Deep learning achieves outstanding results in many machine learning tasks. Nevertheless, it is vulnerable to backdoor attacks that modify the training set to embed a secret functionality in the trained model. The modified training samples have a secret property, i. e., a trigger. At inference time, the secret functionality is activated when the input contains the trigger, while the model functions correctly in other cases. While there are many known backdoor attacks (and defenses), deploying a stealthy attack is still far from trivial. Successfully creating backdoor triggers depends on numerous parameters. Unfortunately, research has not yet determined which parameters contribute most to the attack performance. This paper systematically analyzes the most relevant parameters for the backdoor attacks, i.e., trigger size, position, color, and poisoning rate. Using transfer learning, which is very common in computer vision, we evaluate the attack on state-of-the-art models (ResNet, VGG, AlexNet, and GoogLeNet) and datasets (MNIST, CIFAR10, and TinyImageNet). Our attacks cover the majority of backdoor settings in research, providing concrete directions for future works. Our code is publicly available to facilitate the reproducibility of our results.
Purpose: In curriculum learning, the idea is to train on easier samples first and gradually increase the difficulty, while in self-paced learning, a pacing function defines the speed to adapt the training progress. While both methods heavily rely on the ability to score the difficulty of data samples, an optimal scoring function is still under exploration. Methodology: Distillation is a knowledge transfer approach where a teacher network guides a student network by feeding a sequence of random samples. We argue that guiding student networks with an efficient curriculum strategy can improve model generalization and robustness. For this purpose, we design an uncertainty-based paced curriculum learning in self distillation for medical image segmentation. We fuse the prediction uncertainty and annotation boundary uncertainty to develop a novel paced-curriculum distillation (PCD). We utilize the teacher model to obtain prediction uncertainty and spatially varying label smoothing with Gaussian kernel to generate segmentation boundary uncertainty from the annotation. We also investigate the robustness of our method by applying various types and severity of image perturbation and corruption. Results: The proposed technique is validated on two medical datasets of breast ultrasound image segmentation and robotassisted surgical scene segmentation and achieved significantly better performance in terms of segmentation and robustness. Conclusion: P-CD improves the performance and obtains better generalization and robustness over the dataset shift. While curriculum learning requires extensive tuning of hyper-parameters for pacing function, the level of performance improvement suppresses this limitation.
We consider gradient-based optimisation of wide, shallow neural networks, where the output of each hidden node is scaled by a positive parameter. The scaling parameters are non-identical, differing from the classical Neural Tangent Kernel (NTK) parameterisation. We prove that for large such neural networks, with high probability, gradient flow and gradient descent converge to a global minimum and can learn features in some sense, unlike in the NTK parameterisation. We perform experiments illustrating our theoretical results and discuss the benefits of such scaling in terms of prunability and transfer learning.
Neurophysiological time-series recordings of brain activity like the electroencephalogram (EEG) or local field potentials can be decoded by machine learning models in order to either control an application, e.g., for communication or rehabilitation after stroke, or to passively monitor the ongoing brain state of the subject, e.g., in a demanding work environment. A typical decoding challenge faced by a brain-computer interface (BCI) is the small dataset size compared to other domains of machine learning like computer vision or natural language processing. The possibilities to tackle classification or regression problems in BCI are to either train a regular model on the available small training data sets or through transfer learning, which utilizes data from other sessions, subjects, or even datasets to train a model. Transfer learning is non-trivial because of the non-stationary of EEG signals between subjects but also within subjects. This variability calls for explicit calibration phases at the start of every session, before BCI applications can be used online. In this study, we present arguments to BCI researchers to encourage the use of embeddings for EEG decoding. In particular, we introduce a simple domain adaptation technique involving both deep learning (when learning the embeddings from the source data) and classical machine learning (for fast calibration on the target data). This technique allows us to learn embeddings across subjects, which deliver a generalized data representation. These can then be fed into subject-specific classifiers in order to minimize their need for calibration data. We conducted offline experiments on the 14 subjects of the High Gamma EEG-BCI Dataset [1]. Embedding functions were obtained by training EEGNet [2] using a leave-one-subject-out (LOSO) protocol, and the embedding vectors were classified by the logistic regression algorithm. Our pipeline was compared to two baseline approaches: EEGNet without subject-specific calibration and the standard FBCSP pipeline in a within-subject training. We observed that the representations learned by the embedding functions were indeed non-stationary across subjects, justifying the need for an additional subject-specific calibration. We also observed that the subject-specific calibration indeed improved the score. Finally, our data suggest, that building upon embeddings requires fewer individual calibration data than the FBCSP baseline to reach satisfactory scores.
