Science topic

Deep Learning - Science topic

Explore the latest questions and answers in Deep Learning, and find Deep Learning experts.
Questions related to Deep Learning
Question
4 answers
Hi,
Is there any single-stage deep learning based instance segmentation method that outputs single channel mask where each pixel indicates instance labels (assume we are considering one class: person)?
To be more concise, Is it possible that I have an input image with multiple instances of a person class and deep learning model outputs a mask for all instances present in input image by predicting 1,... k labels in single channel output mask
I hope my question is clear now
Relevant answer
Answer
Yes it is possible, I guess I have tried the same thing with U-net. but I don't know for some reason when the number of classes increases the model accuracy drops. but I assume for one class should be working fine.
Question
3 answers
Is there some methodology or reasoning that we can follow for this selection? I understand the difference that for two-tailed are of the tail is equally distributed on both ends.
But is there any need to select one over the other? How to decide if I encounter new data, what test to go with?
Relevant answer
Answer
There are a few considerations to make:
if you are able to previously predict your outcome i.e. you expect to see an increase or a decrease, then you should use a one-sided alternative hypothesis. However, if you cannot anticipate the direction of your result, then a two-sided alternative is more appropriate, where it tells you if there is any significant association between your tested groups, but cannot tell you about an increase or a decrease i.e. cannot give you a direction.
Another consideration is to at the level of the sample size: if you have a small sample size, it's better to go with a one-sided alternative. However, large sample-sized experiments are suitable for the two-sided alternative.
The last consideration would your decision of how strict you want your test to be. if you want a more strict i.e. a difficult-to-achieve statistical significance i.e. when you reach it, this means that your result is so significant, then use a two-sided alternative i.e. don't divide your p-value by 2. And if you want to conduct a less strict i.e. an easier to reach statistical significance, then use a one-sided alternative i.e. divide your p-value by 2.
I hope you benefit from the comment,
Best
  • asked a question related to Deep Learning
Question
7 answers
Hi,
I'm a software engineering undergraduate.
I have a dataset which includes the numerical values of variables x,y,z and output r.
I want to create an algorithm which basically uses a prediction algorithm using neural networks and finds the relationship/correlation between x,y,z and predict r and/or also i want to forecast the r value ?
what type of algorithms should i look into?
Thank you.
Relevant answer
Answer
you can go for regression analysis. I think that is the best possible solution for the prediction you want to do for three variable as input.
Question
9 answers
I am training a model, where class distribution is unbalanced. So, I thought of applying augmentation technique, to improve the performance. I am wondering,is it fair to apply augmentation only on class with low samples, or Do I apply augmentation on all data?
If we apply augmentation on only one class, will it create bias while training?
Please provide your valuable insights on this.
Relevant answer
Answer
Here's an interesting paper that talks about how a change in class prior (i.e., class balance) from training to applying (e.g., testing) can affect a classifier's performance.
Another way data augmentation can be useful is to build models that can generalize better. Often in computer vision tasks, people would augment the dataset, for example, with rotated versions of the training images. Here's a paper talking about that type of augmentation:
  • asked a question related to Deep Learning
Question
4 answers
Hi,
I'm a software engineering undergraduate.
I have a dataset which includes the numerical values of variables x,y,z and output r.
I want to create an algorithm which basically uses a prediction algorithm using neural networks and finds the relationship between x,y,z and predict r, such that when i give a new set of inputs it should predict z, this is a reactive approach.
Therefore to make it proactive, i want to look at the past data of x,y,z and r and forecast future values of z.
Can i combine two models like this and make a hybrid algorithm because sometimes we can't just use past behavior to determine? are there any usecases or papers which have used prediction and forecasting together?
Also any idea how long will it take to train both models?
Thank you.
  • asked a question related to Deep Learning
Question
4 answers
Generative Adversarial Networks have had a great deal of success with tasks such as Image Generation, but I've been having a hard time looking through the research projects on GANs that work with sequential data such as time series, music tracks, etc. I wasn't able to compare the methods properly as they don't give any solid metric the compare their performances. I was wondering if anyone here has checked the recent studies on this subject and could give me a summary and guide to the most promising models and approaches.
I have a dataset of power amplitudes of a signal and I'm trying to generate some synthetic data that resemble the original values closely.
Relevant answer
Answer
Generating images from word descriptions is a challenging task: Nowadays this task is to try accomplished by combining the ConvLSTM + GAN (Attention Mechanism)
Generating Image Sequence from Description with LSTM Conditional GAN
Question
8 answers
Computer vision, Deep learning
Relevant answer
Answer
I agree with Pravin to a certain extent. But don't you think GANs are designed for the generation of new data & discrimination such as fake news, rather than image annotation.
  • asked a question related to Deep Learning
Question
4 answers
I am working on a CV project where I am trying to extract key frames from videos. The videos are of bottles containing text labels, now the criteria for key frame in my case, is "to extract those frames such that the frames cover all the text on the bottle". So as you can see the criteria for choosing key frames is more text driven here.
I know that we generally either use frame clustering , shot detection or compare histograms of frames to extract the key frames but I am not sure if that is the best approach for this particular use case, given that the colour intensity may not vary much from frame to frame(Black/White text written on white label)
So have anyone of you worked on such a problem before or any pointers as to what could be a better way to approach this
Relevant answer
Answer
Video is a collection of frames. Texture techniques can be used to achieve this objective.
Question
1 answer
I am studying 2D and 3D registrations. I need help and guidance in producing DRR from CT images. Is there an article or program that fully explains how to generate DRR from CT images? .
I want to use CNN networks for registration whose input is a fluoroscopic and DRR images.
thanks
Relevant answer
Answer
Recently, I have tried to apply the same approach, along with textures as a co-author in oral cavity cancer classification: https://www.mdpi.com/1424-8220/20/20/5780, maybe it will help you, I know its different applications but can be applied to Ct images too.
  • asked a question related to Deep Learning
Question
7 answers
Hi,
I extracted the features of Left and right eeg signals and classified into left and right hand from BCI compettition III dataset IVa. . I got accuracy of 60%. In matlab, can you please guide me how to correct my errors /misclassification by deep learning methods(or) machine learning methods which helps to improve 100% accuracy...
Relevant answer
Answer
To increase accuracy you can also play with some pre-processing techniques, such as stationary subspace analysis (SSA). Check out for more details. I compared different approaches and you can check my implementation presented in the paper
Question
5 answers
I am working on developing a predictive drive cycle for a specific city. I have gathered driving characteristics parameters for my target city. I have obtained similar data from previous research works and their corresponding drive cycles. I would like to train a network using the driving characteristic data obtained from the research papers keeping the time stamped drive cycle obtained from the corresponding researches as target so that the network can predict a drive cycle based on my collected data. I have experience of working on ML using TensorFlow and Keras.
Relevant answer
Answer
I think it would be appropriate to try a solution with RNN. Note that the standard RNN has a problem with the so-called fading gradient. I recommend using the LSTM network on a non-sequential input data set, it is enhanced by introducing an additional closed and cellular state.
Question
5 answers
I'm starting a master's research about on-board image processing for small satellites but struggling to find a niche or problem to solve or discuss. I've read less than 15 articles about related works and would hopefully find one nice topic to pursue if I continue reading, but I just want to ask some senior researchers in this field what would they recommend.
Relevant answer
Answer
1- Various large smallsat constellations in low Earth orbit (LEO)
LEO has been commercially active in both broadband and imaging constellations of 100 satellites or more. This scenario takes mainly into account "mega-constellations" of broadband, providing a low latency to accessible global broadband Internet.
_______________________________________________________________
_______________________________________________________________
2- Smalls with Larger Remote Sensing Satellites
In this scenario a increasing number of countries have close parallel access to technology with large satellites in distant context , given the remote sensing capabilities that are available commercially and beyond the United States.
_______________________________________________________________
_______________________________________________________________
3- Insecure for LEO Satellite Operation
In this scenario a increasing number of countries have close parallel access to technology with large satellites in distant context , given the remote sensing capabilities that are available commercially and beyond the United States.
_______________________________________________________________
_______________________________________________________________
4- On the Space Maintenance, Assembly & Development
In these circumstances, many persistent LEO and GEO platforms for on-orbit operation, assembly and development (OSAM) are used by governments and the private sector.
