Figure 3 - uploaded by Y. Bengio
Content may be subject to copyright.
Source publication
Catastrophic forgetting is a problem faced by many machine learning models
and algorithms. When trained on one task, then trained on a second task, many
machine learning models "forget'' how to perform the first task. This is widely
believed to be a serious problem for neural networks. Here, we investigate the
extent to which the catastrophic forge...
Context in source publication
Context 1
... not improved for 100 epochs. After running all 25 randomly configured experiments for all 8 conditions, we make a possibilities frontier curve showing the minimum amount of test error on the new task obtaining for each amount of test error on the old task. Specifically, these plots are made by drawing a curve that traces out the lower left frontier of the cloud of points of all (old task test error, new task test error) pairs encountered by all 25 models during the course of training on the new task, with one point generated after each pass through the training set. Note that these test set errors are computed after training on only a subset of the training data, because we do not train on the validation set. It is possible to improve further by also training on the validation set, but we do not do so here because we only care about the relative performance of the different methods, not necessarily obtaining state of the art results. (Usually possibilities frontier curves are used in sce- narios where higher values are better, and the curves trace out the higher edge of a convex hull of scatterplot. Here, we are plotting error rates, so the lower values are better and the curves trace out the lower edge of a convex hull of a scatterplot. We used error rather than accuracy so that log scale plots would com- press regions of bad performance and expand regions of good performance, in order to highlight the differences between the best-performing methods. Note that the log scaling sometimes makes the convex regions apear non-convex) Many naturally occurring tasks are highly similar to each other in terms of the underlying structure that must be understood, but have the input presented in a different format. For example, consider learning to understand Italian after already learning to understand Spanish. Both tasks share the deeper underlying structure of being a natural language understanding problem, and fur- thermore, Italian and Spanish have similar grammar. However, the specific words in each language are different. A person learning Italian thus benefits from having a pre-existing representation of the general struc- ture of the language. The challenge is to learn to map the new words into these structures (e.g., to attach the Italian word “sei” to the pre-existing concept of the second person conjugation of the verb “to be”) without damaging the ability to understand Spanish. The ability to understand Spanish could diminish if the learning algorithm inadvertently modifies the more abstract definition of language in general (i.e., if neurons that were used for verb conjugation before now get re-purposed for plurality agreement) rather than exploiting the pre-existing definition, or if the learning algorithm removes the associations between individual Spanish words and these pre-existing concepts (e.g., if the net retains the concept of there being a second person conjugation of the verb “to be” but forgets that the Spanish word “eres” corresponds to it). To test this kind of learning problem, we designed a simple pair of tasks, where the tasks are the same, but with different ways of formatting the input. Specifically, we used MNIST classification, but with a different permutation of the pixels for the old task and the new task. Both tasks thus benefit from having concepts like penstroke detectors, or the concept of penstrokes being combined to form digits. However, the meaning of any individual pixel is different. The net must learn to associate new collections of pixels to penstrokes, without significantly disrupting the old higher level concepts, or erasing the old connections between pixels and penstrokes. The classification performance results are presented in Fig. 1. Using dropout improved the two-task validation set performance for all models on this task pair. We show the effect of dropout on the optimal model size in Fig. 2. While the nets were able to basically succeed at this task, we don’t believe that they did so by mapping different sets of pixels into pre-existing concepts. We visualized the first layer weights of the best net (in terms of combined validation set error) and their apparent semantics do not noticeably change between when training on the old task concludes and training on the new task begins. This suggests that the higher layers of the net changed to be able to ac- comodate a relatively arbitrary projection of the input, rather than remaining the same while the lower layers adapted to the new input format. We next considered what happens when the two tasks are not exactly the same, but semantically similar, and using the same input format. To test this case, we used sentiment analysis of two product categories of Amazon reviews (Blitzer et al., 2007) as the two tasks. The task is just to classify the text of a product review as positive or negative in sentiment. We used the same preprocessing as (Glorot et al., 2011b). The classification performance results are presented in Fig. 3. Using dropout improved the two-task validation set performance for all models on this task