Parameter-efficient fine-tuning approaches have recently garnered a lot of attention. Having considerably lower number of trainable weights, these methods can bring about scalability and computational effectiveness. In this paper, we look for optimal sub-networks and investigate the capability of different transformer modules in transferring knowledge from a pre-trained model to a downstream task. Our empirical results suggest that every transformer module in BERT can act as a winning ticket: fine-tuning each specific module while keeping the rest of the network frozen can lead to comparable performance to the full fine-tuning. Among different modules, LayerNorms exhibit the best capacity for knowledge transfer with limited trainable weights, to the extent that, with only 0.003% of all parameters in the layer-wise analysis, they show acceptable performance on various target tasks. On the reasons behind their effectiveness, we argue that their notable performance could be attributed to their high-magnitude weights compared to that of the other modules in the pre-trained BERT.
A fundamental problem in manifold learning is to approximate a functional relationship in a data chosen randomly from a probability distribution supported on a low dimensional sub-manifold of a high dimensional ambient Euclidean space. The manifold is essentially defined by the data set itself and, typically, designed so that the data is dense on the manifold in some sense. The notion of a data space is an abstraction of a manifold encapsulating the essential properties that allow for function approximation. The problem of transfer learning (meta-learning) is to use the learning of a function on one data set to learn a similar function on a new data set. In terms of function approximation, this means lifting a function on one data space (the base data space) to another (the target data space). This viewpoint enables us to connect some inverse problems in applied mathematics (such as inverse Radon transform) with transfer learning. In this paper we examine the question of such lifting when the data is assumed to be known only on a part of the base data space. We are interested in determining subsets of the target data space on which the lifting can be defined, and how the local smoothness of the function and its lifting are related.
The scarcity of labeled data is a long-standing challenge for many machine learning tasks. We propose our gradient flow method to leverage the existing dataset (i.e., source) to generate new samples that are close to the dataset of interest (i.e., target). We lift both datasets to the space of probability distributions on the feature-Gaussian manifold, and then develop a gradient flow method that minimizes the maximum mean discrepancy loss. To perform the gradient flow of distributions on the curved feature-Gaussian space, we unravel the Riemannian structure of the space and compute explicitly the Riemannian gradient of the loss function induced by the optimal transport metric. For practical applications, we also propose a discretized flow, and provide conditional results guaranteeing the global convergence of the flow to the optimum. We illustrate the results of our proposed gradient flow method on several real-world datasets and show our method can improve the accuracy of classification models in transfer learning settings.
A goal of artificial intelligence is to construct an agent that can solve a wide variety of tasks. Recent progress in text-guided image synthesis has yielded models with an impressive ability to generate complex novel images, exhibiting combinatorial generalization across domains. Motivated by this success, we investigate whether such tools can be used to construct more general-purpose agents. Specifically, we cast the sequential decision making problem as a text-conditioned video generation problem, where, given a text-encoded specification of a desired goal, a planner synthesizes a set of future frames depicting its planned actions in the future, after which control actions are extracted from the generated video. By leveraging text as the underlying goal specification, we are able to naturally and combinatorially generalize to novel goals. The proposed policy-as-video formulation can further represent environments with different state and action spaces in a unified space of images, which, for example, enables learning and generalization across a variety of robot manipulation tasks. Finally, by leveraging pretrained language embeddings and widely available videos from the internet, the approach enables knowledge transfer through predicting highly realistic video plans for real robots.