As large satellites become the standard for competitive and host payload platforms, the satellite industry is versatile in designing , constructing and utilizing satellites that best fit a given application. drives the perception of profitability and consequently injection of funding and talent into other drivers, including the development of new technology, low-cost approaches, and infrastructure.
_______________________________________________________________
_______________________________________________________________
5- Market Demand
-- The most important technology for the four scenarios examined comprises optical imagery, radio frequency interference (RFI), spectrum use, small platform optical communication, and propulsion system miniaturization.
-- Low cost approaches to development , implementation, and robotics and alternative business models, such as modularity and standardization are included. are included.
-- Professional practice of infrastructure include space awareness technologies and systems (SSA), ground stations networks and space relays, and low cost ground antennas and terminals. Infrastructure drivers include
_______________________________________________________________
_______________________________________________________________
6- Space Access
The fact that the four scenarios can be realised in the interim is driven
by the cost of their launch as well as the availability of reliable launch options.
_______________________________________________________________
_______________________________________________________________
7- Alternatives to Compete
Smallsats may either render or ignore the relative value proposition for smellsats of alternatives such as terrestrial and airborne platforms and gradual and advanced developments in large satellites.
_______________________________________________________________
_______________________________________________________________
8- Government Policies
The spectrum allocation policies of governments; RFI; protectionism/ mercantilism; debris mitigation; on-orbit regulation; and the management
of space traffic drive the private sector 's involvement in the smallsat ecosystem, both positive and negative.
_______________________________________________________________
_______________________________________________________________
  • asked a question related to Deep Learning
Question
6 answers
I'm thinking about doing a research in satellite image processing field. Can I calculate the area covered of cloud/forest/vegetation/rural/water using deep learning/AI? If so, can someone provide links of related literatures. It would be very helpful, thank you!
Relevant answer
Answer
It is an interesting hot field of research with many applications, see e.g., https://www.satimagingcorp.com/applications/ and .
A s for as computational algorithm for experimental process is concern, you will understand a suitable model after literature reviews (Deep learning is an application for training, it is not suitable for every data, neither it is yet really understood).
Question
6 answers
according to you what are Hot problems currentlyworld is facing? And what solutions you would suggest to solve them?
Relevant answer
Answer
  1. Now, I agree with prof. Madhukar Baburao Deshmukh.
Question
9 answers
Currently working on to detect acoustically radiating source that consists of both narrowband and broadband signal components in the presence of non – Gaussian noise and low SNR condition.
So, for this MFCC method has already been used for feature extraction, now i have to find some different algorithm for feature extraction which can give improved performance in discriminating between signal and noise.
Also, i would like to know that what features will give me the best performance, for discriminating the signal and the noise, when the signal has narrowband and broadband components and noise can be of any types i.e what features researchers are looking to discriminate the signal from the noise.
Any help or research would be much helpful.
Thanks
Relevant answer
Answer
I think, the best way is to leverage or adopt some effective filtering technique, e.g., Butterworth Filtering, Kalman Filtering, Particle Filtering or the most recent is Grassmann Filtering. However., all this is based on your task and nature of data.
Question
9 answers
From a scientific perspective, is this acceptable to compare the ?active contours technique with the deep learning technique for medical image segmentation?
and if its yes.
So what are the scientific benchmarks for comparison because these two techniques differ in approach?
Does it have scientific value?
Kindly correct me any concept that I used if it was not in a scientific way.
Relevant answer
Answer
Thank you Dr. Larbi Messaouda and Dr. Muhammad Ali for your support
Question
3 answers
LSTM tools of the Matlab allow us to perform classification of the time-series. It is assumed, that different time-series can have different length, but they should have same interval of measurements (e.g., min., days, weeks, etc.). In real situations different time-series can have different times of measurement – e.g., one time-series is measured on times 1, 3, 7, 9 and other on the times 2, 6, 8. Certainly, we can perform interpolation and to transform our time-series to the time-series with same measurement times (e.g. 2, 5, 8), but may be you can propose more effective solution? Thanks a lot for your answers beforehand. Regards, Sergey.
Relevant answer
Answer
Hi Sergey,
I recommend the following 2 papers that perfectly match your needs. Both are based on Neural ODEs by Chen et al., 2018:
Regards
  • asked a question related to Deep Learning
Question
10 answers
I search for applications about recommender systems.
The Idea in general is to use algorithms which gives according to historical data of the classifications of soil. then make a mapping spell between the existing data of soil, weather and crops and an input entered, user to know the type of crop adapted to his soil in addition to that the amount of appropriate fertilizer.
If you have research/review articles comparative studies or support it will be very helpful .
Relevant answer
Answer
We still need basic data , which will come from crop-nutrient repsonse studies in field or the basic dataset in field need to be subjected to machine learning duly cross validated in order to set right the suitability for practical application in filed...
  • asked a question related to Deep Learning
Question
7 answers
I need to create a simple?Recurrent Neural Network RNN or Long short-term memory (LSTM), which is specific type of RNN.?After searching through examples and forums I haven't come across many applications that are similar enough to my project.? I tried to implement a simple RNN to classify the ECG signals into two classes, but no luck so far. I'm confused about?What is the correct input shape for the model?.
Let say I have 20 ECG record.?My input data is such that I have m=20 signals each with n=7000 data points, and L=2 output labels that I am trying to learn. How I should structuring my input data. Is it true to make the input as (m,n,1) and output targets as (m,L)!.
I attempt to use the following Matlab toolboxes for building the RNN:
Please help!
Relevant answer
Answer
There’s nothing more entertaining than people barging in on a message board like this and giving their priceless advice on subjects they know nothing about. Typically, astronomer (“exomoon hunter”) Alex Teachey giving – without anyone asking him to - his enlightened opinion on the machine-learning classification of heartbeats from ECG data. “It’s a time series, so the LSTM RNN is potentially useful”: yeah, sure. “Going from location A to location B is a transportation problem, so the airplane is potentially useful”: yeah, sure. Nevertheless, knowing that location A is three miles away from location B might help refrain from taking a sledge hammer to crush a tiny nut.
There are thousands of published papers on LSTMs and RNNs applied to time series; they should provide you with as a huge number of “input and output shapes of LSTM RNNs”, whatever that means. My bookmarked favorite is “Applying LSTM to time series predictable through time-window approaches” (ICANN 2001, doi: 10.1007/3-540-44668-0_93); in that paper, Jürgen Schmiedhuber (the inventor of LSTMs) and his collaborators report that LSTMs have no advantage over other methods for a specific class of time series – probably the only paper in the machine learning literature acknowledging that the authors’ own method is not universally superior to all others. The paper is about laser signals, about which, of course, you "could not care less”.
I must apologize deeply for having replied, three years ago, to Huda Da’s question on heartbeat classification by discussing heartbeat classification, despite the fact that you “could not care less whether heartbeats are an appropriate application”; obviously, I should have anticipated that. Sorry for wasting your invaluable time reading my worthless answer to Huda Da’s question, and thanks for a good laugh!
Question
15 answers
Hi Everyone. Here I provided SCI or SCIE journals with Free or Low Cost in the area of Plant Sciences and Food Science & Technology. I think it will be helpful for those who are economically poor. If anyone knows any other journal names with low-cost APC kindly provided in the reply section so that some people will get benefited. Thank you
Journal title
1. AGRICULTURAL AND FOOD SCIENCE (Free)
2. CEREAL FOODS WORLD(Free)
3. COMMUNICATIONS IN SOIL SCIENCE AND PLANT ANALYSIS(Free)
4. COMPREHENSIVE REVIEWS IN FOOD SCIENCE AND FOOD SAFETY (free if you join as a IFT member with 50$)
5. Journal Of Food Science(free if you join as a IFT member with 50$)
6. CRITICAL REVIEWS IN PLANT SCIENCES (Free)
7. CZECH JOURNAL OF FOOD SCIENCES (320Eur)
8. FOOD QUALITY AND SAFETY (Free)
9. FOOD SCIENCE AND TECHNOLOGY (300$)
10. FOOD SCIENCE AND TECHNOLOGY RESEARCH (200$)
11. INTERNATIONAL FOOD RESEARCH JOURNAL (250$)
12. INVASIVE PLANT SCIENCE AND MANAGEMENT (120$)
13. POLISH JOURNAL OF FOOD AND NUTRITION SCIENCES (40,000INR)
14. TURKISH JOURNAL OF AGRICULTURE AND FORESTRY (Free)
Relevant answer
Answer
Find this link: https://mjl.clarivate.com/search-results. My recommendation is to opt for high-impact factor quality journals and publishers only.