pair. We show the effect of dropout on the optimal model size in Fig. 6. We next considered what happens when the two tasks are semantically similar. To test this case, we used Amazon reviews as one task, and MNIST classification as another. In order to give both tasks the same output size, we used only two classes of the MNIST dataset. To give them the same validation set size, we randomly subsampled the remaining examples of the MNIST validation set (since the MNIST validation set was originally larger than the Amazon validation set, and we don’t want the estimate of the performance on the Amazon dataset to have higher variance than the MNIST one). The Amazon dataset as we preprocessed it earlier has 5,000 input features, while MNIST has only 784. To give the two tasks the same input size, we reduced the dimensionality of the Amazon data with PCA. Classification performance results are presented in Our experiments have shown that training with dropout is always beneficial, at least on the relatively small datasets we used in this paper. Dropout improved performance for all eight methods on all three task pairs. Dropout works the best in terms of performance on the new task, performance on the old task, and points along the tradeoff curve balancing these two extremes, for all three task pairs. Dropout’s resistance to forgetting may be explained in part by the large model sizes that can be trained with dropout. On the input-reformatted task pair and the similar task pair, dropout never decreased the size of the optimal model for any of the four activation functions we tried. However, dropout seems to have additional properties that can help prevent forgetting that we do not yet have an explanation for. On the dissimilar tasks experiment, dropout improved performance but reduced the size of the optimal model for most of the activation functions, and on the other task pairs, it occasionally had no effect on the optimal model size. The only recent previous work on catastrophic forget- ting(Srivastava et al., 2013) argued that the choice of activation function has a significant effect on the catastrophic forgetting properties of a net, and in particular that hard LWTA outperforms logistic sigmoid and rectified linear units in this respect when trained with stochastic gradient descent. In our more extensive experiments we found that the choice of activation function has a less consistent effect than the choice of training algorithm. When we performed experiments with different kinds of task pairs, we found that the ranking of the activation functions is very problem dependent. For example, logistic sigmoid is the worst under some conditions but the best under other conditions. This suggests that one should always cross-validate the choice of activation function, as long as it is computationally feasible. We also re- ject the idea that hard LWTA is particular resistant to catastrophic forgetting in general, or that it makes the standard SGD training algorithm more resistant to catastrophic forgetting. For example, when training with SGD on the input reformatting task pair, hard LWTA’s possibilities frontier is worse than all activation functions except sigmoid for most points along the curve. On the similar task pair, LWTA with SGD is the worst of all eight methods we considered, in terms of best performance on the new task, best performance on the old task, and in terms of attaining points close to the origin of the possibilities frontier plot. However, hard LWTA does perform the best in some cir- cumstances (it has the best performance on the new task for the dissimilar task pair ). This suggests that it is worth including hard LWTA as one of many activation functions in a hyperparameter search. LWTA is however never the leftmost point in any of our three task pairs, so it is probably only useful in sequential task settings where forgetting is an issue. When computational resources are too limited to experiment with multiple activation functions, we rec- ommend using the maxout activation function trained with dropout. This is the only method that appears on the lower-left frontier of the performance tradeoff plots for all three task pairs we considered. We would like to thank the developers of Theano (Bergstra et al., 2010; Bastien et al., 2012), Pylearn2 (Goodfellow et al., 2013a). We would also like to thank NSERC, Compute Canada, and Calcul Québec for providing computational resources. Ian Goodfellow is supported by the 2013 Google Fellowship in Deep Learning. Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Bergstra, James, Goodfellow, Ian J., Bergeron, Arnaud, Bouchard, Nicolas, and Bengio, Yoshua. Theano: new features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop, 2012. Bergstra, James and Bengio, Yoshua. Random search for hyper-parameter optimization. ...
Similar publications
Nonnegative Matrix Factorization (NMF) was first introduced as a low-rank
matrix approximation technique, and has enjoyed a wide area of applications.
Although NMF does not seem related to the clustering problem at first, it was
shown that they are closely linked. In this report, we provide a gentle
introduction to clustering and NMF before reviewi...
This paper presents a framework for exact discovery of the most interesting
sequential patterns. It combines (1) a novel definition of the expected support
for a sequential pattern - a concept on which most interestingness measures
directly rely - with (2) SkOPUS: a new branch-and-bound algorithm for the exact
discovery of top-k sequential patterns...
