackle long summaries; and (3) evaluating on diverse summaries generated by multiple summarization systems, ranging from pre-transformer methods to SOTA pretrained models. Our experiments demonstrate that prompting LLMs is able to outperform the previous best factuality systems in all settings, by up to 12.2 absolute points in terms of the binary classification accuracy on inconsistency detection.
When facing an unsatisfactory prediction from a machine learning model, users can be interested in investigating the underlying reasons and exploring the potential for reversing the outcome. We ask: To flip the prediction on a test point $x_t$, how to identify the smallest training subset $\mathcal{S}_t$ that we need to relabel? We propose an efficient algorithm to identify and relabel such a subset via an extended influence function for binary classification models with convex loss. We find that relabeling fewer than 2% of the training points can always flip a prediction. This mechanism can serve multiple purposes: (1) providing an approach to challenge a model prediction by altering training points; (2) evaluating model robustness with the cardinality of the subset (i.e., $|\mathcal{S}_t|$); we show that $|\mathcal{S}_t|$ is highly related to the noise ratio in the training set and $|\mathcal{S}_t|$ is correlated with but complementary to predicted probabilities; and (3) revealing training points lead to group attribution bias. To the best of our knowledge, we are the first to investigate identifying and relabeling the minimal training subset required to flip a given prediction.
We introduce a novel loss function to minimize the outage probability of an ML-based resource allocation system. A single-user multi-resource greedy allocation strategy constitutes our application scenario, for which an ML binary classification predictor assists in selecting a resource satisfying the established outage criterium. While other resource allocation policies may be suitable, they are not the focus of our study. Instead, our primary emphasis is on theoretically developing this loss function and leveraging it to train an ML model to address the outage probability challenge. With no access to future channel state information, this predictor foresees each resource's likely future outage status. When the predictor encounters a resource it believes will be satisfactory, it allocates it to the user. Our main result establishes exact and asymptotic expressions for this system's outage probability. These expressions reveal that focusing solely on the optimization of the per-resource outage probability conditioned on the ML predictor recommending resource allocation (a strategy that appears to be most appropriate) may produce inadequate predictors that reject every resource. They also reveal that focusing on standard metrics, like precision, false-positive rate, or recall, may not produce optimal predictors. With our result, we formulate a theoretically optimal, differentiable loss function to train our predictor. We then compare predictors trained using this and traditional loss functions namely, binary cross-entropy (BCE), mean squared error (MSE), and mean absolute error (MAE). In all scenarios, predictors trained using our novel loss function provide superior outage probability performance. Moreover, in some cases, our loss function outperforms predictors trained with BCE, MAE, and MSE by multiple orders of magnitude.
Randomized smoothing is a technique for providing provable robustness guarantees against adversarial attacks while making minimal assumptions about a classifier. This method relies on taking a majority vote of any base classifier over multiple noise-perturbed inputs to obtain a smoothed classifier, and it remains the tool of choice to certify deep and complex neural network models. Nonetheless, non-trivial performance of such smoothed classifier crucially depends on the base model being trained on noise-augmented data, i.e., on a smoothed input distribution. While widely adopted in practice, it is still unclear how this noisy training of the base classifier precisely affects the risk of the robust smoothed classifier, leading to heuristics and tricks that are poorly understood. In this work we analyze these trade-offs theoretically in a binary classification setting, proving that these common observations are not universal. We show that, without making stronger distributional assumptions, no benefit can be expected from predictors trained with noise-augmentation, and we further characterize distributions where such benefit is obtained. Our analysis has direct implications to the practical deployment of randomized smoothing, and we illustrate some of these via experiments on CIFAR-10 and MNIST, as well as on synthetic datasets.
Decision Trees (DTs) are commonly used for many machine learning tasks due to their high degree of interpretability. However, learning a DT from data is a difficult optimization problem, as it is non-convex and non-differentiable. Therefore, common approaches learn DTs using a greedy growth algorithm that minimizes the impurity locally at each internal node. Unfortunately, this greedy procedure can lead to inaccurate trees. In this paper, we present a novel approach for learning hard, axis-aligned DTs with gradient descent. The proposed method uses backpropagation with a straight-through operator on a dense DT representation, to jointly optimize all tree parameters. Our approach outperforms existing methods on binary classification benchmarks and achieves competitive results for multi-class tasks. The method is available under: https://github.com/s-marton/GradTree
Most deepfake detection methods focus on detecting spatial and/or spatio-temporal changes in facial attributes and are centered around the binary classification task of detecting whether a video is real or fake. This is because available benchmark datasets contain mostly visual-only modifications present in the entirety of the video. However, a sophisticated deepfake may include small segments of audio or audio-visual manipulations that can completely change the meaning of the video content. To addresses this gap, we propose and benchmark a new dataset, Localized Audio Visual DeepFake (LAV-DF), consisting of strategic content-driven audio, visual and audio-visual manipulations. The proposed baseline method, Boundary Aware Temporal Forgery Detection (BA-TFD), is a 3D Convolutional Neural Network-based architecture which effectively captures multimodal manipulations. We further improve (i.e. BA-TFD+) the baseline method by replacing the backbone with a Multiscale Vision Transformer and guide the training process with contrastive, frame classification, boundary matching and multimodal boundary matching loss functions. The quantitative analysis demonstrates the superiority of BA-TFD+ on temporal forgery localization and deepfake detection tasks using several benchmark datasets including our newly proposed dataset. The dataset, models and code are available at https://github.com/ControlNet/LAV-DF.
Computerized cognitive training (CCT) is a scalable, well-tolerated intervention that has promise for slowing cognitive decline. Outcomes from CCT are limited by a lack of effective engagement, which is decreased by factors such as mental fatigue, particularly in older adults at risk for dementia. There is a need for scalable, automated measures that can monitor mental fatigue during CCT. Here, we develop and validate a novel Recurrent Video Transformer (RVT) method for monitoring real-time mental fatigue in older adults with mild cognitive impairment from video-recorded facial gestures during CCT. The RVT model achieved the highest balanced accuracy(78%) and precision (0.82) compared to the prior state-of-the-art models for binary and multi-class classification of mental fatigue and was additionally validated via significant association (p=0.023) with CCT reaction time. By leveraging dynamic temporal information, the RVT model demonstrates the potential to accurately measure real-time mental fatigue, laying the foundation for future personalized CCT that increase effective engagement.
As an important task in remote sensing image analysis, remote sensing change detection (RSCD) aims to identify changes of interest in a region from spatially co-registered multi-temporal remote sensing images, so as to monitor the local development. Existing RSCD methods usually formulate RSCD as a binary classification task, representing changes of interest by merely feature concatenation or feature subtraction and recovering the spatial details via densely connected change representations, whose performances need further improvement. In this paper, we propose STNet, a RSCD network based on spatial and temporal feature fusions. Specifically, we design a temporal feature fusion (TFF) module to combine bi-temporal features using a cross-temporal gating mechanism for emphasizing changes of interest; a spatial feature fusion module is deployed to capture fine-grained information using a cross-scale attention mechanism for recovering the spatial details of change representations. Experimental results on three benchmark datasets for RSCD demonstrate that the proposed method achieves the state-of-the-art performance. Code is available at https://github.com/xwmaxwma/rschange.