  • asked a question related to Deep Learning
Question
3 answers
I am working on a project where I want to implement time stretching and pitch shifting with deep learning methods as a first thing. I tried searching the internet but I haven't come across many papers or articles to start with. If anyone has the clue, a little help would be really appreciated.
Relevant answer
Answer
Hello!
I personally found the following papers quite helpful for having a basic overview on audio processing in conjunction with deep learning.
  • asked a question related to Deep Learning
Question
2 answers
I am using a U-Net architecture with Xception net as encoder . The visual area of the segmentation mask is very small and after training, it is giving a lot of false positives. I am thinking of changing the kernel size of 3 by 3 to 5 by 5. What precautions/care shall I take while creating the model? I am getting the issue of dimensionality mismatch if just change 3 by 3 to 5 by 5 kernel size.
Also,while concatenating the skip connections, the dimensions don't match
Relevant answer
Answer
Hello!
I think the issue regarding dimensional mismatch can be fixed by modulating the stride and padding. You can look into the initial layers of AlexNet architecture (provided below) for better understanding.
model=Sequential()
model.add(Conv2D(filters=96, kernel_size=(11,11), strides=(4,4), padding="valid", activation="relu", input_shape=(224,224,3)))
model.add(MaxPooling2D(pool_size=(3,3),strides=(2,2)))
model.add(BatchNormalization())
model.add(Conv2D(filters=256,kernel_size=(5,5),strides=(1,1), padding="valid", activation="relu"))
model.add(MaxPooling2D(pool_size=(3,3), strides=(2,2)))
  • asked a question related to Deep Learning
Question
3 answers
I am currently preparing a paper for a conference, describing a deep learning model for Human Activity Recognition.
The model is tested on several known datasets.
Is a resulting confusion matrix supposed to formed on a single random training/test split, since this information is lacking on most of the papers I have read?
What I am currently doing is a stratified 10-fold cross-validation, for each training/test split reinitializing the model, training it on the training set and calculating a confusion matrix on the test-set and then adding all 10 confusion matrices together. The idea here being that this would result in a general accuracy over all test sets.
I am aware that cross-validation is usually used for finding optimal hyperparameters, but lets say I have found them in this scenario and I'm not changing them.
I would also like to include a plot of the training accuracy/loss for each batch/epoch, which would show the training process, but this would take too much space in the paper because I would have 10 plots.
Relevant answer
Answer
A common approach is also to work with 3 sets - training, validation and test. You evaluate for the best model using cross-validation on the training and validation pairs and then do your final analysis on the test set.
  • asked a question related to Deep Learning
Question
7 answers
I have an idea of
1. RNN Embedding
2. RNN with pack padded sequence
3. FastTest
4. Bi-LSTM
5. CNN
6. CNN-LSTM
7. BERT Transformer
these models.
I am looking model apart form these.
Relevant answer
Answer
You can use MobileBERT, it's a compact BERT model open sourced on GitHub. ... or the other Google open source models using projection methods, namely SGNN, PRADO and pQRNN. pQRNN is much smaller than BERT, but is quantized, and can nearly achieve BERT-level performance, despite being 300x smaller and being trained on only supervised data.
Question
14 answers
I am trying to implement a CNN (U-Net) for semantic segmentation of similar large greyscale ~4600x4600px medical images. The area I want to segment is the empty space (gap) between a round object in the middle of the picture and an outer object, which is ring-shaped, and contains the round object. This gap is "thin" and is only a small proportion of the whole image. In my problem having a small a gap is good, since then the two objects have a good connection to each other.
My questions:
  1. Is it possible to feed such large images on a CNN? Downscaling the images seems a bad idea since the gap is thin and most of the relevant information will be lost. From what I've seen CNN are trained on much smaller images.
  2. Since the problem is symmetric in some sense is it a good idea to split the image in 4 (or more) smaller images?
  3. Are CNNs able to detect such small regions in such a huge image? From what I've seen in the literature, mostly larger objects are segmented such as organs etc.
I would appreciate some ideas and help. It is my first post on the site so hopefully I didn't make any mistakes.
Cheers
Relevant answer
Answer
?The Fully-Convolutional Net (FCN) implementation is the most popular architecture for semantic segmentation?. FCNs only have convolutional and pooling layers which give them the ability to make predictions on arbitrary-sized inputs. Another option is R-CNN (Regions with CNN feature). It performs the semantic segmentation based on the object detection results. R-CNN first utilizes selective search to extract a large quantity of object proposals and then computes CNN features for each of them. Finally, it classifies each region using the class-specific linear SVMs. Compared with traditional CNN structures which are mainly intended for image classification, R-CNN can address more complicated tasks, such as object detection and image segmentation, and it even becomes one important basis for both fields.
Question
3 answers
I have textual data, I need to annotate for further experiments but I do not have sufficient domain expertise to label it myself or financial resource to hire annotators. Is there any other method like standard machine learning/deep learning approach to annotate the data. The data is on sentiment analysis and I need to classify this data into three classes based in the sentiment. I will be using this annotated data for further deep learning architectures. Is there a way other manual annotation or hiring annotators.The data is contains Indian language
  • asked a question related to Deep Learning
Question
20 answers
Why most researchers are shifting from tensorFlow to Pytorch?
Relevant answer
Answer
Tensorflow creates static graphs, PyTorch creates dynamic graphs.
In Tensorflow, you have to define the entire computational graph of the model and then run your ML model. In PyTorch, you can define/manipulate/adapt your graph as you work. This is particularly helpful while using variable length inputs in RNNs.
Tensorflow has a steep learning curve. Building ML models in PyTorch feels more intuitive. PyTorch is a relatively new framework as compared to Tensorflow. So, in terms of resources, you will find much more content about Tensorflow than PyTorch. This I think will change soon.
Tensorflow is currently better for production models and scalability. It was built to be production ready. PyTorch is easier to learn and work with and, is better for some projects and building rapid prototypes.
  • asked a question related to Deep Learning
Question
4 answers
It seems that using machine/deep learning to solve PDEs is very popular (actually, not only in scientific computing, but also in all fields). So I want to know the reasons behind this. And is the prospect cheerful?
Relevant answer
Answer
In short its the best method to solve PDE issues
Question
3 answers
Hybride intelligent system is good techniques used for solving diffrent optomization problems. Like classification, recognition ....etc
Deep learning can be used with and hybrid with other machine learning techniques. How we can optomized deep learning method by combining with swarm or fuzzy?
Relevant answer
Answer
This can happen by creating decentralised AI working model. In that decentralised model, swarm intelligence can be used for automating the resources allocation and better decisions making process. And the deep learning can be used for data or resources processing.
  • asked a question related to Deep Learning
Question
1 answer
We have proposed a work on hybrid deep learning. I think it will be helpful for future research on COVID-19. You can read this paper and cite us.
Relevant answer
Answer
It might interesting to you to see what learnings we retrieved from COVID-19
  • asked a question related to Deep Learning
Question
6 answers
Dear researchers i am working on content based video retrieval, i need popular datasets for evaluation of my method. I have downloaded UCF101,HMDB datasets but those datasets has very small clips specially collected for action recognition.
Relevant answer
  • asked a question related to Deep Learning
Question
2 answers
CCGrid 2021:- 21st IEEE/ACM international Symposium on Cluster, Cloud and Internet Computing welcomes high quality submissions. Deadline: 8 December 2020 (Final paper submission : 15 Dec). Link: http://cloudbus.org/ccgrid2021/
Relevant answer
Answer
Thank you for your invitation ??