Adversarial nets have proved to be powerful in various domains including generative modeling (GANs), transfer learning, and fairness. However, successfully training adversarial nets using first-order methods remains a major challenge. Typically, careful choices of the learning rates are needed to maintain the delicate balance between the competing networks. In this paper, we design a novel learning rate scheduler that dynamically adapts the learning rate of the adversary to maintain the right balance. The scheduler is driven by the fact that the loss of an ideal adversarial net is a constant known a priori. The scheduler is thus designed to keep the loss of the optimized adversarial net close to that of an ideal network. We run large-scale experiments to study the effectiveness of the scheduler on two popular applications: GANs for image generation and adversarial nets for domain adaptation. Our experiments indicate that adversarial nets trained with the scheduler are less likely to diverge and require significantly less tuning. For example, on CelebA, a GAN with the scheduler requires only one-tenth of the tuning budget needed without a scheduler. Moreover, the scheduler leads to statistically significant improvements in model quality, reaching up to $27\%$ in Frechet Inception Distance for image generation and $3\%$ in test accuracy for domain adaptation.
The adoption of deep neural networks (DNNs) in safety-critical contexts is often prevented by the lack of effective means to explain their results, especially when they are erroneous. In our previous work, we proposed a white-box approach (HUDD) and a black-box approach (SAFE) to automatically characterize DNN failures. They both identify clusters of similar images from a potentially large set of images leading to DNN failures. However, the analysis pipelines for HUDD and SAFE were instantiated in specific ways according to common practices, deferring the analysis of other pipelines to future work. In this paper, we report on an empirical evaluation of 99 different pipelines for root cause analysis of DNN failures. They combine transfer learning, autoencoders, heatmaps of neuron relevance, dimensionality reduction techniques, and different clustering algorithms. Our results show that the best pipeline combines transfer learning, DBSCAN, and UMAP. It leads to clusters almost exclusively capturing images of the same failure scenario, thus facilitating root cause analysis. Further, it generates distinct clusters for each root cause of failure, thus enabling engineers to detect all the unsafe scenarios. Interestingly, these results hold even for failure scenarios that are only observed in a small percentage of the failing images.
Early diagnosis of Alzheimer's disease (AD) is essential in preventing the disease's progression. Therefore, detecting AD from neuroimaging data such as structural magnetic resonance imaging (sMRI) has been a topic of intense investigation in recent years. Deep learning has gained considerable attention in Alzheimer's detection. However, training a convolutional neural network from scratch is challenging since it demands more computational time and a significant amount of annotated data. By transferring knowledge learned from other image recognition tasks to medical image classification, transfer learning can provide a promising and effective solution. Irregularities in the dataset distribution present another difficulty. Class decomposition can tackle this issue by simplifying learning a dataset's class boundaries. Motivated by these approaches, this paper proposes a transfer learning method using class decomposition to detect Alzheimer's disease from sMRI images. We use two ImageNet-trained architectures: VGG19 and ResNet50, and an entropy-based technique to determine the most informative images. The proposed model achieved state-of-the-art performance in the Alzheimer's disease (AD) vs mild cognitive impairment (MCI) vs cognitively normal (CN) classification task with a 3\% increase in accuracy from what is reported in the literature.
Variational quantum algorithms represent a promising approach to quantum machine learning where classical neural networks are replaced by parametrized quantum circuits. However, both approaches suffer from a clear limitation, that is a lack of interpretability. Here, we present a variational method to quantize projective simulation (PS), a reinforcement learning model aimed at interpretable artificial intelligence. Decision making in PS is modeled as a random walk on a graph describing the agent's memory. To implement the quantized model, we consider quantum walks of single photons in a lattice of tunable Mach-Zehnder interferometers trained via variational algorithms. Using an example from transfer learning, we show that the quantized PS model can exploit quantum interference to acquire capabilities beyond those of its classical counterpart. Finally, we discuss the role of quantum interference for training and tracing the decision making process, paving the way for realizations of interpretable quantum learning agents.
