Memory Consumption Research Articles

Strongly solved Ostle: calculating a strong solution helps compose high-quality puzzles for recent games.

Pure strategy board games such as chess are popular intellectual activities, and solving them is a challenging task in computer science. In addition to traditional games, many new board games have gained popularity in recent years. Ostle is one such unsolved game published in 2017. It is based on simple rules but is highly competitive. It is a two-player zero-sum game with perfect information in which the game-theoretical values of all game states can be obtained. In this study, we strongly solved Ostle by retrograde analysis. Utilizing various known techniques, including bitboards and succinct indexable dictionaries, significantly reduced the memory consumption in the analyses. We confirmed that the initial position is a draw and found some fundamental properties of Ostle. Additionally, we manually composed a tactical Ostle puzzle with the help of extracted outputs of the analyses. The result demonstrates that solving recent games provided helpful information to compose high-quality problems.

Targeted mining of top-[formula omitted] high utility itemsets

Finding high-importance patterns in data is an emerging data mining task known as High-utility itemset mining (HUIM). Given a minimum utility threshold, a HUIM algorithm extracts all the high-utility itemsets (HUIs) whose utility values are not less than the threshold. This can reveal a wealth of useful information, but the precise needs of users are not well taken into account. In particular, users often want to focus on patterns that have some specific items rather than find all patterns. To overcome that difficulty, targeted mining has emerged, focusing on user preferences, but only preliminary work has been conducted. For example, the targeted high-utility itemset querying algorithm (TargetUM) was proposed, which uses a lexicographic tree to query itemsets containing a target pattern. However, selecting the minimum utility threshold is difficult when the user is not familiar with the processed database. As a solution, this paper formulates the task of targeted mining of the top-k high-utility itemsets and proposes an efficient algorithm called TMKU based on the TargetUM algorithm to discover the top-k target high-utility itemsets (top-k THUIs). At the same time, several pruning strategies are used to reduce memory consumption and execution time. Extensive experiments show that the proposed TMKU algorithm has good performance on real and synthetic datasets.

A scalable second order optimizer with an adaptive trust region for neural networks

We introduce Tadam (Trust region ADAptive Moment estimation), a new optimizer based on the trust region of the second-order approximation of the loss using the Fisher information matrix. Despite the enhanced gradient estimations offered by second-order approximations, their practical implementation requires sizable batch sizes to estimate the second-order approximation matrices and perform matrix inversions. Consequently, integrating second-order approximations entails additional memory consumption and imposes substantial computational demands due to the inversion of large matrices. In light of these challenges, we have devised a second-order approximation algorithm that mitigates these issues by judiciously approximating the pertinent large matrix, requiring only a marginal increase in memory usage while minimizing the computational burden. Tadam approximates the loss up to the second order using the Fisher information matrix. Since estimating the Fisher information matrix is expensive in both memory and time, Tadam approximates the Fisher information matrix and reduces the computational burdens to the O(N) level. Furthermore, Tadam employs an adaptive trust region scheme to reduce approximate errors and guarantee stability. Tadam evaluates how well it minimizes the loss function and uses this information to adjust the trust region dynamically. In addition, Tadam adjusts the learning rate internally, even if we provide the learning rate as a fixed constant. We run several experiments to measure Tadam’s performance against Adam, AMSGrad, Radam, and Nadam, which have the same space and time complexity as Tadam. The test results show that Tadam outperforms the benchmarks and finds reasonable solutions fast and stably.