In this position paper, I first describe a new perspective on machine
learning (ML) by four basic problems (or levels), namely, "What to learn?",
"How to learn?", "What to evaluate?", and "What to adjust?". The paper stresses
more on the first level of "What to learn?", or "Learning Target Selection".
Towards this primary problem within the four le...
Interoperability among heterogeneous systems is a key chal-lenge in today's networked environment, which is charac-terised by continual change in aspects such as mobility and availability. Automated solutions appear then to be the only way to achieve interoperability with the needed level of flex-ibility and scalability. While necessary, the techni...
Entity resolution (ER), an important and common data cleaning problem, is
about detecting data duplicate representations for the same external entities,
and merging them into single representations. Relatively recently, declarative
rules called matching dependencies (MDs) have been proposed for specifying
similarity conditions under which attribute...
Citations
... LVT results are only reported after the last task, since no code has been released ual learning considers learning from a non-IID stream of data. Applying a naive finetuning approach to such data results in a phenomenon called catastrophic forgetting which results in a drastic drop in performance on previous tasks (Goodfellow et al., 2014). The main goal of continual learning algorithms is to maximize the stability-plasticity trade-off (Mermillod et al., 2013), i.e. to mitigate forgetting of previously learned classes while maintaining the plasticity required to learn new ones. ...
Vision transformers (ViTs) have achieved remarkable successes across a broad range of computer vision applications. As a consequence, there has been increasing interest in extending continual learning theory and techniques to ViT architectures. We propose a new method for exemplar-free class incremental training of ViTs. The main challenge of exemplar-free continual learning is maintaining plasticity of the learner without causing catastrophic forgetting of previously learned tasks. This is often achieved via exemplar replay which can help recalibrate previous task classifiers to the feature drift which occurs when learning new tasks. Exemplar replay, however, comes at the cost of retaining samples from previous tasks which for many applications may not be possible. To address the problem of continual ViT training, we first propose gated class-attention to minimize the drift in the final ViT transformer block. This mask-based gating is applied to class-attention mechanism of the last transformer block and strongly regulates the weights crucial for previous tasks. Importantly, gated class-attention does not require the task-ID during inference, which distinguishes it from other parameter isolation methods. Secondly, we propose a new method of feature drift compensation that accommodates feature drift in the backbone when learning new tasks. The combination of gated class-attention and cascaded feature drift compensation allows for plasticity towards new tasks while limiting forgetting of previous ones. Extensive experiments performed on CIFAR-100, Tiny-ImageNet and ImageNet100 demonstrate that our exemplar-free method obtains competitive results when compared to rehearsal based ViT methods.(Code:https://github.com/OcraM17/GCAB-CFDC)
... In the prompt, we explicitly provide the fault locations (i.e., activation functions and optimizer), while ChatGPT ignores the hint and locates another fault (i.e., the number of units) instead. This situation may happen due to the catastrophic forgetting (Ramasesh et al. 2022;Korbak et al. 2022;Ramasesh et al. 2020;Kemker et al. 2018;Goodfellow et al. 2013;Arora et al. 2019), which is a common problem for the long input sequence. ...
The emergence of large language models (LLMs) such as ChatGPT has revolutionized many fields. In particular, recent advances in LLMs have triggered various studies examining the use of these models for software development tasks, such as program repair, code understanding, and code generation. Prior studies have shown the capability of ChatGPT in repairing conventional programs. However, debugging deep learning (DL) programs poses unique challenges since the decision logic is not directly encoded in the source code. This requires LLMs to not only parse the source code syntactically but also understand the intention of DL programs. Therefore, ChatGPT’s capability in repairing DL programs remains unknown. To fill this gap, our study aims to answer three research questions: (1) Can ChatGPT debug DL programs effectively? (2) How can ChatGPT’s repair performance be improved by prompting? (3) In which way can dialogue help facilitate the repair? Our study analyzes the typical information that is useful for prompt design and suggests enhanced prompt templates that are more efficient for repairing DL programs. On top of them, we summarize the dual perspectives (i.e., advantages and disadvantages) of ChatGPT’s ability, such as its handling of API misuse and recommendation, and its shortcomings in identifying default parameters. Our findings indicate that ChatGPT has the potential to repair DL programs effectively and that prompt engineering and dialogue can further improve its performance by providing more code intention. We also identified the key intentions that can enhance ChatGPT’s program repairing capability.
