วีดีโอ

RE: To address the issues of false backtesting and overfitting, a suitable technique is to employ k-fold cross-validation. This approach involves dividing the sample data into K subsets, or folds, of equal size. During each iteration, one subset is kept as validation data, while the model is trained and tested on the remaining K-1 subsets. During each iteration, steps include: (1) Reserve a subset as validation data, and test on the remaining K-1 subsets. (2) Observe how the model performs on the validation sample. (3) Score model performance based on output data quality. This method can mathematically reduce the impact of small sample size, single sample, noise data and other factors on the prediction results. Additionally, it is important to consider the potential overfitting phenomenon, which may arise from the relatively small sample size used in this study (half a year's data). To mitigate this, increasing the sample size can be considered to ensure a more robust and reliable analysis, helping to reduce the risk of false backtesting and overfitting. A larger sample size allows the model to better capture underlying patterns and generalizations, leading to more accurate predictions and more trustworthy investment strategies.

RE: For the convenience of investors to make judgments, the ai forecasting model can omit the selection of forecasting methods, namely momentum strategy and reversal strategy, and instead directly give future growth trends and investment suggestions. In practice, on the basis of the code in this study, the result interpretation and personalized stock input and output windows are added. The result interpretation automatically assesses the goodness-of-fit using R square and selects an appropriate plan after the forecast is complete. It also provides forecasts on whether the stocks are expected to rise or fall. The input and output window allows investors to enter stock codes, retrieve relevant stock information, and receive forecast results and investment suggestions in the output window. These additional features have been designed based on the research to offer consumers more user-friendly and intuitive investment advice, making it easier for them to interpret the results and make informed investment decisions. This AI forecasting model aims to provide a convenient and valuable tool for investors seeking reliable insights into the stock market.

RE: Efficiency and Accuracy are important indicators for comparing the two methods, but they are not the only indicators. Others like flexibility, which can enable the ease of flexible customization for specific cases or according to different datasets. For example, for the size requirements of the data set, different methods may have different requirements for the training set. Model size, for example, will be more important in some application scenarios with limited memory. In short, the comparison between the two may not only include the two indicators of Efficiency and Accuracy, but also requires multi-dimensional evaluation according to application requirements.

RE: There are various theoretical frameworks for analyzing the two mainstream methods, such as backbone network, performance metrics, computational efficiency, robustness, hyperparameters, loss functions, etc., each of which can expand a huge topic. The main structure proposed in the paper is the backbone structure, such as VGG, Darknet, ResNet, etc. among them. Different network structures play an important role in feature extraction. YOLO uses a custom backbone architecture called DarkNet. DarkNet is a lightweight convolutional neural network designed for YOLO, focusing on efficiency and speed. Due to its characteristics, YOLO takes up less memory in use and has better real-time performance. VGG is the backbone structure of Fast R-CNN. This is a network with a deep architecture. Due to its characteristics, the accuracy of the Fast R-CNN method will be better. For real-time applications, VGG's deep architecture may make Fast R-CNN slower than YOLO, especially on resource-constrained devices. In addition, performance metrics can be used as an analysis framework, including precision (the ratio of true positive detections to the total number of positive predictions), Recall (the ratio of true positive detections to the total number of ground truth positive instances). There is also Mean Average Precision (mAP), It computes the average precision (AP) for each class and then takes the mean over all classes.

RE: It is still not resolved as to whether or not markets exhibit “tipping points” or not, in the sense of the critical phenomena studied in physics. Part of this lies in the fact that markets are generally assumed to be in equilibrium, and from there develop the necessary structural solutions given that assumption, so if the equilibrium assumption doesn’t hold then the remainder of the analysis is not well founded. With that in mind, claiming that markets go through a non-equilibrium transition is a strong claim that requires validation, and so this paper provides evidence for this strong claim of non-equilibrium transitions.

RE: The innovation here is in the computational aspect of the work: to date the sliding window method for the analysis of critical events had not been developed or tested on real data, hence the extensive back-testing described in the paper. The computational development is reasonably straightforward in principle, but like many such techniques it needs extensive validation to make sure there are no inconsistencies, in part this is provided by the statistical tests of the tine series, followed by the criticality measures developed by Michael Small and his group (whose group lent me Ayham Zaitouny to compute and validate that aspect of the work).