  • asked a question related to Deep Learning
Question
4 answers
I am training a U-Net for semantic segmentation of large medical images (4096x4096px). The two classes are "too" unbalanced. The white pixels are just about 0.1% (or less) of the whole image. The Dice Coeff loss function seems to not be working since it predicts always black pixels.
  • Is there any specialized loss function for such unbalanced data? I can not find anything that works.
  • Is the U-Net arcitecture suitable for such segmentation tasks?
I have tried to train with the following setup:
Epochs: 50 Batch size: 4 Learning rate: 1e-05 Training size: 451 Validation size: 23 Checkpoints: True Device: cuda Images scaling: 0.25
and also with batch size of 1 and learning rate of 10^4. I would appreciate some help.
Cheers
Relevant answer
Answer
Also when using any Conv Net on such High-resolution data it is important to consider the maximum receptive field size of the neurons in your network. What is the minimum spatial resolution in your network and what are the filter sizes?
How much spatial context do you think your network needs to make a good decision?
Question
5 answers
I'm quite a newbie when it comes to practically working with Deep Learning techniques, although I studied them quite a lot theoretically in the last months. However, now I'm facing my first practical task.
The setting is the following: There is an air hockey game that exists both physically and as a simulation which could be run in parallel. Currently, the robotic opponent is programmed with a fixed algorithm and should further on be replaced by a trainable AI. Note that we are operating in a low-dimensional space as the environment is solely represented as a two-dimensional surface (x and y coordinates of the mallets and the puck).
Following my research, I decided to go for an approach incorporating Model-Based Reinforcement Learning in order to achieve a descent sample efficiency as the simulation won't run much faster than real time. Therefore, I plan to pre-train the network using classic Supervised Learning methods with data generated by the already existing simulation to learn about the environment's model beforehand.
As I'm lacking some practical knowledge and I read about several problems regarding the convergence of deep Q-learning methods, I'm not sure how to proceed with the Reinforcement Learning part. Also, I'm pretty overwhelmed by the publications made in this field throughout the last years. Thus, I would appreciate some practical recommendations, which kind of algorithms (specific policy gradient, Q-learning or hybrid methods) might be promising for my application.
Thank you in advance!
Relevant answer
Answer
Seeing how Air-hockey is similar to the old Atari game Pong, I recommend you to workup on Q Learning with function approximation. Further, you can follow the github repository on Deep Reinforcement Learning: Pong from Pixels for more information on practical implementation and quick convergence of the algorithm.
  • asked a question related to Deep Learning
Question
3 answers
Hi,
What is the easier deep learning method to use to predict salt and water in sea cucumbers?
Is the xgboost method easier and better than random forest?
I'am very grateful if you guys want to help me.
Thank you.
  • asked a question related to Deep Learning
Question
8 answers
If I have collected data regarding say food-preferences from multiple sources and merged them.
How can I decide what kind of clustering to do if I want to find related preferences?
Whether to go for K means, hierarchical, density-based, etc. ?
Is there any process of selecting the clustering technique?
Relevant answer
Answer
If you know (or any rough idea) how many clusters you need, then you may try K-means or some other variants of it like K-means++ or MinMax K-means clustering algorithm. But if you do not have prior knowledge about number of clusters then you can try with different values of k, and assess the goodness of the results using some cluster validity indices like DB index, C index, CH index, Dunn index etc. Otherwise you can DBSCAN clustering algorithm.
  • asked a question related to Deep Learning
Question
4 answers
I am looking for a animal pose estimation code that predicts the pose of an animal given 100-200 annotated frames from scratch using deep learning on a frame-by-frame basis. Is there any such code? I am not looking for something like DeepLabCut or DeepPoseKit or LEAP/SLEAP.ai tools. Looking for a simple baseline preferably written in PyTorch that could be easily modified.
Relevant answer
Answer
Hello Umair,
Hope you are doing well.
There are some useful sources related to your question.
DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning. (https://elifesciences.org/articles/47994 )
SLEAP: Social LEAP Estimates Animal Poses (https://github.com/talmo/leap)
BW
  • asked a question related to Deep Learning
Question
4 answers
Currently, I am working on my final year project which involves depth estimation form multiple camera inputs for video analysis purposes. I am new in this field and struggling to look a the problem from the scratch. Can anyone share some work or research papers(classical work and recent work) to look at the problem.
Thanks,
Relevant answer
Answer
You must look into stereo vision, some use costs functions to match points between the images and some use machine learning and convolutional neural networks to identify depths. Look into something called the correspondence problem.
  • asked a question related to Deep Learning
Question
9 answers
i design a neural network based WSN to achieve real time operation, i simulate the part of neural in matlab and the issue of WSN in a network simulator. how can i start with neural part?
and is there any effect in the design of neural network due to using it in a WSN?
thanks
jannat
  • asked a question related to Deep Learning
Question
9 answers
Is square matrix-shaped kernels provide some advantages than others?
Can we use other shapes like a rectangular matrix for CNNs in some situations?
Relevant answer
Answer
In comparison to rectangular matrices, square matrices have very beautiful computational properties., e.g., symmetry, closure properties w.r.t multiplication, addition etc. See e.g.,
Question
13 answers
Hello,
I work on data projects (data hub, repositories like Single Customer View) and digital analytics (Web analytics, trackign implementation, dataLayer) where marketing and/or UX peronas are be useful but should be more relevant with data on real persons.
I started to test StyleGAN for artistic & design projects and find the parametrics options interesting to populate dataset (for testing or other purposes).
Have you considered using StyleGAN to generate realistic faces for the personas based on data and parameters like we can use it on Artbreeder.com for example (I attached an example)?
Kind Regards,
Laurent Berry
Relevant answer
Answer
produce stereotypes of real persons is one of the risks when we use personas. Using representation like realistic photography could let people think that this representation is "THE user or customer". And of coure we could have gender or racial representation issues which is a serious issue in some museums or part of the world.
For example, In Guadeloupe, West Indies, we have Mémorial ACTe Museum "to contribute to the construction of a collective memory of slavery and to finally healing its abhorrent wounds and hurts, through historical awareness, values of tolerance and contemporary creation." http://memorial-acte.fr/the-founding-acte-memory-holds-the-future
As the local population is mixed and that the Museum is about slavery, caribbean area and colonial history, use only a kind of persona (white or black etc.) will be a bad idea. So it's probably better to use:
  • a panel of representations (like a matrix of people generated by exploring the latent space of racial types axis)
  • or a very simplistic design which merge differents persons profiles
  • or no representation at all
I encountered this kind of issues with personas designed:
  • with marketing team with data and informations not updated over time (we know tht they change over time)
  • for/with technical team which doesn't really want to work with a persona or UX-centered approach
  • by UX designer to try to identify the main profiles of users implicated in a project or everyday (more like a program)
From last year, I work on a Data Hub Project (in a Data Roadmap where I included User / Customer Experience) and started by a typology of persons (Physical and/or Legal) to design the project not on stereotypes but on role (more Activity-centered ergonomics than UX-centered). After, with data and dynamic personas generation, it' possible to classify persons (thounsands if you have data) under each role or profile.
Roles can help to limit biases and trigger process design in organisation (in the top menu of the Memorial ACTe Museum, they added profiles entries: ADHéRENT / FAMILLE / ENSEIGNANT / GROUPE / ETUDIANT / TOURISTE / PROFESSIONNEL ET ASSOCIATION)
  • asked a question related to Deep Learning
Question
6 answers
I found so many papers aplying the Deep Deterministic Policy Gradient (DDPG) algorithm implementing a critic neural network (NN) architecture where the action vector skips the first layer. That is, the state vector is connected to the first layer, but the actions are connected directly to the second layer of the critic NN.
Actually, in the original DDPG paper ("CONTINUOUS CONTROL WITH DEEP REINFORCEMENT LEARNING", Lillicrap 2016) they do that. But they do not explain why.
So... why is this? Which are the advantages of this architecture?
Thanks in advance.
Relevant answer
Answer
Look the link, maybe useful.
Regards,
Shafagat
  • asked a question related to Deep Learning
Question
8 answers
I am doing my phd work on plant disease identification using deep learning. So i am searching for data set of plant leaves of soybean and wheat please help me.