... For instance, Peng et al. [14], building on Faster-RCNN, introduced a multi-network adaptive distillation method and designed an efficient, end-to-end incremental object detection model. This approach addresses the problem of catastrophic forgetting [15] in deep learning models, enabling the model to learn new knowledge without forgetting previously acquired knowledge. Ren et al. [16] replaced the fully connected layers in Fast/Faster-RCNN's classification head with convolutional layers. ...
Tomato plant diseases often first manifest on the leaves, making the detection of tomato leaf diseases particularly crucial for the tomato cultivation industry. However, conventional deep learning models face challenges such as large model sizes and slow detection speeds when deployed on resource-constrained platforms and agricultural machinery. This paper proposes a lightweight model for detecting tomato leaf diseases, named LT-YOLO, based on the YOLOv8n architecture. First, we enhance the C2f module into a RepViT Block (RVB) with decoupled token and channel mixers to reduce the cost of feature extraction. Next, we incorporate a novel Efficient Multi-Scale Attention (EMA) mechanism in the deeper layers of the backbone to improve detection of critical disease features. Additionally, we design a lightweight detection head, LT-Detect, using Partial Convolution (PConv) to significantly reduce the classification and localization costs during detection. Finally, we introduce a Receptive Field Block (RFB) in the shallow layers of the backbone to expand the model’s receptive field, enabling effective detection of diseases at various scales. The improved model reduces the number of parameters by 43% and the computational load by 50%. Additionally, it achieves a mean Average Precision (mAP) of 90.9% on a publicly available dataset containing 3641 images of tomato leaf diseases, with only a 0.7% decrease compared to the baseline model. This demonstrates that the model maintains excellent accuracy while being lightweight, making it suitable for rapid detection of tomato leaf diseases.
... Addressing the missing modality requires immediate compensation strategies. It is also important to consider catastrophic forgetting, which refers to the rapid forgetting of previously acquired information when new data is introduced (Goodfellow et al., 2013). In dynamic environments, the periodic absence of modalities can disrupt the learning process, making long-term learning and memory mechanisms crucial (Y. ...
Recently, machine learning technologies have been successfully applied across various fields. However, most existing machine learning models rely on unimodal data for information inference, which hinders their ability to generalize to complex application scenarios. This limitation has resulted in the development of multimodal learning, a field that integrates information from different modalities to enhance models' capabilities. However, data often suffers from missing or incomplete modalities in practical applications. This necessitates that models maintain robustness and effectively infer complete information in the presence of missing modalities. The emerging research direction of incomplete multimodal learning (IML) aims to facilitate effective learning from incomplete multimodal training sets, ensuring that models can dynamically and robustly address new instances with arbitrary missing modalities during the testing phase. This paper offers a comprehensive review of methods based on IML. It categorizes existing approaches based on their information sources into two main types: based on internal information and external information methods. These categories are further subdivided into data-based, feature-based, knowledge transfer-based, graph knowledge enhancement-based, and human-in-the-loop-based methods. The paper conducts comparative analyses from two perspectives: comparisons among similar methods and comparisons among different types of methods. Finally, it offers insights into the research trends in IML.
... Fine-tuning of pretrained models allows the development of more accurate models with a small amount of data and a greatly reduced training time (Goodfellow et al., 2013). During fine-tuning the user performs additional training with a dataset containing many targeted examples related to a given task, aiming to increase performance in this domain. ...
Pretrained foundation models learn embeddings that can be used for a wide range of downstream tasks. These embeddings optimise general performance, and if insufficiently accurate at a specific task the model can be fine-tuned to improve performance. For all current methodologies this operation necessarily degrades performance on all out-of-distribution tasks. In this work we present 'fill-tuning', a novel methodology to generate datasets for continued pretraining of foundation models that are not suited to a particular downstream task, but instead aim to correct poor regions of the embedding. We present the application of roughness analysis to latent space topologies and illustrate how it can be used to propose data that will be most valuable to improving the embedding. We apply fill-tuning to a set of state-of-the-art materials foundation models trained on data points and show model improvement of almost 1% in all downstream tasks with the addition of only 100 data points. This method provides a route to the general improvement of foundation models at the computational cost of fine-tuning.
