Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 5 years ago.
Improve this question
There's two things I don't know how to do about them.
I trained the network with alexnet, it couldn't converge. And there was no overfiting. However, when I trained the network with lenet, it converged. Why is that?
I trained the network to solve a problem. It converged, however not enough. The softmax loss can't be lower(0.2). And how can I deal about it.
An inability to overfit may be caused by a lack of network capacity or a bug in your code itself. Are you training the model from scratch? Did you write the model yourself? Are you working on a task at a similar scale to Imagenet classification?
Plateauing of a loss is to be expecting during any kind of model training. But without any learning curves showing the loss across epochs, a loss of "0.2" is meaningless.
Without knowing the exact task your're working on, the framework you're using, source code, learning curves, and/or any debugging steps you've taken, nobody will be able to really answer your questions. I suggest you go back to the machine learning basics in the cs231 lectures (particularly Lectures 4,5) or Andrew Ng's Coursera course for a more rudimentary knowledge.
Related
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
We don’t allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Closed 5 years ago.
Improve this question
I am new to Machine Learning. I had learnt about ANN, SVR and ANN optmised with GA and PSO, but I realised that ANN involving long run time due to high computational effort. Is there any other machine learning toolbox in Matlab that perform faster than SVR?
Thanks :)
Sorry i don't know a lot on Matlab toolbox, but I can maybe help you on machine learning on a more general way. Do you need only a faster algorithm or better performing overall ?
Kind of a beginner too, but i'd say it really depends if :
it's a regression, classification or clustering task.
The complexity of the goal you want to achieve.
The Dataset you have.
I came across a microsoft post to help you chose your algorithm that is quite cool as a starting point, you should take a look, there is a graph that answers your question :
https://learn.microsoft.com/en-us/azure/machine-learning/studio/algorithm-choice
Some other will have better advices than me but heh that's a start.
Good luck,
Marc.
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 7 years ago.
Improve this question
I can access MatLab on school computers, but it would be more like a convenience factor as of right now. I want to get into programming using raspberry pi later on, but is it worth buying in the meantime? Will it be useful to me as a computer engineer? Thanks.
I have used Matlab as a student and employee for over 10 years now. I have recently taken an interest in Python as a free open-source alternative to Matlab and have been extremely impressed. Matlab is a very powerful and convenient tool, but has a significant cost, especially if toolboxes are needed. Python seems to cover all the same capability needs that I have in Matlab (with toolboxes), and if I ever need help doing something in python that I know how to do in matlab, I just do a quick google search and 9 times out of 10 I end up on this website, taking advantage of the questions that others have already answered.
With your question on Raspberry pi, it can actually run python code on board, but not the case for matlab.
I am not a computer engineer, but am of the opinion that python would likely benefit you more than matlab in the end as it is free, more widely used and can be used on devices such as raspberry pi. For reference, I have degrees in physics and electrical engineering and spend a great deal of time at work building optical sensors, operating them, and processing their data.
One more note: I like the Anaconda distribution of python, highly recommended
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 9 years ago.
Improve this question
I am looking for ideas for my undergraduate project and I quite like the area of High Performance Computing , has got a lot of scope for research . Are there any ideas / already existing open source projects worth looking at ?
One hot field right now is in the area of algorithmic trading. You can sign up for $3000 (if you're under 21 -- it's $10k for over 21) at InteractiveBrokers.com and they will give you a free paper trading account (which is fake money traded using realtime data) of $10,000,000. They have API's in C#, C++, VB, Java and reasonable support... You could write your own stock pair trading algorithm. They have good documentation on how to get started.
You can scale this as high as you want, also a lot of people do high frequency trading which requires hpc and in-depth knowledge of Unix and C++.
Worth looking into, my 2 cents.
Perhaps massively parallel processing? Libraries like Cuda, OpenCL, and DirectCompute are just blossoming, and have a high likelihood of becoming commonplace. In my company, we are researching uses for OpenCL, and we're finding that it has the potential to revolutionize our industry.
Just a thought.
I would suggest looking at Sandia National Labs's SST (The Structural Simulation Toolkit). Its a highly parallel simulator framework used for HPC. It uses and incorporates other simulators from academia and industry. For instance, it currently integrates GEM5, QSim, MacSim, DRAMSim, Merlin, Portals, DRAMSim2, Iris, etc. Moreover, it is open source so you can contribute to the development.
You could work on integrating other academia components into SST, improve the interface of one of the components with SST, or just improve of the components themselves.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 8 years ago.
Improve this question
I'm learning about multilayer perceptrons, and looking to tackle some real-world problems in Matlab. Perhaps something like medical diagnosis, or speech recognition... But I'm not really sure where to start here. How should I tackle this?
Geoff Hinton is the man when it comes to multilayer perceptrons. His Science paper from 2006 used a special class of MLP called an "autoencoder" that was successful in digit recognition, facial recognition, and document classification (all of which have real world applications):
Reducing the Dimensionality of Data with Neural Networks
Fortunately, they also published the
Matlab code.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 5 years ago.
Improve this question
I'm preparing to give a talk about some topic I choose from artificial intelligence area (neural networks). I'm looking for something interesting, used in a real life and preferably not too complicated (the simpler it is, the easier it is for students to understand and the more interested they will be). I thought that it's a good place to look for advice ;)
Code applying neural networks to text recognition.
I think the concept of text recognition is interesting and understandable.
Toby Segaran's interesting book "Programming Collective Intelligence" contains a simple neural net example for learning search results relevancy. He offers the code from the book free on his site.
The neural net is in chapter4 code. Not sure if you could figure out the code without the text - if you don't mind spending a little money, the book certainly wouldn't hurt.
Learn your neural net the sine wave. It's simple. You only need 4 neurons. And the weights will clearly show how it's working. It was the example that made it click for me.
The "Real life applications" section of the (english) Wikipedia article about "Artificial neural networks" lists some (quite general) applications of neural networks.