Video anomaly detection based on cross-frame prediction mechanism and spatio-temporal memory-enhanced pseudo-3D encoder

Intelligent video anomaly detection (VAD) methods play a crucial role in conserving human resources, reducing the financial burden on governments, and promptly and accurately identifying abnormal behaviors. Frame prediction, which performs VAD by reasonably predicting normal video and distorting predicted anomalous video, is a popular and efficient method. Although Auto-Encoders (AE) show excellent performance in video frame prediction methods, the ability of these methods to use temporal information and poorly reconstruct anomalous videos is insufficient. To improve the shortcomings of AE in VAD, we propose a VAD method based on the cross-frame prediction mechanism and the spatio-temporal memory-enhanced pseudo-3D encoder. This method significantly enhances the recognition capability of abnormal activities in surveillance videos. Firstly, we use the prediction mechanism that uses multiple past frames with intervals to predict future frames. It can broaden the extra information and limit the memory consumption. Then we design the pseudo-3D encoder to encode the spatio-temporal information in videos, avoiding the problems that the 2D encoder cannot obtain the temporal dimensional information and the 3D encoder has complicated structure and overmuch parameters. Finally, we design the spatio-temporal memory block with three loss functions to store the spatio-temporal information of normal videos, which can expand the predicted differences between normal and abnormal examples. Experiments on UCSD Ped2, CUHK Avenue and ShanghaiTech datasets experimentally show that the proposed method achieves 99.4%, 90.5% and 74.3% of the AUC values. Our method shows excellent performance among single-stage semi-supervised anomaly detection methods.

A dynamical particle merging and splitting algorithm for Particle-In-Cell simulations

In Particle-In-Cell simulation, macro-particles represent clusters of numerous physical particles. Rational merging or splitting of these macro-particle clusters can significantly improve the efficiency of simulation and reduce unnecessary computational requirements and memory consumption. Here, we propose a dynamical particle merging and splitting algorithm. For macro-particle in different density regions, regional macro-particle clusters are obtained by dividing them with the Minkowski metric function. Dynamic merging and splitting of macro-particles is achieved by dynamically changing the number of macro-particle clusters to reduce the number of macro-particles without substantially distorting the physical description of the system. To test the capability of the algorithm, we also compare its performance under three types of classical plasma cases: two-stream instability (1D), QED cascades (2D), and magnetic shower (3D). The results show a good agreement with expectations.

TGMIL: A hybrid multi-instance learning model based on the Transformer and the Graph Attention Network for whole-slide images classification of renal cell carcinoma

The pathological diagnosis of renal cell carcinoma is crucial for treatment. Currently, the multi-instance learning method is commonly used for whole-slide image classification of renal cell carcinoma, which is mainly based on the assumption of independent identical distribution. But this is inconsistent with the need to consider the correlation between different instances in the diagnosis process. Furthermore, the problem of high resource consumption of pathology images is still urgent to be solved. Therefore, we propose a new multi-instance learning method to solve this problem. In this study, we proposed a hybrid multi-instance learning model based on the Transformer and the Graph Attention Network, called TGMIL, to achieve whole-slide image of renal cell carcinoma classification without pixel-level annotation or region of interest extraction. Our approach is divided into three steps. First, we designed a feature pyramid with the multiple low magnifications of whole-slide image named MMFP. It makes the model incorporates richer information, and reduces memory consumption as well as training time compared to the highest magnification. Second, TGMIL amalgamates the Transformer and the Graph Attention's capabilities, adeptly addressing the loss of instance contextual and spatial. Within the Graph Attention network stream, an easy and efficient approach employing max pooling and mean pooling yields the graph adjacency matrix, devoid of extra memory consumption. Finally, the outputs of two streams of TGMIL are aggregated to achieve the classification of renal cell carcinoma. On the TCGA-RCC validation set, a public dataset for renal cell carcinoma, the area under a receiver operating characteristic (ROC) curve (AUC) and accuracy of TGMIL were 0.98±0.0015,0.9191±0.0062, respectively. It showcased remarkable proficiency on the private validation set of renal cell carcinoma pathology images, attaining AUC of 0.9386±0.0162 and ACC of 0.9197±0.0124. Furthermore, on the public breast cancer whole-slide image test dataset, CAMELYON 16, our model showed good classification performance with an accuracy of 0.8792. TGMIL models the diagnostic process of pathologists and shows good classification performance on multiple datasets. Concurrently, the MMFP module efficiently diminishes resource requirements, offering a novel angle for exploring computational pathology images.