... At the end of each period, the EI can be trained solely on this new domain dataset, which is then discarded to free up memory resources on the ED, preparing for subsequent domain datasets. Nevertheless, such sequential training suffers from the catastrophic forgetting problem [20], [ tends to rapidly forget its knowledge of previous domains when trained on the current domain dataset. This problem severely hinders the EI from gaining the cross-domain sensing capability and, consequently, limits its potential for enabling ubiquitous sensing applications in resource-constrained EDs. ...
... To achieve this goal, we employ the herding method [44] to select the rest of (1 − β)E exemplars so that the overall mean of the selected feature vectors has a minimal distance fromf c . It is equivalent to solving the following optimization problem: (20) is a challenging combinational optimization problem, it can be solved approximately using a greedy algorithm. Specifically, we employ the nearest-classmean algorithm in [44], which iteratively finds each feature vector to minimize the current distance. ...
... Consequently, the CSI data corresponding to the E feature vectors obtained by (19) and (20) are selected as the exemplars, forming the core-set of class c, i.e., We then proceed to deriving the regularization function. Since the empirical CE loss approximates the negative logarithmic a posterior distribution, it is intuitive to adopt the CE loss over the core-set as the regularization function. ...
In wireless networks with integrated sensing and communications (ISAC), edge intelligence (EI) is expected to be developed at edge devices (ED) for sensing user activities based on channel state information (CSI). However, due to the CSI being highly specific to users' characteristics, the CSI-activity relationship is notoriously domain dependent, essentially demanding EI to learn sufficient datasets from various domains in order to gain cross-domain sensing capability. This poses a crucial challenge owing to the EDs' limited resources, for which storing datasets across all domains will be a significant burden. In this paper, we propose the EdgeCL framework, enabling the EI to continually learn-then-discard each incoming dataset, while remaining resilient to catastrophic forgetting. We design a transformer-based discriminator for handling sequences of noisy and nonequispaced CSI samples. Besides, we propose a distilled core-set based knowledge retention method with robustness-enhanced optimization to train the discriminator, preserving its performance for previous domains while preventing future forgetting. Experimental evaluations show that EdgeCL achieves 89% of performance compared to cumulative training while consuming only 3% of its memory, mitigating forgetting by 79%.
... To evaluate the performance of quantum-inspired swarm optimization in addressing complex problem-solving, we consider the following datasets: a hierarchical text generation dataset that facilitates learning representations of dialogue messages [31], a dataset offering human-centered face similarity judgments designed for creating low-dimensional embeddings [32], a study on catastrophic forgetting in neural networks highlighting the utility of the dropout algorithm [33], the TWEETQA dataset for question answering focused on social media content [34], and an analysis of single-layer networks that underscores the importance of hidden nodes and feature extraction for high performance in unsupervised learning tasks [35]. ...
Quantum-Inspired Swarm Optimization (QISO) represents a novel approach to tackling complex problem-solving across various fields by merging swarm intelligence with quantum principles. This method utilizes collective agent behaviors to navigate solution landscapes, addressing challenges that traditional optimization methods often encounter. In specific applications like engineering design, data clustering, and scheduling, QISO outperforms conventional approaches, achieving notable reductions in computation time and improvements in solution quality. Experimental evaluations substantiate the advantages of QISO, highlighting its versatility and resilience in intricate environments. By integrating quantum concepts into swarm optimization, QISO introduces innovative strategies that significantly enhance computational performance, particularly in complex and dynamic problem areas.
... CL [18], also known as life-long learning, describes the ability of a network to learn a stream of data continuously. In CL, the data stream is often divided into distinct tasks, where each task represents a set of data learned by the model at a specifc point in time. ...