RE: I think high-performance computing and simulation will evolve further and become different from current approaches. Moore's Law is slowing down and we need to change the way we compute. For scientific applications, I believe that hardware and software will become stronger linked to one another. For instance, neuromorphic computers have certain advantages for specific applications, i.e., they can be more efficient and less energy-consuming when it comes to simulating certain types of neural networks. Regarding applications, there are many more potential avenues than in the past. I am personally very much interested in working with agent-based computational models, which I used for studying neural development and cancer. In the future, I would like to use high-performance computing with agent-based modelling to simulate initially individual brain areas, up to an entire human brain.

RE: In my view, the most challenging part is to obtain sufficient relevant information from experimental data. Biological systems can be extremely complex on multiple levels, both spatially (e.g., intra-cellular, extra-cellular, tissue-level, organ-level, whole-organism level, etc.) as well as temporally (e.g., electrical activity, cellular proliferation and synaptic learning take place on very different timescales). As a computational modeller, it can be very difficult to determine suitable parameters to generate a meaningful computational model. Fortunately, current experimental techniques in disciplines such as MRI, transcriptomics or proteomics have become very powerful and it is becoming easier to access large databases or other datasets from collaborators. For instance, I'm currently working with data from the UK Biobank, which has over half a million of participants. In future, stronger collaboration and networking between experimentalists, computational biologists, neuroscientists, mathematicians, physicists and others are needed to help determine plausible model parameters.

RE: In general, as you have more machines, tail latency is going to get worse because of the overhead of coordination. If multiple messages need to be processed simultaneously (which is called batching in our domain), the latency will increase as well due to the overhead of serializing and deserializing. In general, batching is used to increase throughout by sacrificing tail latency.

RE: As modern distributed system is very complicated, there is no general solution for reducing tail latency. It’s still an ongoing research area that attracts effort from both academia and research. My vision is that there will be some cross-stack and hardware-software approaches to tackle the latency.

1. Response- Vulnerabilities in NLP models can yield a severe safety risk. In this work, we describe and empirically validate a new technique based on the K-means clus- tering technique to detect backdoor attacks in the NLP domain. 2.To evaluate the performance of our approach and validate its effectiveness, we employed real- world datasets and numerous network architectures, including a transformer- based model, WordCNN, and LSTM. Our empirical findings indicate that our K-means technique can detect backdoors with an accuracy of up to 92.59 on the YELP dataset for the sentiment analysis task. To validate the competi- tiveness of our approach against existing work, we tested our technique against three attack recipes using the same datasets and models above; our empirical results illustrate that our technique outperforms state-of-theart solutions with F 1 detection accuracy scores of up to 94.8 . We are intrigued to see how our technique performs against out-of-domain datasets. In particular, we are inter- ested in evaluating the performance of our approach in the malware anomaly detection domain of cybersecurity which is a future research direction worth pursuing.

1. The operation of averaging the collected EEG signal gives better results and is free from any type of bias. 2. Although the subjects are carrying out the same task, but the signals generated will definitely be different based upon their mental as well as physical abilities.

how so

RE: Due to the significant number of 3D U-net parameters, it is difficult to train the model with insufficient training data.

RE: MRI scan is commonly used in hospitals for stroke examination to diagnose acute ischemic stroke within 12 hours of the stroke symptoms. Therefore, it is more practical to study the segmentation of stroke lesions based on MRI data. We decided to use the ATLAS open dataset, consisting of T1-weighted stroke MRI scans, to study an automatic lesion segmentation algorithm. We will improve the network structure and improve the prediction method to achieve a more accurate 3D lesion segmentation based on the U-net network.

RE: The reason for selecting this project is social media gives some impact on people both positively and negatively. We can make a person feel up and down. Here memes take the major role. Through this project we learn how fastly memes spread the information quickly. That's why we plan to separate a meme as a troll and non-troll as a filter. And also we got some experience on creating and classifying memes dataset. Through this project we got a clear idea about the Convolutional Neural Network. How it works, which fields need CNN support and what are the models available in CNN likewise we got an overview on CNN.

RE: In the project first step we have to collect the data, next step is data augmentation here we do a random rotation and zoom techniques, next step is Data preprocessing here we resize the image, next step we train the MobileNet, ResNet and AlexNet models with the dataset, then we develop a Modified AlexNet architecture and then train the Modified AlexNet also. Then we plot a graph for those algorithms based on the accuracy.