A novel image deep learning-based sub-centimeter pulmonary nodule management algorithm to expedite resection of the malignant and avoid over-diagnosis of the benign.

With the popularization of chest computed tomography (CT) screening, there are more sub-centimeter (≤ 1cm) pulmonary nodules (SCPNs) requiring further diagnostic workup. This area represents an important opportunity to optimize the SCPN management algorithm avoiding "one-size fits all" approach. One critical problem is how to learn the discriminative multi-view characteristics and the unique context of each SCPN. Here, we propose a multi-view coupled self-attention module (MVCS) to capture the global spatial context of the CT image through modeling the association order of space and dimension. Compared with existing self-attention methods, MVCS uses less memory consumption and computational complexity, unearths dimension correlations that previous methods have not found, and is easy to integrate with other frameworks. In total, a public dataset LUNA16 from LIDC-IDRI, 1319 SCPNs from 1069 patients presenting to a major referral center, and 160 SCPNs from 137 patients from three other major centers were analyzed to pre-train, train, and validate the model. Experimental results showed that performance outperforms the state-of-the-art models in terms of accuracy and stability and is comparable to that of human experts in classifying precancerous lesions and invasive adenocarcinoma. We also provide a fusion MVCS network (MVCSN) by combining the CT image with the clinical characteristics and radiographic features of patients. This tool may ultimately aid in expediting resection of the malignant SCPNs and avoid over-diagnosis of the benign ones, resulting in improved management outcomes. In the diagnosis of sub-centimeter lung adenocarcinoma, fusion MVCSN can help doctors improve work efficiency and guide their treatment decisions to a certain extent. • Advances in computed tomography (CT) not only increase the number of nodules detected, but also the nodules that are identified are smaller, such as sub-centimeter pulmonary nodules (SCPNs). • We propose a multi-view coupled self-attention module (MVCS), which could model spatial and dimensional correlations sequentially for learning global spatial contexts, which is better than other attention mechanisms. • MVCS uses fewer huge memory consumption and computational complexity than the existing self-attention methods when dealing with 3D medical image data. Additionally, it reaches promising accuracy for SCPNs' malignancy evaluation and has lower training cost than other models.

Self-supervised feature matched virtual try-on

Abstract Virtual try-on is a technology that enables users to preview the effect of wearing a target garment without wearing the actual garment. However, existing image-based virtual try-on methods often require additional human parsing or segmentation operations to generate intermediate representations required for garment deformation and texture fusion. These operations not only increase the computational complexity and memory consumption, but also limit the real-time and portability of virtual try-on. Additionally, inaccurate parsing results can lead to misleading final generated images. To overcome these challenges, we propose a self-supervised feature matched virtual try-on network, which can directly generate high-quality try-on results from human body images and target clothing images without any additional input. Specifically, we design an optical flow warp module, which focuses on the optical flow changes between the person image and the clothing image to achieve accurate clothing alignment and deformation. Furthermore, a feature refine warp module is designed to enhance the features of the extracted optical flow information and the original character segmentation and analysis operations, reducing the influence of background clutter features on the content, and ensuring that the wrinkles and deformation of the replacement clothes are close to the original clothes. The feature match module is developed to calculate the feature matching loss of the converted clothing and the generated results of the teacher network and the student network, and the corresponding knowledge is distilled and passed to the student network to assist in self-supervised training. We conduct experiments on the VITON dataset and show that our model can generate high quality and high resolution, and our proposed method outperforms the state-of-the-art virtual try-on methods both qualitatively and quantitatively.