Originally proposed as a method for knowledge transfer from one model to another, some recent studies have suggested that knowledge distillation (KD) is in fact a form of regularization. Perhaps the strongest argument of all for this new perspective comes from its apparent similarities with label smoothing (LS). Here we re-examine this stated equivalence between the two methods by comparing the predictive confidences of the models they train. Experiments on four text classification tasks involving models of different sizes show that: (a) In most settings, KD and LS drive model confidence in completely opposite directions, and (b) In KD, the student inherits not only its knowledge but also its confidence from the teacher, reinforcing the classical knowledge transfer view.
Flooding is one of the most destructive and costly natural disasters, and climate changes would further increase risks globally. This work presents a novel multimodal machine learning approach for multi-year global flood risk prediction, combining geographical information and historical natural disaster dataset. Our multimodal framework employs state-of-the-art processing techniques to extract embeddings from each data modality, including text-based geographical data and tabular-based time-series data. Experiments demonstrate that a multimodal approach, that is combining text and statistical data, outperforms a single-modality approach. Our most advanced architecture, employing embeddings extracted using transfer learning upon DistilBert model, achieves 75\%-77\% ROCAUC score in predicting the next 1-5 year flooding event in historically flooded locations. This work demonstrates the potentials of using machine learning for long-term planning in natural disaster management.
Continual learning aims to avoid catastrophic forgetting and effectively leverage learned experiences to master new knowledge. Existing gradient projection approaches impose hard constraints on the optimization space for new tasks to minimize interference, which simultaneously hinders forward knowledge transfer. To address this issue, recent methods reuse frozen parameters with a growing network, resulting in high computational costs. Thus, it remains a challenge whether we can improve forward knowledge transfer for gradient projection approaches using a fixed network architecture. In this work, we propose the Restricted Orthogonal Gradient prOjection (ROGO) framework. The basic idea is to adopt a restricted orthogonal constraint allowing parameters optimized in the direction oblique to the whole frozen space to facilitate forward knowledge transfer while consolidating previous knowledge. Our framework requires neither data buffers nor extra parameters. Extensive experiments have demonstrated the superiority of our framework over several strong baselines. We also provide theoretical guarantees for our relaxing strategy.
AI-Generated Content (AIGC) has recently gained a surge in popularity, powered by its high efficiency and consistency in production, and its capability of being customized and diversified. The cross-modality nature of the representation learning mechanism in most AIGC technology allows for more freedom and flexibility in exploring new types of art that would be impossible in the past. Inspired by the pictogram subset of Chinese characters, we proposed PaCaNet, a CycleGAN-based pipeline for producing novel artworks that fuse two different art types, traditional Chinese painting and calligraphy. In an effort to produce stable and diversified output, we adopted three main technical innovations: 1. Using one-shot learning to increase the creativity of pre-trained models and diversify the content of the fused images. 2. Controlling the preference over generated Chinese calligraphy by freezing randomly sampled parameters in pre-trained models. 3. Using a regularization method to encourage the models to produce images similar to Chinese paintings. Furthermore, we conducted a systematic study to explore the performance of PaCaNet in diversifying fused Chinese painting and calligraphy, which showed satisfying results. In conclusion, we provide a new direction of creating arts by fusing the visual information in paintings and the stroke features in Chinese calligraphy. Our approach creates a unique aesthetic experience rooted in the origination of Chinese hieroglyph characters. It is also a unique opportunity to delve deeper into traditional artwork and, in doing so, to create a meaningful impact on preserving and revitalizing traditional heritage.
