I'm just starting with Torch and neural networks and just glancing at a lot of sample code and tutorials, I see a lot of variety in the how people structure their neural networks. There are layers like Linear(), Tanh(), Sigmoid() as well as criterions like MSE, ClassNLL, MultiMargin, etc.
I'm wondering what kind of factors people keep in mind when creating the structure of their network? For example, I know that in a ClassNLLCriterion, you want to have the last layer of your network be a LogSoftMax() layer so that you can input the right log probabilities.
Are there any other general rules or guidelines when it comes to creating these networks?
Thanks
Here is a good webpage which contains the pros and cons of some of the main activation functions;
http://cs231n.github.io/neural-networks-1/#actfun
It can boil down to the problem at hand and knowing what to do when something goes wrong. As an example, if you have a huge dataset and you can't churn through it terribly quickly then a ReLU might be better in order to quickly get to a local minimum. However you could find that some of the ReLU units "die" so you might want to keep a track on the proportion of activated neurons in that particular layer to make sure this hasn't happened.
In terms of criterions, they are also problem specific but a bit less ambiguous. For example, binary cross entropy for binary classification, MSE for regression etc. It really depends on the objective of the whole project.
For the overall network architecture, I personally find it can be a case of trying out different architectures and seeing which ones work and which don't on your test set. If you think that the problem at hand is terribly complex and you need a complex network to solve the problem then you will probably want to try making a very deep network to begin with, then add/remove a few layers at a time to see if you have under/overfitted. As another example, if you are using convolutional network and the input is relatively small then you might try and use a smaller set of convolutional filters to begin with.
Related
Recently, I was asked about how to pre-train a deep neural network with unlabeled data, meaning, instead of initializing the model weight with small random numbers, we set initial weight from a pretrained model (with unlabeled data).
Well, intuitively, I kinda get it, it probably helps with the vanishing gradient issue and shorten the training time when there are not too much labeled data available. But still, I don't really know how it is done, how can you train a neural network with unlabeled data? Is it something like SOM or Boltzmann machine?
Has anybody heard about this? If yes, can you provide some links to sources or papers. I am curious. Greatly appreciate!
There are lots of ways to deep-learn from unlabeled data. Layerwise pre-training was developed back in the 2000s by Geoff Hinton's group, though that's generally fallen out of favor.
More modern unsupervised deep learning methods include Auto-Encoders, Variational Auto-Encoders, and Generative Adversarial Networks. I won't dive into the details of all of them, but the simplest of these, auto-encoders, work by compressing an unlabeled input into a low dimensional real-valued representation, and using this compressed representation to reconstruct the original input. Intuitively, a compressed code that can effectively be used to recreate an input is likely to capture some useful features of said input. See here for an illustration and more detailed description. There are also plenty of examples implemented in your deep learning library of choice.
I guess in some sense any of the listed methods could be used as pre-training, e.g for preparing a network for a discriminative task like classification, though I'm not aware of that being a particularly common practice. Initialization methods, activation functions, and other optimization tricks are generally advanced enough to do well without more complicated initialization procedures.
I've already trained the neural network in Keras for detecting two classes of images (cats and dogs) and got accuracy on test data. Is it enough for the conclusion in the master thesis or should I do other actions for evaluating the quality of network (for instance, cross-validation)?
Not really, I would expect more than just accuracy from my students in any classification setup. Accuracy only evaluates that particular network on that particular test set but you would have to some extent justify the design choices you've made in building that network. Here are some things to consider:
Presumably you have some hyper-parameters you've fixed, you can investigate how these affect your results. How many filters? How many layers? and most importantly why?
An important aspect of object classification is how your model handles noise. Depending on your dataset, one simple way would be to pre-process the test data, blur it, invert colours etc and you'll see that your performance will drop. Why does it do that? How does the confusion matrix look like then?
What is the performance of the network? Is it fast, slow compared to another system, say VGG?
When you evaluate your project in general not just the network, asking why things worked helps a lot, not just why things didn't work.