Human activity recognition method using joint deep learning and acceleration signal

Many studies have been conducted on human activity recognition (HAR) in the last decade. Accordingly, deep learning (DL) algorithms have been given more attention in terms of classification of human daily activities. Deep neural networks (DNNs) compute and extract complex features on voluminous data through some hidden layers that require large memory and powerful graphics processing units (GPUs). So, this study proposes a new joint learning (JL) approach to classify human activities using inertial sensors. To this end, a large complex donor model based on a convolutional neural network (CNN) is used to transfer knowledge to a smaller model based on CNN referred to as the acceptor model. The acceptor model can be deployed on mobile devices and low-power hardware due to decreased computing costs and memory consumption. The wireless sensor data mining (WISDM) dataset is used to test the proposed model. According to the experimental results, the HAR system based on the JL algorithm outperforms than other methods.

Symbolic Minimization on Relational Data

The current wave of AI is heavily driven by data, especially for cognitive capabilities. Minimization of data semantics not only reveals core information but also becomes a guide in a wide range of domains. However, scalability is theoretically weak in pure semantic methodologies. In order to cooperate with large DBs, expressiveness is over-sacrificed in existing techniques. Thus, the quality of discovered patterns and redundancies are far from satisfactory. In this article, we formalize symbolic minimization on relational DBs and prove its NP-Completeness. A lossless technique is proposed by inducing generic first-order Horn rules that infer a subset of records from the others. More importantly, we further improve the scalability via effective caching and pruning without sacrificing the expressiveness of first-order Horn rules. A concrete system is implemented and comprehensively evaluated. Experiments show that our technique removes up to 70% contents and outperforms the state-of-the-art on minimization and scalability. The optimizations reduce up to 96% memory consumption and accelerate the performance by two orders. Our technique shows the practicality of pure semantic approaches in database mining.

Model-Free Neural Counterfactual Regret Minimization With Bootstrap Learning

Counterfactual Regret Minimization (CFR) has achieved many fascinating results in solving large-scale Imperfect Information Games (IIGs). Neural network approximation CFR (neural CFR) is one of the promising techniques that can reduce computation and memory consumption by generalizing decision information between similar states. Current neural CFR algorithms have to approximate cumulative regrets. However, efficient and accurate approximation in a large-scale IIG is still a tough challenge. In this paper, a new CFR variant, Recursive CFR (ReCFR), is proposed. In ReCFR, Recursive Substitute Values (RSVs) are learned and used to replace cumulative regrets. It is proven that ReCFR can converge to a Nash equilibrium at a rate of $O({1}/{\sqrt{T}})$. Based on ReCFR, a new model-free neural CFR with bootstrap learning, Neural ReCFR-B, is proposed. Due to the recursive and non-cumulative nature of RSVs, Neural ReCFR-B has lower-variance training targets than other neural CFRs. Experimental results show that Neural ReCFR-B is competitive with the state-of-the-art neural CFR algorithms at a much lower training cost.

A LIGHTWEIGHT ENCRYPTION ALGORITHM TO SECURE IOT DEVICES

Lightweight cryptography is an effective solution for ensuring confidentiality and privacy in resource-constrained environments, particularly in Internet of Things (IoT) systems. These systems consist of a large number of IoT devices, such as sensor nodes and embedded devices that handle sensitive data and are connected via the Internet, making them susceptible to various types of attacks. Considering the constrained processing capability, memory size, and energy of these devices, this article proposes a lightweight block cipher that aims to provide security for IoT devices, with a focus on achieving efficient software implementation, particularly in devices with 32-bit processors. The structural design of the suggested algorithm is based on Type-2 Generalized Feistel Networks (GFN), which are based on four branches of 32-bit and simple key scheduling. To enhance the diffusion and strength of the GFN structure, a partial permutation layer has been added to the cipher. Moreover, to reduce the implementation cost, a simple round function has been designed using low-cost operations like Addition, XOR, and Rotation. Security of the suggested algorithm has been validated through security analysis, and the results show meeting the criteria for avalanche effect with 50.97% Key Sensitivity and 50.58% Plaintext Sensitivity, and the correlation coefficienttest results indicate a weak correlation between the plaintext and ciphertext, this makes it difficult to retrieve the plaintext from the ciphertext. When designing lightweight algorithms, one of the primary objectives is to enhance performance. To evaluate the algorithm's performance, we tested its execution time, code size, memory consumption, and throughput on real IoT devices, such as ESP8266 and ESP32, as an embedded cryptographic module. Based on the experimental results and security analysis, the suggested algorithm showed a good grade of security and higher performance in encryption and decryption operations than other lightweight algorithms, which makes it capable of providing a framework for securing data in the restricted IoT environment.