Relevant answer
Answer
Check this review paper regarding your question, good luck
Question
5 answers
Is there some rule of thumb or method to balance the number of samples vs the sample size. e.g. like a minimum of 10 % sample size or a number of samples = 40% ( like 400 samples for a size of 1000).
Is there a way top decide those.
Relevant answer
Answer
May be this article will help you. Link:
Question
5 answers
Hi All, I'm looking for a database of Raman spectra of bacteria. To give a background to this - we want to use Raman spectra and deep learning to distinguish between several microbial strains. I have found numerous papers using the same strategy, so I've been wondering if we could use datasets from some of such studies to test our deep-learning analysis on realistic data before our setup is ready and we can start collecting spectra. Thank you!
Relevant answer
Answer
Thanks, Luca Parisi , this looks very relevant to what we want to do.
Question
9 answers
Hello all,
Could someone guide me in looking for an optimized combination of Machine Learning and Deep Learning methods/models to deal with medical data?
Thanks in advance.
Relevant answer
Answer
Many thanks Luca Parisi .
Question
4 answers
Is it possible to determine any Physics Law from a moving object using deep learning model algorithm or can I train my model to detect it ??
Relevant answer
Answer
Question
4 answers
What algorithms deal with deep learning based on a convolutional neural network for medical image fusion?
  • asked a question related to Deep Learning
Question
6 answers
Dear Colleague and experts.
I am looking for new laptop with a GPU card for deep learning mission.
is there any criteria to make me choose the proper GPU card based on my dataset size and image resolution. is there any rough estimation to expected training time based on card parameters and dataset size?
Thanks for any help
Relevant answer
Question
12 answers
Hi every one. Anyone pls suggest me what is the best segmentation algorithm for segmenting leaf region from the real field images. Consider that real field images contains too many unwanted objects like other leaf images, branches, human parts ( like fingers, shoes) etc. And also kindly provide me MATLAB code for this if possible. Thank you.
  • asked a question related to Deep Learning
Question
13 answers
Accurate image captioning with the use of multimodal neural networks has been a hot topic in the field of Deep Learning. I have been working with several of these approaches and the algorithms seem to give very promising results.
But when it comes to using image captioning in real world applications, most of the time only a few are mentioned such as hearing aid for the blind and content generation.
I'm really interested to know if there are any other good applications (already existing or potential) where image captioning can be used either directly or as a support process. Would love to hear some ideas.
Thanks in advance.
  • asked a question related to Deep Learning
Question
48 answers
I search through the internet it seems to me that Python is better to work on it. Share your opinion with experience if you agree or not.
Relevant answer
Answer
Python is the easiest language to work with Deep Learning.
  • asked a question related to Deep Learning
Question
4 answers
I intend to ensemble Multi-layer perceptron (MLP) and Convolutional Neural Network for urban land use and land cover classification and change detection but don't have enough high resolution image for time series analysis. The free Sentinel-2 image with 10m resolution is only about 6 years now and am looking at monitoring changes from 1996 to date. The only available open access image i could find is Landsat images with 30m resolution. So i don't know how suitable this can be in Deep learning such as CNN.
Relevant answer
Answer
Thank you all for your valid answers and contributions
Question
4 answers
What is the difference between using Matlab and Python in the field of deep learning?
Relevant answer
Answer
Advantages of Matlab
MatLab has a large number of committed users which include many universities and a few companies who have the budget to buy a license for the program. Even though it is used in many universities, Matlab is easy for beginners who are just starting to learn about programming language because the package, when purchased, includes all that you will need.
When using Python you are required to install extra packages. One part of MatLab is a product called Simulink, which is a core part of the MatLab package for which there does not yet exist a good alternative in other programming languages.
Disadvantages of Matlab
Disadvantage is its cost of License. Its very costly user has to buy each and every module and pay for it. Disadvantage is during cross compiling or converting Matlab to other language code is very difficult. Its very difficult or requires deep devel Matlab knowledge to deal with all errors.
Matlab is not suggested to make any product. Because, Matlab doesn’t create application deployment like task (like setup files and other executable which copies during installation).
Advantages of Python
The Python language has diversified application in the software development companies such as in gaming, web frameworks and applications, language development, prototyping, graphic design applications, etc.
– User Friendly and Easy to learn – Cross platform supported – Vast community support – Very powerful – Open source
– Python Packages Index ( PyPI ) – hosts thousands of third-party modules for python.
Applications
  • Web and Internet Development
  • Database Access
  • Desktops GUIs
  • Scientific and Numeric
  • Education
  • Network Programming
  • Software and Game Development
This provides the language a higher plethora over other programming languages used in the industry. Some of its advantages in details are-
Extensive Support Libraries It provides large standard libraries that include the areas like string operations, Internet, web service tools, operating system interfaces and protocols. Most of the highly used programming tasks are already scripted into it that limits the length of the codes to be written in Python.
Integration Feature Python integrates the Enterprise Application Integration that makes it easy to develop Web services by invoking COM or CORBA components. It has powerful control capabilities as it calls directly through C, C++ or Java via Jython. Python also processes XML and other markup languages as it can run on all modern operating systems through same byte code.
Improved Programmer’s Productivity The language has extensive support libraries and clean object-oriented designs that increase two to ten fold of programmer’s productivity while using the languages like Java, VB, Perl, C, C++ and C#.
Productivity With its strong process integration features, unit testing framework and enhanced control capabilities contribute towards the increased speed for most applications and productivity of applications. It is a great option for building scalable multi-protocol network applications.
Disadvantages of Python
Python has varied advantageous features, and programmers prefer this language to other programming languages because it is easy to learn and code too.
However, this language has still not made its place in some computing arenas that includes Enterprise Development Shops. Therefore, this language may not solve some of the enterprise solutions, and limitations include-
Difficulty in Using Other Languages The Python lovers become so accustomed to its features and its extensive libraries, so they face problem in learning or working on other programming languages. Python experts may see the declaring of cast “values” or variable “types”, syntactic requirements of adding curly braces or semi colons as an onerous task.
Weak in Mobile Computing Python has made its presence on many desktop and server platforms, but it is seen as a weak language for mobile computing. This is the reason very few mobile applications are built in it like Carbonnelle.
Gets Slow in Speed Python executes with the help of an interpreter instead of the compiler, which causes it to slow down because compilation and execution help it to work normally. On the other hand, it can be seen that it is fast for many web applications too.
Run-time Errors The Python language is dynamically typed so it has many design restrictions that are reported by some Python developers. It is even seen that it requires more testing time, and the errors show up when the applications are finally run.
Underdeveloped Database Access Layers As compared to the popular technologies like JDBC and ODBC, the Python’s database access layer is found to be bit underdeveloped and primitive. However, it cannot be applied in the enterprises that need smooth interaction of complex legacy data.
Let’s consist a small combination of them – following can be incredibly useful –
MATLAB
– Invaluable for signal processing – Incredibly broad array of useful libraries – Simplest and most concise language for anything involving matrix operations – Works very well for anything that is simply represented as a numeric feature matrix – Huge pain to use for anything that isn’t simply represented as a numeric feature matrix – Lacking a good open source ecosyste
Python
– Very fragmented but comprehensive scientific computing stack – Pandas, scikit.learn, numpy, scipy, ipython, & matplotlib are my most-used scientific computing libraries – IPython notebook makes a nice interactive data analysis tool – All the benefits of a general purpose programming language – Unfortunately slow if you don’t drop into C – Some of the scientific computing stack is still stuck in Python 2.7 – Very good for problems that don’t come as a simple feature matrix, between tools like pandas and nltk – Incredible open source ecosystem
Python is most popular language in the AI field.
Why ? Because –
Python comes with a huge amount of libraries. Many of the libraries are for Artificial Intelligence and Machine Learning. Some of the libraries are Tensorflow (which is high-level neural network library), scikit-learn (for data mining, data analysis and machine learning), pylearn2 (more flexible than scikit-learn), etc. The list keeps going and never ends.
For other languages, students and researchers need to get to know the language before getting into ML or AI with that language. This is not the case with python. Even a programmer with vert basic knowledge can easily handle python.