I recently started learning neural networks, and I thought that creating a sudoku solver would be a nice application for NN. I started learning them with backward propagation neural network, but later I figured that there are tens of neural networks. At this point, I find it hard to learn all of them and then pick an appropriate one for my purpose. Hence, I am asking what would be a good choice for creating this solver. Can back propagation NN work here? If not, can you explain why and tell me which one can work.
Thanks!
Neural networks don't really seem to be the best way to solve sudoku, as others have already pointed out. I think a better (but also not really good/efficient) way would be to use an genetic algorithm. Genetic algorithms don't directly relate to NNs but its very useful to know how they work.
Better (with better i mean more likely to be sussessful and probably better for you to learn something new) ideas would include:
If you use a library:
Play around with the networks, try to train them to different datasets, maybe random numbers and see what you get and how you have to tune the parameters to get better results.
Try to write an image generator. I wrote a few of them and they are stil my favourite projects, with one of them i used backprop to teach a NN what x/y coordinate of the image has which color, and the other aproach combines random generated images with ine another (GAN/NEAT).
Try to use create a movie (series of images) of the network learning to create a picture. It will show you very well how backprop works and what parameter tuning does to the results and how it changes how the network gets to the result.
If you are not using a library:
Try to solve easy problems, one after the other. Use backprop or a genetic algorithm for training (whatever you have implemented).
Try to improove your implementation and change some things that nobody else cares about and see how it changes the results.
List of 'tasks' for your Network:
XOR (basically the hello world of NN)
Pole balancing problem
Simple games like pong
More complex games like flappy bird, agar.io etc.
Choose more problems that you find interesting, maybe you are into image recognition, maybe text, audio, who knows. Think of something you can/would like to be able to do and find a way to make you computer do it for you.
It's not advisable to only use your own NN implemetation, since it will probably not work properly the first few times and you'll get frustratet. Experiment with librarys and your own implementation.
Good way to find almost endless resources:
Use google search and add 'filetype:pdf' in the end in order to only show pdf files. Search for neural network, genetic algorithm, evolutional neural network.
Neither neural nets not GAs are close to ideal solutions for Sudoku. I would advise to look into Constraint Programming (eg. the Choco or Gecode solver). See https://gist.github.com/marioosh/9188179 for example. Should solve any 9x9 sudoku in a matter of milliseconds (the daily Sudokus of "Le monde" journal are created using this type of technology BTW).
There is also a famous "Dancing links" algorithm for this problem by Knuth that works very well https://en.wikipedia.org/wiki/Dancing_Links
Just like was mentioned in the comments, you probably want to take a look at convolutional networks. You basically input the sudoku bord as an two dimensional 'image'. I think using a receptive field of 3x3 would be quite interesting, and I don't really think you need more than one filter.
The harder thing is normalization: the numbers 1-9 don't have an underlying relation in sudoku, you could easily replace them by A-I for example. So they are categories, not numbers. However, one-hot encoding every output would mean a lot of inputs, so i'd stick to numerical normalization (1=0.1, 2 = 0.2, etc.)
The output of your network should be a softmax with of some kind: if you don't use softmax, and instead outupt just an x and y coordinate, then you can't assure that the outputedd square has not been filled in yet.
A numerical value should be passed along with the output, to show what number the network wants to fill in.
As PLEXATIC mentionned, neural-nets aren't really well suited for these kind of task. Genetic algorithm sounds good indeed.
However, if you still want to stick with neural-nets you could have a look at https://github.com/Kyubyong/sudoku. As answered Thomas W, 3x3 looks nice.
If you don't want to deal with CNN, you could find some answers here as well. https://www.kaggle.com/dithyrambe/neural-nets-as-sudoku-solvers
How do I approach the problem with a neural network and a intrusion detection system where by lets say we have an attack via FTP.
Lets say some one attempts to continuously try different logins via brute force attack on an ftp account.
How would I set the structure of the NN? What things do I have to consider? How would it recognise "similar approaches in the future"?
Any diagrams and input would be much appreciated.
Your question is extremely general and a good answer is a project in itself. I recommend contracting someone with experience in neural network design to help come up with an appropriate model or even tell you whether your problem is amenable to using a neural network. A few ideas, though:
Inputs need to be quantized, so start by making a list of possible numeric inputs that you could measure.
