Artificial learning systems aspire to mimic human intelligence by continually learning from a stream of tasks without forgetting past knowledge. One way to enable such learning is to store past experiences in the form of input examples in episodic memory and replay them when learning new tasks. However, performance of such method suffers as the size of the memory becomes smaller. In this paper, we propose a new approach for experience replay, where we select the past experiences by looking at the saliency maps which provide visual explanations for the model's decision. Guided by these saliency maps, we pack the memory with only the parts or patches of the input images important for the model's prediction. While learning a new task, we replay these memory patches with appropriate zero-padding to remind the model about its past decisions. We evaluate our algorithm on CIFAR-100, miniImageNet and CUB datasets and report better performance than the state-of-the-art approaches. With qualitative and quantitative analyses we show that our method captures richer summaries of past experiences without any memory increase, and hence performs well with small episodic memory.

The opaque nature and unexplained behavior of transformer-based language models (LMs) have spurred a wide interest in interpreting their predictions. However, current interpretation methods mostly focus on probing models from outside, executing behavioral tests, and analyzing salience input features, while the internal prediction construction process is largely not understood. In this work, we introduce LM-Debugger, an interactive debugger tool for transformer-based LMs, which provides a fine-grained interpretation of the model's internal prediction process, as well as a powerful framework for intervening in LM behavior. For its backbone, LM-Debugger relies on a recent method that interprets the inner token representations and their updates by the feed-forward layers in the vocabulary space. We demonstrate the utility of LM-Debugger for single-prediction debugging, by inspecting the internal disambiguation process done by GPT2. Moreover, we show how easily LM-Debugger allows to shift model behavior in a direction of the user's choice, by identifying a few vectors in the network and inducing effective interventions to the prediction process. We release LM-Debugger as an open-source tool and a demo over GPT2 models.

Knowledge graphs, as the cornerstone of many AI applications, usually face serious incompleteness problems. In recent years, there have been many efforts to study automatic knowledge graph completion (KGC), most of which use existing knowledge to infer new knowledge. However, in our experiments, we find that not all relations can be obtained by inference, which constrains the performance of existing models. To alleviate this problem, we propose a new model based on information retrieval and reading comprehension, namely IR4KGC. Specifically, we pre-train a knowledge-based information retrieval module that can retrieve documents related to the triples to be completed. Then, the retrieved documents are handed over to the reading comprehension module to generate the predicted answers. In experiments, we find that our model can well solve relations that cannot be inferred from existing knowledge, and achieve good results on KGC datasets.

Retraining deep neural networks when new data arrives is typically computationally expensive. Moreover, certain applications do not allow such costly retraining due to time or computational constraints. Fast Bayesian updates are a possible solution to this issue. Therefore, we propose a Bayesian update based on Monte-Carlo samples and a last-layer Laplace approximation for different Bayesian neural network types, i.e., Dropout, Ensemble, and Spectral Normalized Neural Gaussian Process (SNGP). In a large-scale evaluation study, we show that our updates combined with SNGP represent a fast and competitive alternative to costly retraining. As a use case, we combine the Bayesian updates for SNGP with different sequential query strategies to exemplarily demonstrate their improved selection performance in active learning. Extending a dataset with new samples to train a deep learning model typically poses two problems. Updating a trained model may cause catastrophic forgetting while retraining may require high computational effort. Although the generalization performance typically justifies exhaustive retraining procedures, in some applications, retraining is not possible due to, for example, (1) the high number of retraining procedures in applications where data arrives sequentially and immediate updates are beneficial, e.g., in active learning (Settles, 2009) or when working on data streams (Sahoo et al., 2018), (2) the lack of computational power, e.g., for execution on embedded hardware (Taylor et al., 2018), (3) privacy reasons, e.g., when new data cannot be sent to distributed computing units (Taylor et al., 2018). Therefore, we suggest using Bayesian neural networks (BNNs, Fortuin, 2022) in the above examples as they not only provide additional uncertainty estimates or out-of-distribution detection capabilities but also allow updating the predictions with additional data without retraining the network (Kirsch et al., 2022). In this article, we develop a fast Bayesian update algorithm for BNNs. Figure 1 (left) shows the idea of sampling an ensemble of probabilistic hypotheses, each representing a possible true solution for the learning task (white samples).

We propose an ensemble learning framework with Poisson sub-sampling to effectively train a collection of teacher models to issue some differential privacy (DP) guarantee for training data. Through boosting under DP, a student model derived from the training data suffers little model degradation from the models trained with no privacy protection. Our proposed solution leverages upon two mechanisms, namely: (i) a privacy budget amplification via Poisson sub-sampling to train a target prediction model that requires less noise to achieve a same level of privacy budget, and (ii) a combination of the sub-sampling technique and an ensemble teacher-student learning framework that introduces DP-preserving noise at the output of the teacher models and transfers DP-preserving properties via noisy labels. Privacy-preserving student models are then trained with the noisy labels to learn the knowledge with DP-protection from the teacher model ensemble. Experimental evidences on spoken command recognition and continuous speech recognition of Mandarin speech show that our proposed framework greatly outperforms existing DP-preserving algorithms in both speech processing tasks.