Apart from that, the time someone spends on writing and debugging code in python is way less when compared to C, C++ or Java. This is exactly the students of AI and ML wants. They don’t want to spend time on debugging the code for syntax errors, they want to spend more time on their algorithms and heuristics related to AI and ML. Not just the libraries but their tutorials, handling of interfaces are easily available online. People build their own libraries and upload them on GitHub or elsewhere to be used by others
Python has a solid claim to being the fastest growing major programming language. Recommended to check ground breaking statistics on incredible growth of python and why is python growing so quickly from stack overflow.
Advantages of Python over Matlab
1. Python code is more compact and easier to read than Matlab code —- Unlike Matlab, which uses end statement to indicate the end of a block, Python determines block size based on indentation. —- Python uses square brackets for indexing and parentheses for functions and methods, whereas Matlab uses parentheses for both, making Matlab more difficult to differentiate and understand. —- Python’s better readability leads to fewer bugs and faster debugging.
2. While most programming languages, including Python, use zero-based indexing, Matlab uses one-based indexing making it more confusing for users to translate.
3. The object-oriented programming (OOP) in Python is simple flexibility while Matlab’s OOP scheme is complex and confusing
4. Python is free and open —- While Python is open source programming, much of Matlab is closed —- The developers of Python encourage users to input suggestions for the software, while the developers of Matlab offer no such interaction
5. There is no Matlab counterpart to Python’s import statement 6. Python offers a wider set of choices in graphics package and toolsets
In Steve Hanly’s research on the speed test between Python and MATLAB for vibration analysis
Utilization of Python
Python has been gaining momentum as being the programming language for novice users. Highly ranked Computer Science departments at MIT and UC Berkeley use Python to teach their novice programming language students. The three largest Massive Open Online Course (MOOC) providers (edX, Coursera andUdacity) all use Python as their programming language for their beginning courses in programming. A variety of professors in other disciplines now utilize the need for novice students to understand Python and its key features.
Conclusion
There is no such thing as a ‘best language for machine learning’
Popularity is not a good yardstick to use when selecting a programming language for machine learning and data science. There is no such thing as a ‘best language for machine learning’ and it all depends on what you want to build, where you’re coming from and why you got involved in machine learning.
In most cases developers port the language they were already using into machine learning, especially if they are to use it in projects adjacent to their previous work?—?such as engineering projects for C/C++ developers or web visualizations for JavaScript developers.
Question
4 answers
I want to implement only a Gate similar to forget gate or output gate of LSTM cell. Mind it...!!!
I am not implementing the entire LSTM cell rather I have certain input and previous output and I want to pass them through a Gate with tanh or sigmoid as activation. Is there a way to do that in Keras or in Pytorch. Please suggest.
Relevant answer
Answer
Yes, It can be implemented with python in keras. Although i have not implemented it personally, but I can suggest something which might be useful. You can write a class in python for the GATE that you want to program. You can use normal numpy operations to write the function. Algorithm would look something like this:
Input: prev_output, current_input;
Output: activated_output;
preactivated_output=prev_output+current_input
activated_output=sigmoid(preactivated_output)
You might have to do vector embedding so that the size of the vectors are consistent. You can try this approach and good luck.
Question
6 answers
As in the attached diagram in the image below, it shows the amount of data and performance in using traditional methods or deep learning.
What are the limits to the amount of data with which we can make the decision to use traditional methods or deep learning?
Question
10 answers
What are the motivations for using deep learning instead of traditional techniques in Medical image segmentation?
Relevant answer
Answer
All the previous respondents addressed interesting issues but
@Shipra Suman has touched the subject of feature engineering. Deep learning allows to handle features that cannot be currently understood. And the dimensionality is high.
The fact that many problems are solved fast and well should not obliterate meaning. We need this type of analysis to handle semantics and semiotics.
  • asked a question related to Deep Learning
Question
9 answers
I have tried tabula-py library and java tool so far but it results in many false positives ( i.e. telling that a table is present when not the case).
Some of the cases were
content 1 content 3
content 2 content 4
If text is written in the above manner, then also it marks it as tabular data. Is there any solution that does the task better and handles the above problem. ( including Deep learning or other techniques).
Relevant answer
Answer
The Excalibur, which is built on top of camelot:
Best Software to Extract Tables from PDF (and export them to Excel, CSV, …)
  • asked a question related to Deep Learning
Question
3 answers
Recently, I read about 'PyCaret' ( https://pycaret.org/ ), a low code ML library for python that claims to be easy and fastest for users. Is there any other platforms too available?? Any suggestions??
Relevant answer
Answer
You are welcome.
Question
5 answers
The universal approximation theorem states that: " For any continuous bounded function with compact domain X and any threshold ε we can find a neural network N with a single hidden layer that gives an approximation of the function to the specific threshold ε. "
But, we also know that Taylor and Fourier series expansions can be used to achieve the same purpose. So why are deep neural networks seem to be considered more efficient on many tasks. For instance, the face recognition task.
The universal approximation theorem seems general and doesn't say a lot about the nature of the function to approximate. Is there a mathematical explanation for such success of deep learning?
Relevant answer
Answer
Deep learning methods have changed the artificial intelligence world over the last few years. The skills and strategies that humans once believed were special to us have started to fall under the assault of ever more powerful machines.
One by one. In such tasks as facial reconnaissance and object recognition, deep neural networks are much stronger than humans. The old Games were perfected and the best human players thrashed.
Yet a dilemma persists. There are no mathematical explanations why layered
networks should be so good at this. They're flummoxed mathematicians. No one knows how they succeed despite the massive success of deep neural networks.
That is changing today through Henry Lin 's work at Harvard and Max Tegmark 's work at MIT. These guys claim that the source of mathematicians' shame is that the answer is decided by the existence of the universe. In other words, the answer is not mathematics but physics.
First, let's set the issue with a megabit grayscale classification example
to figure out if a cat or a dog is showing it. A million-pixel image consists of one of 256 gray-scale values. Thus in principle, 256100000 images can be made, and if it displays a cat or a dog for each, it must be measured. Neural networks, with just thousands or millions of parameters, always handle this classification role in some way easily.
Neural networks operate in the language of mathematics by approximating complex math functions with simpler functions. When the images of cats and dogs are categorised, the neural network must perform a process that takes a million grayscale pixels as input and outputs the distribution of probabilities.
The problem is that order of magnitude exists for approximating more mathematical functions than possible networks. But somehow deep neural
networks get the right reply. Lin and Tegmark are now saying that they worked out why. The response is that a limited sub-set of all possible functions rules the universe. In other words, when mathematical writing is carried out of the laws of
physics, all functions with a remarkable set of simple properties can be defined.
Therefore profound neural networks should only approximate a small subset of mathematical functions. Take the order of a polynomial function that is the height of its highest exponent into account. Thus a quadratic equation like y = x2 is 2, the equation y = x24 is 24 etc.
The number of orders obviously is limitless and yet in physics laws only a small subset of polynomials appear. "Our universe can precisely be represented by low-order polynomial Hamiltonians due to reasons that are still not fully understood," Lin and Tegmark claim. The polynomials representing physics laws usually have orders ranging from 2 to 4.
Physics laws have other major characteristics. In rotation and translation, for
instance, they are typically symmetrical. Turn a cat or dog by 360 degrees and it looks the same, translate by about 10 metres. Translate it by 100 metres. This makes it easier to approximate the identification process for cats or dogs.
These characteristics mean that neural networks do not actually have to approximate an infinity of potential mathematics, just a small subset of the easiest functions.
The neural networks manipulate another property of the universe. This is the systemic hierarchy. 'Primary particles form atoms that in turn form chemical, cellular, organism, planetary, nuclear, galaxy, etc.' Lin and Tegmark say so. In several instances, complex structures are built in a series of simple steps.
That is why it is also important to structure neural networks, because each stage in the causal sequence can be approximated by these networks.
The Big Bang echoes that pervade the universe, Lin and Tegmark give the example of the cosmic radiation from the background microwave. Different spacecrafts have mapped this radiation in growing resolution in recent years.
And physicists have, of course, perplexed as to why these maps are taking shape.
Tegmark and Lin stress that, for whatever reason, the causal hierarchy is inevitably the result. "The power spectrum of variations in density within our universe is determined by a number of cosmological parameters (density of dark matter, etc.) which in turn determine the pattern of cosmic radiation from our early universe which is coupled with our galactic earth radio noise to make the frequency-dependent cell maps reported by the satellite-based .