Outputs also need to be quantized and you probably can't generate a simple "Yes/no" response. Most likely you'll want to generate one or more numbers that represent a rough probability of it being an attack, perhaps broken down by category.
You'll need to accumulate a large set of training data that has been analyzed and quantized into the inputs and outputs you've designed. Figuring out the process of doing this quantization is a huge part of the overall problem.
You'll also need a large set of validation data, which should be quantized in the same way as the training data, but that should not take any part in the training, as otherwise you will simply force a correlation network that may well be completely meaningless.
Once you've completed the above, you can think about how you want to structure your network and the specific algorithms you want to use to train it. There is a wide range of literature on this topic, but, honestly, this is the simpler part of the problem. Representing the problem in a way that can be processed coherently is much more difficult.
Seeing that as as far as we know, one half of your brain is logical and the other half of your brain is emotional, and that the wants of the emotional side are fed to the logical side in order to fulfill those wants; has there been any research done in connecting two separate neural networks to one another (one trained to be emotional, and one trained to be logical) to see if it would result in almost a free-will sort of "brain"?
I don't really know anything about neural networks except that they were modeled after the biological synapses in the human brain, which is why I ask.
I'm not even sure if this would be possible considering that even a trained neural network sometimes doesn't act logically (a.k.a. do what you thought you trained it to do).
First, most modern neural networks aren't really modeled after biological synapses. They use an Artificial Neuron which allowed Back Propagation to work rather than a Perceptron which is a much more accurate representation.
When you feed the output of one network into the input of another network, you've really just created one larger network, not two separate networks. It just happens that in this case portions of the networks would be trained independently.
That said, all neural networks have to be trained. Which means you need sample input and sample output. You are looking to create a decision engine of sorts I suppose. So you would need to create a dataset where it makes sense that there might be an emotional and rational response, such as purchasing an item. You'd have to train the 'rational' network to accept as a set of inputs the output of an 'emotional' network. Which means you are really just training the rational decision engine to be responsive based on the leve of 'distress' caused by the emotional network.
Just my two cents.
I have also heard of one hemisphere being called "divergent" and one "convergent". This may not make any more sense than emotional vs logical, but it does hint at how you might model it more easily. I don't know how the brain achieves some of the impressive computational feats it does, but I wouldn't be very surprised if all revolved around balance, but maybe that is just one of the baises you have when you are a brain with two hemipheres (or any even number) :D
A balance between convergence and divergence is the crux of the creativity inherent in evolution. Replicating this with neural nets sounds promising to me. Suppose you make one learning system that generalizes and keeps representations of only the typical groups of patterns it is shown. Then you take another and make it generate all the in-betweens and mutants of the patterns it is shown. Then you feed them to eachother in a circle, and poof, you have made something really interesting!
It's even more complex than that, unbelievably. The left hemisphere works on a set of logical rules, it uses these to predict its environment and categorize input. It also infers rules and stores them for future use. The right hemisphere is based, as you said, on emotion, but also on memory of single, unique or emotionally relevant occurrences. A software implementation should also be able to retrieve and store these two data types and exchange "opinions" about them.
While the left hemisphere of the brain may be more involved in making emotional decisions, emotion itself is unlikely to occur exclusively in one side of the brain, and the interplay between emotions and rational thought within the brain is likely to be substantially more complex than having two completely separate circuits. For instance, a study on rhesus macaques found that dopamine and other hormones associated with emotional responses essentially implements temporal difference learning within the brain (I'm still looking for a link to it). This suggests that separating emotional and rational thought into two separate neural networks probably wouldn't be practical, even if we had the resources to build neural networks on the scale of brain hemispheres (which we don't, or at least not within most research budgets).
This idea is supported by Sloman and Croucher's suggestion that emotion will likely be an unavoidable emergent property of a sufficiently advanced intelligent system. Such systems (discussed in detail in the paper) will be much more complex than straight-up neural nets. More importantly, though, the emotions won't be something that you can localize to one part of the system.