RE: Difficult part of the project is text segmentation. First we try to split the image and text in the image, Which is a very huge task even though we don't get any work on text segmentation. That's why we move on to the CNN model where text is automatically recognized by the model which is already trained with that. It is easy for us to move ahead.

RE: Can be increased with additional training of disease images

RE: If we employ this system, the diseases in tomato can be identified earlier and there is a possibility to stop further spreading and progress of the disease. So the yield can be increased.

RE: Sure. It is possible by training with corresponding dataset. If we go for brinjal .... with its training image we can follow

RE: I have met difficulties in designing the obstacles on the map. The map size was originally 100*100. There were too many squares. It was difficult to add obstacles one by one. To cope with this issue, I learned a function online, which could randomly generate obstacles, starting point and ending point. I adjusted it for a few time, getting it able to produce an ideal map and finally locked it as my overall test map.

RE: This paper does not concern the data transmission rate, but the examples should follow approximately the same as the results mentioned in paper RDC-SLAM: REAL-TIME DISTRIBUTED COOPERATIVE SLAM SYSTEM BASED ON 3D LiDAR. Should not be larger than 5kB/s.

RE: Great question to hear. As is known, complex scenes are typically hard to align, which frequently gives large noise in the point clouds. As the detections are misleading, some denoising techniques should be introduced to mitigate the bias. As denoising is not robust, we still expect the average system error higher than normal level. For details, I strongly recommend this paper: FSD-SLAM: a fast semi-direct SLAM algorithm by Xiang Dong.

RE: The ICP algorithm is one of the most important machine learning-based heuristic algorithms for unifying the coordinate systems. Arguably, matching the coordinate systems will be continuously ruled by heuristic methods, because they’re both interpretable and fast. However, ICP algorithm is sometimes misused because the researcher does not have strong insight of the algorithm, which reduces the performance of the matching process, which gives the motivation for my research.

RE: None for now. We have submitted our proposal to Sony Research Grant previously but did not get reply from them.

RE: Our multi-stage framework takes inspiration from the cascaded network where each stage is being trained separately. Our framework adds to that by attempting to split the training set based on the difficulty so that different stages of training will be able to focus on the difficulty aspect of the training data.

RE: "The research is mainly limited by the MIT-BIH dataset on 2 reasons: Even though the MIT-BIH Arrhythmia Database is well-annotated, but it has limited number of subjects (48 patients from Boston's Beth Israel Hospital from year 1975-1979). The data quality can vary among different institutions in real practice. So more investigations are needed to test the external validity of this method in different regions/countries. In MIT-BIH dataset, certain arrhythmia types have limited samples and it hinders the model to learn useful representation/information of minority class(es). We need a larger well-annotated dataset to include more samples for uncommon arrhythmia types."

RE: It is achieving a higher detection accuracy, compared to existing models, by implementing the hybrid CNN and LSTM deep learning model with optimized hyperparameter values

RE: The project is able to design a deep learning model, train the model using a repository of spam messages and detect spam sms with a high accuracy of 99.77%

RE: The biggest challenge currently faced is the inability of existing spam detection systems to accurately detect unreported spam sms.

RE: Basically, our target was to work in an industrial ergonomics domain, the work in this domain is nascent. The explicit review of literature related to the digital design of the clinical health care systems highlights the importance of typography along with other factors associated with the operator and user interface. The advent of the COVID-19 pandemic hugely augmented the use of Ventilator systems, and the functioning of this machine in a highly critical environment grasped our attention.

RE: Over time researchers agreed and countered one another's findings related to the selection of the typography for the digital systems working in different environments. We feel that these disparities do not falsify any research findings; instead, they emphasize the challenges of understanding digital readability. We note that well-selected typography for the digital systems that optimize readability, establishes a strong visual hierarchy, over and above that saves even a small amount of time, number of touches per task, and poses low use errors could be clinically significant, especially when dealing with critical healthcare systems.

RE: This model is more accurate. By implementing this concept in a business (commercial application prospect), we can automate the laborious process of classification of product reviews by developers, which solves their issues by removing the need for constant monitoring and generating revenue while also saving time, money, and resources.

RE: This system translates the text to sign language, it may be used for learning and training purpose. Computer vision may be used to translate the signs into text.

RE: The presented system reduces the cost in terms of translation and network bandwidth. Furthermore, it may translate sentences instead of just one keyword. Moreover, it may be used for training and teaching purposes.