Each of these causal layers gradually contains more data. The maps and the noise that they provide are made up of milliards of numbers just a handful of cosmological parameters. Physics is aimed at evaluating large numbers in a way that displays smaller numbers.
This work, with major consequences, is fascinating and relevant. Neural artificial networks are well-known for organic networks. So Lin and Tegmark not only explain why deep learning machines work so well but also why human brains can make sense of the universe. Evolution has focused on an optimal brain system to dismantle the mystery of the universe.
This work opens the way for important artificial intelligence advances. Now that we understand finally why deep neural networks function so well, mathematicians
can explore the unique math properties that cause them to perform so well.
Lin and Tegmark say, "Reinforcing the empirical awareness of deep learning will suggest ways to develop it."
  • asked a question related to Deep Learning
Question
2 answers
This is more of a survey question than a query for precise mathematical detailing. Opinions are welcome!
Relevant answer
Answer
Follow
Question
14 answers
Hi everyone. I found that none of the dataset available publicly for identification and classification of plant leaf diseases except PlantVillage dataset. So I started this discussion to help researchers who works on this area. And who are already completed their work in this area kindly provide your dataset link here so that it will helpful for the researchers who are having intrest in this area. Thank you.
  • asked a question related to Deep Learning
Question
10 answers
Hi,
as deep learning is a data-driven approach, the crucial is to have quality data. There exist a lot of datasets for free, but they differ in the quality of labels.
I'm now working on an index, which can tell a researcher quality of the labels, so the researcher may decide if such a dataset is useful nor not. I do have established a pipeline on how to produce such an index in a fully autonomous way. Note, I'm focusing on object detection tasks only, i.e., labels given as bounding-boxes.
The question is: does such the index exist already? I googled a lot and find nothing. It would be nice to compare our approach with existing ones.
Relevant answer
  • asked a question related to Deep Learning
Question
20 answers
Where can I find be any data set whichever human, animal, face thermal camera image dataset?
Relevant answer
Answer
Dear;
You can read it in the following link:
Regards
  • asked a question related to Deep Learning
Question
5 answers
I am working on a real time text extraction problem , where I have the option of either capturing image of an object or take a video of the object and then do the text extraction.
I am trying to understand the advantages and disadvantages of both the methods, like in taking a photo of the object, the problem could be image quality, the advantage could be the time taken to process the image.
Similarly in video the image quality may be better, but selecting the best frame could be a challenge, also computationally it looks to be more intensive.
Can anyone list down potential advantages or disadvantages with both the approach?
Relevant answer
Answer
Text information present in pictures and video contain valuable info. Text extraction from image has stages of detection the text from given image, finding the text location, extraction, improvement and recognition of text from the given image. But variations of text just like the variations in orientation, size, style, alignment; low size image distinction and a lot of difficult background create the matter of automatic text extraction extraordinarily troublesome.
The number of techniques and Methodology are planned to this downside.
You can take a look in these articles:
  • asked a question related to Deep Learning
Question
2 answers
hi everybody
I converted each word into a vector with one-hot encoding. So that each haplotype has become a two-dimensional matrix. I need CNN to classify them.
I appreciate your help.
Best regards
Relevant answer
Answer
You can use pre-trained models such as
VGG16,Xception,VGG19,ResNet50,ResNet101,InceptionV3,MobileNet,MobileNetV2,DenseNet121,DenseNet169,DenseNet201,ResNet50V2,ResNet101V2,ResNet152V2,NASNetMobile,NASNetLarge,EfficientNetB0,EfficientNetB7
  • asked a question related to Deep Learning
Question
9 answers
I have been actively working on medical image classification since my undergrad thesis started and collaborated in two projects on Covid-19 Chest x-ray image classification. Now I am looking for new ideas besides classification. Can anyone please help me?
Relevant answer
Answer
In my view, you should start by collecting the real medical data of COVID 19 patients with their pre-medical conditions. Maybe this will help you to make a connection between the previous diseases and their effect on a COVID 19 patient. You can develop a Deep learning scheme based on these data sets, which may be extremely helpful in the future for a COVID patient to analyze their vulnerability.
  • asked a question related to Deep Learning
Question
14 answers
Hi,
I want to implement a binary classifier (lesion yes/no) with the DeepLesion dataset. Therefore I also need ct images of healthy subjects because the DeepLesion dataset only includes images with lesions.
Thanks for your suggenstions!
  • asked a question related to Deep Learning
Question
8 answers
I'm looking forward to using the latest PlantVillage dataset to detect plant diseases using deep learning technique. However, I'm only able to access previous versions of the dataset with less classes/records. Can somebody please direct me to the latest PlantVillage dataset which comprises 58 classes and 87,848 images?
Question
8 answers
I plan to implement 3D image segmentation. I've features extracted from an unsupervised feature extraction method. The features have lower dimension than that of the input image. Which segmentation method suits best for this use case? I plan to implement it in python.
  • asked a question related to Deep Learning
Question
15 answers
When i start doing literature i observed that most of the authors used pre-trained CNN models for the identification and classification of plant leaf diseases. Why none of the authors concentrated to designed customised CNN model for this problem? is there any particular reason?
Relevant answer
Answer
To make a robust CNN model it requires a large number of pre-labeled data sets. Which is quite difficult while tackling a practical problem. Moreover, the training of a model is a time-consuming process and requires a large storage space, which is quite costly in general.
You are right that a pre-trained model is highly customized with its parameters and size but the impact is low when you trained over a large number of data sets.
  • asked a question related to Deep Learning
Question
5 answers
I have searched and found many available datasets for Software Defects Prediction. But all of them are module(Class,File) level defects detection datasets.
Are there any datasets available for Statement Level Defects Prediction. Or that can point to specific location of error in file?
Relevant answer
Answer
With some dataset that have features of some certain bug or that are targeting specific variable type error like Null Pointer dereference. Can predict if that bug is there or not which is statement level that is what i was trying to say.
In this case if you know such dataset which target a specific bug or variable level for defect prediction please mention it.
I have read some papers who did this without providing the dataset.
  • asked a question related to Deep Learning
Question
3 answers
I am looking at the following factors:
How seriously they take healthcare?
What factors affect their behavior like severity or cost, etc?
Also, other parameters that help in deciding the patient's mindset.
Relevant answer
Answer
you can find data on kaggle and uci repositary.
  • asked a question related to Deep Learning
Question
9 answers
Hi Everyone. Please suggest me the SCI journals list which don't contains article processing charges.
Relevant answer
Answer
Dear Srinivas Talasila, you can check here to find out your desired SCI Journal without APC: https://www.letpub.com/index.php?page=journalapp
Question
7 answers
I was performing a binary classification problem with 15000 RGB images using a scratch build CNN model. While it comes to evaluate the model, I can do it in two ways:
1. Splitting data Train and Test and use 10 fold cross-validation for the training data. Later with the best model, I would use the unseen Test data. In this way I got appx. 91.5% avg. accuracy for both test and validation.
2. Just use 10 fold cross-validation and got 92.5% avg accuracy(slightly better result than the previous one.)
Which option would be the best for reporting the performance of my model in the research article?
TIA
Relevant answer
Answer
Muhammad Ali Edward Peake Thanks for sharing. I will check them out.
  • asked a question related to Deep Learning
Question
3 answers
Can anyone provide the full python code for implementing alexnet models in COVID-19 XRAY image dataset?
Question
7 answers
Matlab tools for Deeep Learning allow us to solve classification tasks with static features (e.g., pixels values of images) and tasks with dynamic features (time-series). For time-series it is recommended to use LSTM (Long Short Time Memory) layer. But how to build model, if we want to use simultaneously time-series parameters and some static parameters?
Naturally, it is possible to create a fictitious time series (with constant values) from each value of a single static parameter, but this will be very inefficient.
Thanks for your answers beforehand. Regards, Sergey.
Relevant answer
Answer
Hi Sergey,
Have you taken a look at so-called spatial-temporal deep learning models?