In this paper, we study the problem of bandits with knapsacks (BwK) in a non-stationary environment. The BwK problem generalizes the multi-arm bandit (MAB) problem to model the resource consumption associated with playing each arm. At each time, the decision maker/player chooses to play an arm, and s/he will receive a reward and consume certain amount of resource from each of the multiple resource types. The objective is to maximize the cumulative reward over a finite horizon subject to some knapsack constraints on the resources. Existing works study the BwK problem under either a stochastic or adversarial environment. Our paper considers a non-stationary environment which continuously interpolates between these two extremes. We first show that the traditional notion of variation budget is insufficient to characterize the non-stationarity of the BwK problem for a sublinear regret due to the presence of the constraints, and then we propose a new notion of global non-stationarity measure. We employ both non-stationarity measures to derive upper and lower bounds for the problem. Our results are based on a primal-dual analysis of the underlying linear programs and highlight the interplay between the constraints and the non-stationarity. Finally, we also extend the non-stationarity measure to the problem of online convex optimization with constraints and obtain new regret bounds accordingly.

Creativity, i.e., the process of generating and developing fresh and original ideas or products that are useful or effective, is a valuable skill in a variety of domains. Creativity is called an essential 21st-century skill that should be taught in schools. The use of educational technology to promote creativity is an active study field, as evidenced by several studies linking creativity in the classroom to beneficial learning outcomes. Despite the burgeoning body of research on adaptive technology for education, mining creative thinking patterns from educational data remains a challenging task. In this paper, to address this challenge, we put the first step towards formalizing educational knowledge by constructing a domain-specific Knowledge Base to identify essential concepts, facts, and assumptions in identifying creative patterns. We then introduce a pipeline to contextualize the raw educational data, such as assessments and class activities. Finally, we present a rule-based approach to learning from the Knowledge Base, and facilitate mining creative thinking patterns from contextualized data and knowledge. We evaluate our approach with real-world datasets and highlight how the proposed pipeline can help instructors understand creative thinking patterns from students' activities and assessment tasks.

Sampling methods, as important inference and learning techniques, are typically designed for unconstrained domains. However, constraints are ubiquitous in machine learning problems, such as those on safety, fairness, robustness, and many other properties that must be satisfied to apply sampling results in real-life applications. Enforcing these constraints often leads to implicitly-defined manifolds, making efficient sampling with constraints very challenging. In this paper, we propose a new variational framework with a designed orthogonal-space gradient flow (O-Gradient) for sampling on a manifold $\mathcal{G}_0$ defined by general equality constraints. O-Gradient decomposes the gradient into two parts: one decreases the distance to $\mathcal{G}_0$ and the other decreases the KL divergence in the orthogonal space. While most existing manifold sampling methods require initialization on $\mathcal{G}_0$, O-Gradient does not require such prior knowledge. We prove that O-Gradient converges to the target constrained distribution with rate $\widetilde{O}(1/\text{the number of iterations})$ under mild conditions. Our proof relies on a new Stein characterization of conditional measure which could be of independent interest. We implement O-Gradient through both Langevin dynamics and Stein variational gradient descent and demonstrate its effectiveness in various experiments, including Bayesian deep neural networks.

In real-world applications, the process generating the data might suffer from nonstationary effects (e.g., due to seasonality, faults affecting sensors or actuators, and changes in the users' behaviour). These changes, often called concept drift, might induce severe (potentially catastrophic) impacts on trained learning models that become obsolete over time, and inadequate to solve the task at hand. Learning in presence of concept drift aims at designing machine and deep learning models that are able to track and adapt to concept drift. Typically, techniques to handle concept drift are either active or passive, and traditionally, these have been considered to be mutually exclusive. Active techniques use an explicit drift detection mechanism, and re-train the learning algorithm when concept drift is detected. Passive techniques use an implicit method to deal with drift, and continually update the model using incremental learning. Differently from what present in the literature, we propose a hybrid alternative which merges the two approaches, hence, leveraging on their advantages. The proposed method called Hybrid-Adaptive REBAlancing (HAREBA) significantly outperforms strong baselines and state-of-the-art methods in terms of learning quality and speed; we experiment how it is effective under severe class imbalance levels too.

Multi-task learning (MTL) aims to improve the generalization performance of multiple tasks by exploiting the shared factors among them. Various metrics (e.g., F-score, Area Under the ROC Curve) are used to evaluate the performances of MTL methods. Most existing MTL methods try to minimize either the misclassified errors for classification or the mean squared errors for regression. In this paper, we propose a method to directly optimize the evaluation metrics for a large family of MTL problems. The formulation of MTL that directly optimizes evaluation metrics is the combination of two parts: (1) a regularizer defined on the weight matrix over all tasks, in order to capture the relatedness of these tasks; (2) a sum of multiple structured hinge losses, each corresponding to a surrogate of some evaluation metric on one task. This formulation is challenging in optimization because both of its parts are non-smooth. To tackle this issue, we propose a novel optimization procedure based on the alternating direction scheme of multipliers, where we decompose the whole optimization problem into a sub-problem corresponding to the regularizer and another sub-problem corresponding to the structured hinge losses. For a large family of MTL problems, the first sub-problem has closed-form solutions. To solve the second sub-problem, we propose an efficient primal-dual algorithm via coordinate ascent. Extensive evaluation results demonstrate that, in a large family of MTL problems, the proposed MTL method of directly optimization evaluation metrics has superior performance gains against the corresponding baseline methods.