This is a hypothetical question.
Assumptions
I am working on a 2 class semantic segmentation task
My ground truths are binary masks
batch size is 1
at an arbitrary point in my network there is a convolution layer called 'conv_5' which has a feature map size of 90 x 45 x 512.
Let's assume I also decide that (during training) I will concatenate the ground truth mask to 'conv_5'. This will result in a new top we can call 'concat_1' which will be a 90 x 45 x 513 dimension feature map.
Assume that the rest of the network follows a normal pattern like a few more convolution layers, a fully connected, and softmax loss.
My question is, can the fully connected layers learn to weigh the first 512 feature channels very low and weigh the last feature channel (which we know is a perfect ground truth) very high?
If this is true then is it true in principle such that if the feature dimension was 1,000,000 channels and I add the last channel as the perfect ground truth it will still learn to effectively ignore all previous 1,000,000 feature channels?
My intuition is that if there is ever a VERY good feature channel passed in then a network should be able to learn to utilize this channel far more than the others. I would also like to think that this is independent to the number of channels.
(In practice I have a scenario where I am passing in a nearly perfect ground truth as the 513th feature map, but it seems to have no impact at all. Then when I examine the magnitudes of the weights across all 513 feature channels, the magnitudes are roughly the same across all channels. This leads me to believe that the "nearly perfect mask" is only being utilized about 1/513th of it's potential. This is what has motivated me to ask the question.)
Hypothetically, if you have a "killing feature" in your disposal, the net should learn to use it and ignore the "noise" from the rest of the features.
BTW, Why are you using a fully connected layer for semantic segmentation? I'm not sure this is "a normal pattern" for semantic segmentation nets.
What may prevent the net from identifying the "killing features"?
- The layers above "conv_5" mess things up: if they reduce resolution (sampling/pooling/striding...) then information is lost, and it is difficult to utilize the information. Specifically, I suspect the fully-connected layer that might globally mess things up.
- A bug: the way you add the "killing feature" is not aligned with the image. Either the mask is added transposed, or you erroneously add the mask of one image to another (do you "shuffle" the training samples?)
An interesting experiment:
You can check if the net has at least a locally optimal weights that uses the "killing features": you can use net surgery to manually set the weights such that "conv_5" is zero for all features but the "killing features" and the weights for the consequent layers are not messing this up. Then you should have very high accuracy and low loss. Training the net from this point should yield very small (if any) gradients and the weights should not change significantly even after many iterations.
Related
tl;dr - I use an autoencoder to try to reduce input dimensions for a reinforcement-learning (RL) agent to learn how to play Atari-KungFu. But it fails at encoding/decoding thrown knives, because they are only a couple pixels and getting them wrong probably has negligible impact on the autoencoder MSE loss (see green arrows in bottom left of image). This will probably permanently hobble the results. I want to figure out if there is a way to solve this -- preferably with a generalized solution, but I'd be happy for now with something specific to this problem.
Background:
I am working on Week5 of the "Practical Reinforcement Learning" course on Coursera (National Research University HSE), and I decided to spend extra time trying to expand performance on the Atari-KungFu assignment using Actor-Critic architecture. This post is not about actor-critic, but more about an interesting sub-problem I ran into related to autoencoders.
I create an encoder which outputs a tanh-64-neuron layer, which is used as a common input to the decoder, policy learner (actor), and value learner (critic). During training, the simulator returns batches of four sequential frames (64 x 144 x 4) and rewards from the last action. Then images are first used to train the autoencoder, then used with the rewards to train the actor & critic branches.
I display some metrics and example frames every 25000 iterations to see how it's doing. If the reconstructed images are accurate, then the inputs to the actor & critic branches should be getting good distilled information for efficient learning.
You can see below that the autoencoder is pretty good except for the thrown knives (see bottom-left). Arguably this is because missing those couple pixels minimally increases the MSE loss of the reconstructed image, so it has little incentive to learn it (and also there's not a lot of frames that have knives). Yet, seeing those knives is critical for the RL agent to learn to how to survive.
I haven't seen this kind of problem addressed before. A tiny artifact in the input images is crucial for learning, but is unlikely to be learned by the autoencoder. Can we fix/improve this?
IMO your problem is loss specific, some things which would probably help autoencoder reconstruct knife as well:
Find knives in input image using image processing techniques. Regions where knives are present should have higher loss value in MSE, say 10 times more. One way to find those semi-automatically could probably be convolution with big kernel; White pixels at the strict center would give more weight and only zeros around it would give it more weight as well. Something along these lines should find a region where only knives are located (throwing guys wouldn't, as they contain too many white pixels and holes). Using some threshold found empirically for the value of this kernel should be enough to correctly find them.
Lower loss for images when no knive was found, say divided by half. This would focus autoencoder harder on rarely seen cases when knive is seen.
On the downside - I suppose it could introduce some artifacts. In such case you may think about usage of pretrained encoder (like some version of ResNet) and increase model's capabilities.
This is found in most implementations I've seen; I don't really understand the purpose? I've heard it's a preprocessing step that helps with classification accuracy? Is it necessary, particularly for non-classification tasks, eg. generating new images, working with image activations?
One of the most popular ways on how to normalize data is to make it have 0 mean and variance 1. It's usually done because:
Computational reasons - most training algorithms need your data points to have a small norm in order to run properly. It's because e.g. gradient stability, etc.
Dataset bias reason - if your data doesn't have a 0 mean - then it means that it constantly pushes network toward the certain direction. This must be compensated by network weights and biases what may slow down training (especially when the norm of outputs are relatively large).