A survey about such models is avalable by following:
It is quite possible to incorporate higher dimensional topological features into these models. To get a better idea, refer to the following papers:
Regards
  • asked a question related to Deep Learning
Question
17 answers
End devices, such as Internet-of-Things sensors, are generating most of the data. The amount of data created by these connected IoT devices will see a compound annual growth rate (CAGR) of 28.7% over the 2018-2025 forecast period, and IDC estimates that there will be 41.6 billion connected IoT devices, or "things," generating 79.4 zettabytes (ZB) of data in 2025[1]. A lot of those data needs to be analyzed in real-time using deep learning models. However, deep learning inference and training require substantial computation resources to work quickly.
There are currently two research directions to deploy deep learning models onto IoT devices to address requirements like scalability, privacy concern, network latency, bandwidth efficiency, etc.
Firstly, at the framework level, where data were trained on devices with higher computational ability, then the deep learning model was compressed to fit onto computationally weak end devices. Examples are DeepSense[2], TinyML[3], DeepThings[4], DeepIoT[5], and others. Secondly, some novel system-on-chip designs aim to solve the demand for deep learning at edge devices at the ASIC chip-level, running the deep learning algorithms from the deep ground. Those chips were designed by highly coupling the computation and memory resources, exploiting deep learning, and computing's vast inherent parallelism and computing at only required numerical precision. Examples are Ergo[6], which can process large neural networks in 20mW and supports a wide variety of currently popular styles of advanced neural networks, including CNNs, RNNs, LSTMs, and more. And also, NDP100/101[7], ECM3531/3532[8], etc.
Both approaches have their advantages regarding the cost, efficiency, effort for implementation, etc. Which one do you think will predominate the development of deep learning on edge devices? I am not asking for an answer to choose A or B, but a discussion of your preference if you are/were working on related topics.
[2] DeepSense: A unified deep learning framework for time-series mobile sensing data processing
[3] TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers
[4] DeepThings: Distributed Adaptive Deep Learning Inference on Resource-Constrained IoT Edge Clusters
[5]. DeepIoT: Compressing Deep Neural Network Structures for Sensing Systems with a Compressor-Critic Framework
Relevant answer
Answer
Interesting question, and from what you write in the description, you should add to it the issue of "IoT at the Edge with mMTC" (massive machine type communication).
There may be an interesting scenario to consider: Quantization at the Edge prior to either aggregated distributed analysis or centralised analysis.
For instance a parameter value is between vmin and vmax, or above, or below. This is quantizable by a ternary model (-1,0,1) as states.
You could have a more refined n level quantizer, or if you collect vectors instead of just values, vector quantizers (see vectir quantization, Voronoi diagram, etc).
The algorithms on quantized data can be made light and fast (replace * by + and integer based representations, GF(n) for Galois fields also noted Z/nZ structures). This gives great algebra decompositions.
It fits on different hardware/VLSI.
You get one more degree of freedom through this quantized approach.
See my articles (on researchgate) on variable rate speech coding, and my patent on algebraic speech coding (method generalising to vector processing)
  • asked a question related to Deep Learning
Question
8 answers
I'd like to find a simple way to understand the deep learning vs machine learning
Relevant answer
Answer
Dear Fahad
Thanks for your answer
Question
10 answers
So far deep learning models depend on powerful computing resources.
Do you think deep learning architectures have the capacity to be incorporated in the resource constrained devices or edge devices such that we can have some sort of on-device analytics?
If so, which method would be best suited for that? Any research on this area?
Relevant answer
Answer
In the last two years, there has been several frameworks published aiming to compress the neural network model for an IoT edge devices, examples are DeepSense[1], TinyML[2], DeepThings[3], DeepIoT[4]. There are also some system on chip design, aiming to address the deep learning model on edge IoT at the chip level, like Ergo, NDP100/101, ECM3531/3532. There are indeed plenty of use cases at the edge IoT devices waiting for a deep learning solution, to solve the problem of privacy concern, network latency, etc. The problem is which solution you would like to choose for the cases.
If you would like to discuss this question, please share your ideas to this link, so everybody in this community can read your discussion:
[1] DeepSense: A unified deep learning framework for time-series mobile sensing data processing
[2] TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers
[3] DeepThings: Distributed Adaptive Deep Learning Inference on Resource-Constrained IoT Edge Clusters
[4]. DeepIoT: Compressing Deep Neural Network Structures for Sensing Systems with a Compressor-Critic Framework
  • asked a question related to Deep Learning
Question
3 answers
Good day friends.
I have a few questions regarding my project.
The project is mainly focused on IoT, network traffic classification, intrusion, machines learning/deep learning.
So, my questions, here, is:
1) Is it possible to test IoT datasets that were captured in Australia in other countries, since I want the study tp base in specific country context?
2) The study settings will be based on a smart city (smart toll tag) or a smart environment. But, the datasets I’m planning to use which are captured in Australia, are datasets of weather station, (which generate air pressure, temperature, etc), smart 顺心彩票 ( motion activated light), smart fridge, remotely activated garages door, smart air conditioner (thermostat), etc.
What do you guys think? Is it possible?
Thanks!
Relevant answer
Answer
Hello, I have doubt around it, maybe becaus I do not know in detail your idea. However, on the one hand, when we are talking about IoT, the idea is associated with a real-time data processing. On the other hand, when we talk about datasets (even very large), we are referring to a batch data processing model. If you want to recreate a new environment using data from monitoring stations...how do you ensure the consistency of both processing environments? Mainly thinking about the kind of available resources and algorithms to use.
Question
4 answers
what useful information can be extracted from a saved model file regarding the training data.
From security perspective too. If someone has access to the saved model what information can they gain?
Question
3 answers
I've implemented a stacked Independent subpace analysis(ISA) network similar to the one in research paper mentioned below for unsupervised feature extraction. It consists of 2 ISA layers stacked one after another. Each ISA layer has two two layers in it. The first layer of the ISA captures square nonlineaity relationships in the input image patch. The second layer groups the response with respect to subpace size. It is very similar to Independent component analysis except it groups dependant sources together and these dependent groups are independent of each other. PCA is used at the beginning of each layer for dimensionality reduction.
My goal is visualize the features (weights at each layer) and the feature maps. Since the dimensions are reduced, the visualization is not straight forward. What is the best way for me to visualize the feature maps/features?
Also, how was figure 4 achieved in this paper?
  • asked a question related to Deep Learning
Question
4 answers
Can anyone provide the full python code for implementing different dense net models in COVID-19 XRAY image dataset?
Relevant answer
Answer
good luck
  • asked a question related to Deep Learning
Question
4 answers
Can anyone provide the full python code for implementing inception net in image dataset
Relevant answer
Answer
good luck
  • asked a question related to Deep Learning
Question
6 answers
The audio recognition is one of the deep learning application
For fast training , the audio signal is divided into set of rows before the feature extraction and running the deep learning model are done
How we can do this division in a python code ??
Relevant answer
Answer
@Reema Ahmad
In my opinion you can use directly or you do it the pre-processing for you train dataset by Wavelet Transfomation.
I suggest you to watch this video for Dr. Ajay Verma
I hope that be Claire for you an helpful.
Question
3 answers
Hello all,
At the moment, I am working on a project about the failure/fault detection on Photovoltaic Module based on Deep Learning using electroluminescence (EL) images. In this project, my proposed method of data reprocessing before applying to the Deep Learning model is inspired on what is currently used in Healthcare for X-ray images: for X-ray images, the reprocessing method is first applying the Intensity Normalization then Contrast Limited Adaptive Histogram Equalization.
My question is: how do I compare between the EL images and the X-ray i.e. compare images from different modalities? The results of this comparison can give me a first glance whether this proposed method is viable.
Relevant answer
Answer
When you have images obtained with different imaging modalities, you need to normalize them in terms of coordinates and image transformations (e.g., translation, rotation, and scaling).
Your images may have some vision artifacts that can impair the comparison.
This is called image pre-processing and it is frequently underestimated in designs.
Regards
Question
1 answer
How to get access to LYON19 dataset? histopathological image dataset of Lymphocytes provided by LYON19 challenge is not accessible from its challenge site https://lyon19.grand-challenge.org/Background/ .
Relevant answer