Data-Driven Tabulation for Chemistry Integration Using Recurrent Neural Networks.

Due to the wide range of time scales involved in the ordinary differential equations (ODEs) describing chemical reaction kinetics, multidimensional numerical simulation of chemical reactive flows using detailed combustion mechanisms is computationally expensive. To confront this issue, this article presents an economic data-driven tabulation algorithm for fast combustion chemistry integration. It uses the recurrent neural networks (RNNs) to construct the tabulation from a series of current and past states to the next state, which takes full advantage of RNN in handling long-term dependencies of time series data. The training data are first generated from direct numerical integrations to form an initial state space, which is divided into several subregions by the K-means algorithm. The centroid of each cluster is also determined at the same time. Next, an Elman RNN is constructed in each of these subregions to approximate the expensive direct integration, in which the integration routine obtained from the centroid is regarded as the basis for a storing and retrieving solution to ODEs. Finally, the alpha-shape metrics with principal component analysis (PCA) are used to generate a set of reduced-order geometric constraints that characterize the applicable range of these RNN approximations. For online implementation, geometric constraints are frequently verified to determine which RNN network to be used to approximate the integration routine. The advantage of the proposed algorithm is to use a set of RNNs to replace the expensive direct integration, which allows to reduce both the memory consumption and computational cost. Numerical simulations of a H₂/CO-air combustion process are performed to demonstrate the effectiveness of the proposed algorithm compared to the existing ODE solver.

CFRLA-Net: A Context-Aware Feature Representation Learning Anchor-Free Network for Pedestrian Detection

High resolution and strong semantic representation are both vital for feature extraction networks of pedestrian detection. The existing high-resolution network (HRNet) has presented a promising performance for pedestrian detection. However, we observed that it still has some significant shortcomings for heavily occluded and small-scale pedestrians. In this paper, we propose to address the shortcomings by extracting semantic and spatial context from HRNet. Specifically, we propose a Context-aware Feature Representation Learning Module (CFRL-Module), which combines a Multi-scale Feature Context Extraction Parallel Block for Convolution and Self-attention (CEPCA-Block) with two parallel paths and an Equivalent FFN (EFFN) Block. The core CEPCA-Block adopts a parallel design to integrate convolution and multi-head self-attention (MHSA) with low parameter computational cost, which can obtain the deep semantic context by convolution path and precise context by MHSA path. Furthermore, to overcome the inefficiency of global MHSA in high-resolution pedestrian detection, we propose a novel local window MHSA, which can significantly reduce memory consumption but barely affect the detection performance. Cascading the proposed CFRL-Module with the anchor-free detection head constitutes our Context-aware Feature Representation Learning Anchor-Free Network (CFRLA-Net). The proposed CFRLA-Net can catch a high-level understanding of the heavily occluded and small-scale pedestrian instances based on HRNet, which can effectively solve the limitation of the insufficient feature extraction ability of HRNet for the hard samples. Experimental results show that CFRLA-Net achieves state-of-the-art performance on CityPersons, Caltech, and CrowdHuman benchmarks.