In recent years Convolutional neural networks (CNN) have made significant progress in computer vision. These advancements have been applied to other areas, such as remote sensing and have shown satisfactory results. However, the lack of large labeled datasets and the inherent complexity of remote sensing problems have made it difficult to train deep CNNs for dense prediction problems. To solve this issue, ImageNet pretrained weights have been used as a starting point in various dense predictions tasks. Although this type of transfer learning has led to improvements, the domain difference between natural and remote sensing images has also limited the performance of deep CNNs. On the other hand, self-supervised learning methods for learning visual representations from large unlabeled images have grown substantially over the past two years. Accordingly, in this paper we have explored the effectiveness of in-domain representations in both supervised and self-supervised forms to solve the domain difference between remote sensing and the ImageNet dataset. The obtained weights from remote sensing images are utilized as initial weights for solving semantic segmentation and object detection tasks and state-of-the-art results are obtained. For self-supervised pre-training, we have utilized the SimSiam algorithm as it is simple and does not need huge computational resources. One of the most influential factors in acquiring general visual representations from remote sensing images is the pre-training dataset. To examine the effect of the pre-training dataset, equal-sized remote sensing datasets are used for pre-training. Our results have demonstrated that using datasets with a high spatial resolution for self-supervised representation learning leads to high performance in downstream tasks.
Unsupervised domain adaptation person re-identification (Re-ID) aims to identify pedestrian images within an unlabeled target domain with an auxiliary labeled source-domain dataset. Many existing works attempt to recover reliable identity information by considering multiple homogeneous networks. And take these generated labels to train the model in the target domain. However, these homogeneous networks identify people in approximate subspaces and equally exchange their knowledge with others or their mean net to improve their ability, inevitably limiting the scope of available knowledge and putting them into the same mistake. This paper proposes a Dual-level Asymmetric Mutual Learning method (DAML) to learn discriminative representations from a broader knowledge scope with diverse embedding spaces. Specifically, two heterogeneous networks mutually learn knowledge from asymmetric subspaces through the pseudo label generation in a hard distillation manner. The knowledge transfer between two networks is based on an asymmetric mutual learning manner. The teacher network learns to identify both the target and source domain while adapting to the target domain distribution based on the knowledge of the student. Meanwhile, the student network is trained on the target dataset and employs the ground-truth label through the knowledge of the teacher. Extensive experiments in Market-1501, CUHK-SYSU, and MSMT17 public datasets verified the superiority of DAML over state-of-the-arts.
Few-shot classification consists of a training phase where a model is learned on a relatively large dataset and an adaptation phase where the learned model is adapted to previously-unseen tasks with limited labeled samples. In this paper, we empirically prove that the training algorithm and the adaptation algorithm can be completely disentangled, which allows algorithm analysis and design to be done individually for each phase. Our meta-analysis for each phase reveals several interesting insights that may help better understand key aspects of few-shot classification and connections with other fields such as visual representation learning and transfer learning. We hope the insights and research challenges revealed in this paper can inspire future work in related directions. Code and pre-trained models (in PyTorch) are available at https://github.com/Frankluox/CloserLookAgainFewShot.
Accurate modeling of student knowledge is essential for large-scale online learning systems that are increasingly used for student training. Knowledge tracing aims to model student knowledge state given the student's sequence of learning activities. Modern Knowledge tracing (KT) is usually formulated as a supervised sequence learning problem to predict students' future practice performance according to their past observed practice scores by summarizing student knowledge state as a set of evolving hidden variables. Because of this formulation, many current KT solutions are not fit for modeling student learning from non-assessed learning activities with no explicit feedback or score observation (e.g., watching video lectures that are not graded). Additionally, these models cannot explicitly represent the dynamics of knowledge transfer among different learning activities, particularly between the assessed (e.g., quizzes) and non-assessed (e.g., video lectures) learning activities. In this paper, we propose Transition-Aware Multi-activity Knowledge Tracing (TAMKOT), which models knowledge transfer between learning materials, in addition to student knowledge, when students transition between and within assessed and non-assessed learning materials. TAMKOT is formulated as a deep recurrent multi-activity learning model that explicitly learns knowledge transfer by activating and learning a set of knowledge transfer matrices, one for each transition type between student activities. Accordingly, our model allows for representing each material type in a different yet transferrable latent space while maintaining student knowledge in a shared space. We evaluate our model on three real-world publicly available datasets and demonstrate TAMKOT's capability in predicting student performance and modeling knowledge transfer.