The identification of surface damage and structural components is critical for structural health monitoring (SHM) in order to evaluate building safety. Recently, deep neural networks (DNNs)–based approaches have emerged rapidly. However, the existing approaches often encounter catastrophic forgetting when the trained model is used to learn new classes of interest. Conventionally, joint training of the network on both the previous and new data is employed, which is time-consuming and demanding for computation and memory storage. To address this issue, we propose a new approach that integrates two continual learning (CL) algorithms, i.e., elastic weight consolidation (EWC) and learning without forgetting (LwF), denoted as EWCLwF. We also investigate two scenarios for a comprehensive discussion, incrementally learning the classes with similar versus dissimilar data characteristics. Results have demonstrated that EWCLwF requires significantly less training time and data storage compared to joint training, and the average accuracy is enhanced by 0.7%–4.5% compared against other baseline references in both scenarios. Furthermore, our findings reveal that all CL-based approaches benefit from similar data characteristics, while joint training not only fails to benefit but performs even worse, which indicates a scenario that can emphasize the advantage of our proposed approach. The outcome of this study will enhance the long-term monitoring of progressively increasing learning classes in SHM, leading to more efficient usage and management of computing resources.
... Humans learn continuously in such environments, but this can lead to catastrophic forgetting, where previously learned tasks are disrupted by the brain's neuro system. Neural variability helps compensate for both accuracy and plasticity in humans [30]. Online learning techniques face the same phenomenon of catastrophic forgetting when working with dynamic data [16]. ...
... Online learning techniques face the same phenomenon of catastrophic forgetting when working with dynamic data [16]. This occurs when the network modifies information related to a previous task due to continuous training on a new task [30][31][32][33][34]. Figure 4 depicts catastrophic forgetting, where the network Fig. 2b (c) Output labels. Due to access feeding of label B, the network forgets label A identity forgets class A due to the recursive occurrence of class B. Section 3 elaborates on the literature related to catastrophic forgetting. ...
... As a result, the memory is optimized to retain important knowledge and forget irrelevant or outdated information. Goodfellow et al. [30] investigated the selection of appropriate learning algorithms and activation functions for different tasks and relationships between tasks to mitigate the catastrophic forgetting effect. They examined the relationship between tasks and found that dropout is the most effective training algorithm for modern feed-forward neural networks. ...
In an era defined by the relentless influx of data from diverse sources, the ability to harness and extract valuable insights from streaming data has become paramount. The rapidly evolving realm of online learning techniques is tailored specifically for the unique challenges posed by streaming data. As the digital world continues to generate vast torrents of real-time data, understanding and effectively utilizing online learning approaches are pivotal for staying ahead in various domains. One of the primary goals of online learning is to continuously update the model with the most recent data trends while maintaining and improving the accuracy of previous trends. Based on the various types of feedback, online learning tasks can be divided into three categories: learning with full feedback, learning with limited feedback, and learning without feedback. This survey aims to identify and analyze the key challenges associated with online learning with full feedback, including concept drift, catastrophic forgetting, skewed learning, and network adaptation, while the other existing reviews mainly focus on a single challenge or two without considering other scenarios. This article also discusses the application and ethical implications of online learning. The results of this survey provide valuable insights for researchers and instructional designers seeking to create effective online learning experiences that incorporate full feedback while addressing the associated challenges. In the end, some conclusions, remarks, and future directions for the research community are provided based on the findings of this review.
... Therefore, we simulated streaming learning, where new information arrives in non-stationary online data streams without clearly defined task boundaries, leading to severe plasticity loss in neural networks. Following previous studies [29,36], we adopted two typical datasets: Input Permuted MNIST [55] and Label Permuted MNIST [29]. We generated 10,000 samples per task, with a total of 100 tasks in sequence. ...
Artificial neural networks face the stability-plasticity dilemma in continual learning, while the brain can maintain memories and remain adaptable. However, the biological strategies for continual learning and their potential to inspire learning algorithms in neural networks are poorly understood. This study presents a minimal model of the fly olfactory circuit to investigate the biological strategies that support continual odor learning. We introduce the fly olfactory circuit as a plug-and-play component, termed the Fly Model, which can integrate with modern machine learning methods to address this dilemma. Our findings demonstrate that the Fly Model enhances both memory stability and learning plasticity, overcoming the limitations of current continual learning strategies. We validated its effectiveness across various challenging continual learning scenarios using commonly used datasets. The fly olfactory system serves as an elegant biological circuit for lifelong learning, offering a module that enhances continual learning with minimal additional computational cost for machine learning.