An Optimized Node Level Lightweight Security Algorithm for Cloud Assisted-IoT

The fastest-evolving technology, the Internet of Things (IoT), will advance the fields of agriculture, defense, and medical electronics. IoT is focused on giving every object a purpose. IoT with cloud assistance offers a potential remedy for the issue of data expansion for individual objects with restricted capabilities. With the increasing use of cloud technology, the Internet of Things (IoT) has encountered additional security hurdles when it comes to exchanging data between two parties. To address this issue, a thorough investigation was conducted into a secure cloud-assisted strategy for managing IoT data, which ensures the safety of data during its collection, storage, and retrieval via the cloud, while also considering the growing number of users. To achieve this, a lightweight security mechanism that is optimized at the node level is implemented in the proposed system. By utilizing our technology, a secure IoT infrastructure can be established to prevent the majority of data confidentiality threats posed by both insiders and outsiders. Using a heartbeat sensor and a node MCU, we create a heartbeat monitoring system. At the node MCU level, giving security to the patient's health data and preventing unauthorized users from attacking it. Smaller key sizes and lightweight security techniques for IoT devices with minimal power, lower power and memory consumption and Execution time, transmission capacity reserve is used to achieve security. In order to achieve this. The performance of the RSA and ECC algorithms in terms of execution time, power consumption, and memory use have been tabulated for this experimental arrangement. The ECC method occurs to produce the best results in tiny devices.

Asymmetric solid burst correcting integer codes

With the development of technology, communication channels are increasingly experiencing burst faults of various forms caused by noise elements. To get around this, an appropriate encoding and decoding mechanism should be designed while taking into account things like redundancy, memory usage, efficiency, etc. Motivated by these facts, in this paper, we present a class of integer codes capable of correct asymmetric solid burst errors. In addition to the theoretical foundations, the paper also derives the expressions for the probabilities of incorrect and correct decoding for the proposed codes. Lastly, we compare the proposed codes with other similar codes in terms of code rate and memory consumption.

Intrusion detection in cloud computing based on time series anomalies utilizing machine learning

The growth of cloud computing is hindered by concerns about privacy and security. Despite the widespread use of network intrusion detection systems (NIDS), the issue of false positives remains prevalent. Furthermore, few studies have approached the intrusion detection problem as a time series issue, requiring time series modeling. In this study, we propose a novel technique for the early detection of intrusions in cloud computing using time series data. Our approach involves a method for Feature Selection (FS) and a prediction model based on the Facebook Prophet model to assess its efficiency. The FS method we propose is a collaborative feature selection model that integrates time series analysis techniques with anomaly detection, stationary, and causality tests. This approach specifically addresses the challenge of misleading connections between time series anomalies and attacks. Our results demonstrate a significant reduction in predictors employed in our prediction model, from 70 to 10 predictors, while improving performance metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Median Absolute Percentage Error (MdAPE), and Dynamic Time Warping (DTW). Furthermore, our approach has resulted in reduced training, prediction, and cross-validation times of approximately 85%, 15%, and 97%, respectively. Although memory consumption remains similar, the utilization time has been significantly reduced, resulting in substantial resource usage reduction. Overall, our study presents a comprehensive methodology for effective early detection of intrusions in cloud computing based on time series anomalies, employing a collaborative feature selection model and the Facebook Prophet prediction model. Our findings highlight the efficiency and performance improvements achieved through our approach, contributing to the advancement of intrusion detection techniques in the context of cloud computing security.

Low-light images enhancement and denoising network based on unsupervised learning multi-stream feature modeling

Inspired by the generative adversarial networks EnlightenGAN, this paper proposes a novel low-light images Enhancement and Denoising model based on unsupervised learning Multi-Stream feature modeling (MSED). The model has two stages: generator network and discriminator network. Generator network includes global and local feature modeling network. First, Swin Transformer Block is innovatively introduced in the global feature modeling of generator network stage. It makes the interaction between the image and the convolutional kernel related to the image content. Its shift window mechanism can model the global feature dependency of the input image with less memory consumption, and extract the color, texture, and shape features of the image, thereby effectively suppressing noise and artifacts. Second, in the local feature modeling, a multi-scale image and feature fusion branch is added. It not only extracts reduced features from large-scale low-light images, but also extracts features from multiple downsampled low-light images, and then combines the two through attention and DSFF. By utilizing the complementary information of the reduced features and downsampled images, various underexposure/ overexposure phenomena caused by low-light images can be effectively avoided. In the discriminator network stage, the deep/shallow feature aggregation module is added to enhance the discrimination ability, and the inconsistencies are suppressed by learning the contradictory information of spatial filter, so that the shallow representation of information and the deep semantic information can guide each other. Thanks to the synergy of the above three innovative work, compared with many existing advanced low-light images enhancement models, MSED can achieve SOTA level performance on several public datasets. However, when dealing with low-light images with blurry content caused by rapid motion, MSED still cannot effectively restore their detailed information.

A novel approach to attention mechanism using kernel functions: Kerformer.

Artificial Intelligence (AI) is driving advancements across various fields by simulating and enhancing human intelligence. In Natural Language Processing (NLP), transformer models like the Kerformer, a linear transformer based on a kernel approach, have garnered success. However, traditional attention mechanisms in these models have quadratic calculation costs linked to input sequence lengths, hampering efficiency in tasks with extended orders. To tackle this, Kerformer introduces a nonlinear reweighting mechanism, transforming maximum attention into feature-based dot product attention. By exploiting the non-negativity and non-linear weighting traits of softmax computation, separate non-negativity operations for Query(Q) and Key(K) computations are performed. The inclusion of the SE Block further enhances model performance. Kerformer significantly reduces attention matrix time complexity from O(N2) to O(N), with N representing sequence length. This transformation results in remarkable efficiency and scalability gains, especially for prolonged tasks. Experimental results demonstrate Kerformer's superiority in terms of time and memory consumption, yielding higher average accuracy (83.39%) in NLP and vision tasks. In tasks with long sequences, Kerformer achieves an average accuracy of 58.94% and exhibits superior efficiency and convergence speed in visual tasks. This model thus offers a promising solution to the limitations posed by conventional attention mechanisms in handling lengthy tasks.

Instance-Based Continual Learning: A Real-World Dataset and Baseline for Fresh Recognition

Real-time learning on real-world data streams with temporal relations is essential for intelligent agents. However, current online Continual Learning (CL) benchmarks adopt the mini-batch setting and are composed of temporally unrelated and disjoint tasks as well as pre-set class boundaries. In this paper, we delve into a real-world CL scenario for fresh recognition where algorithms are required to recognize a huge variety of products to facilitate the checkout speed. Products mainly consists of packaged cereals, seasonal fruits, and vegetables from local farms or shipped from overseas. Since algorithms process instance streams consisting of sequential images, we name this real-world CL problem as Instance-Based Continual Learning (IBCL) . Different from the current online CL setting, algorithms are required to perform instant testing and learning upon each incoming instance. Moreover, IBCL has no task boundaries or class boundaries and allows the evolution and the forgetting of old samples within each class. To promote the researches on real CL challenges, we propose the first real-world CL dataset coined the Continual Fresh Recognition (CFR) dataset, which consists of fresh recognition data streams (766 K labelled images in total) collected from 30 supermarkets. Based on the CFR dataset, we extensively evaluate the performance of current online CL methods under various settings and find that current prominent online CL methods operate at high latency and demand significant memory consumption to cache old samples for replaying. Therefore, we make the first attempt to design an efficient and effective Instant Training-Free Learning (ITFL) framework for IBCL. ITFL consists of feature extractors trained in the metric learning manner and reformulates CL as a temporal classification problem among several most similar classes. Unlike current online CL methods that cache image samples (150 KB per image) and rely on training to learn new knowledge, our framework only caches features (2 KB per image) and is free of training in deployment. Extensive evaluations across three datasets demonstrate that our method achieves comparable recognition accuracy to current methods with lower latency and less resource consumption. Our codes and datasets will be publicly available at https://github.com/detectRecog/IBCL .