When data is not normalized/scaled - some input coordinates (these ones with bigger means and norms) have a much greater impact on a training process. Imagine e.g. two variables - age and a binary indicator if someone had a heart attack. If you don't normalize your data - the fact that age has a higher norm than binary indicator will make this coordinate to influence training process much more than the other one. Is it plausible e.g. for predicting if someone will have another heart attack?
In this case i want to make letter recognition, the letter is scanned from a paper. the result of that process i have 5 x 5 binary matrix. so, it would use 25 input node. but i don't understand how to determine total hidden layer nodes and outputs node for that cases.i want to build the architecture of multilayer perecptron for that cases. thanks for your help!
Every NN has three types of layers: input, hidden, and output.
Creating the NN architecture therefore means coming up with values for the number of layers of each type and the number of nodes in each of these layers.
The Input Layer
Simple--every NN has exactly one of them--no exceptions that I'm aware of.
With respect to the number of neurons comprising this layer, this parameter is completely and uniquely determined once you know the shape of your training data. Specifically, the number of neurons comprising that layer is equal to the number of features (columns) in your data. Some NN configurations add one additional node for a bias term.
The Output Layer
Like the Input layer, every NN has exactly one output layer. Determining its size (number of neurons) is simple; it is completely determined by the chosen model configuration.
Is your NN going running in Machine Mode or Regression Mode (the ML convention of using a term that is also used in statistics but assigning a different meaning to it is very confusing). Machine mode: returns a class label (e.g., "Premium Account"/"Basic Account"). Regression Mode returns a value (e.g., price).
If the NN is a regressor, then the output layer has a single node.
If the NN is a classifier, then it also has a single node unless softmax is used
in which case the output layer has one node per class label in your model.
The Hidden Layers
So those few rules set the number of layers and size (neurons/layer) for both the input and output layers. That leaves the hidden layers.
How many hidden layers? Well if your data is linearly separable (which you often know by the time you begin coding a NN) then you don't need any hidden layers at all. Of course, you don't need an NN to resolve your data either, but it will still do the job.
Beyond that, as you probably know, there's a mountain of commentary on the question of hidden layer configuration in NNs (see the insanely thorough and insightful NN FAQ for an excellent summary of that commentary). One issue within this subject on which there is a consensus is the performance difference from adding additional hidden layers: the situations in which performance improves with a second (or third, etc.) hidden layer are very small. One hidden layer is sufficient for the large majority of problems.
So what about size of the hidden layer(s)--how many neurons? There are some empirically-derived rules-of-thumb, of these, the most commonly relied on is 'the optimal size of the hidden layer is usually between the size of the input and size of the output layers'. Jeff Heaton, author of Introduction to Neural Networks in Java offers a few more.
In sum, for most problems, one could probably get decent performance (even without a second optimization step) by setting the hidden layer configuration using just two rules: (i) number of hidden layers equals one; and (ii) the number of neurons in that layer is the mean of the neurons in the input and output layers.
Optimization of the Network Configuration
Pruning describes a set of techniques to trim network size (by nodes not layers) to improve computational performance and sometimes resolution performance. The gist of these techniques is removing nodes from the network during training by identifying those nodes which, if removed from the network, would not noticeably affect network performance (i.e., resolution of the data). (Even without using a formal pruning technique, you can get a rough idea of which nodes are not important by looking at your weight matrix after training; look weights very close to zero--it's the nodes on either end of those weights that are often removed during pruning.) Obviously, if you use a pruning algorithm during training then begin with a network configuration that is more likely to have excess (i.e., 'prunable') nodes--in other words, when deciding on a network architecture, err on the side of more neurons, if you add a pruning step.
Put another way, by applying a pruning algorithm to your network during training, you can approach optimal network configuration; whether you can do that in a single "up-front" (such as a genetic-algorithm-based algorithm) I don't know, though I do know that for now, this two-step optimization is more common.
Formula
One additional rule of thumb for supervised learning networks, the upperbound on the number of hidden neurons that won't result in over-fitting is:
Others recommend setting alpha to a value between 5 and 10, but I find a value of 2 will often work without overfitting. As explained by this excellent NN Design text, you want to limit the number of free parameters in your model (its degree or number of nonzero weights) to a small portion of the degrees of freedom in your data. The degrees of freedom in your data is the number samples * degrees of freedom (dimensions) in each sample or Ns∗(Ni+No) (assuming they're all independent). So alpha is a way to indicate how general you want your model to be, or how much you want to prevent overfitting.
For an automated procedure you'd start with an alpha of 2 (twice as many degrees of freedom in your training data as your model) and work your way up to 10 if the error for training data is significantly smaller than for the cross-validation data set.
References
Advameg (2016) Comp.Ai.Neural-nets FAQ, part 1 of 7: Introduction. Available at: http://www.faqs.org/faqs/ai-faq/neural-nets/part1/preamble.html
How to choose the number of hidden layers and nodes in a feedforward neural network? (2016a) Available at: https://stats.stackexchange.com/a/136542
How to choose the number of hidden layers and nodes in a feedforward neural network? (2016b) Available at: https://stats.stackexchange.com/a/1097
Legal, H.R. - and Info, C. (2016) Introduction to neural networks for java, 2nd edition. Available at: http://www.heatonresearch.com/book/programming-neural-networks-java-2.html
I am working on Soil Spectral Classification using neural networks and I have data from my Professor obtained from his lab which consists of spectral reflectance from wavelength 1200 nm to 2400 nm. He only has 270 samples.
I have been unable to train the network for accuracy more than 74% since the training data is very less (only 270 samples). I was concerned that my Matlab code is not correct, but when I used the Neural Net Toolbox in Matlab, I got the same results...nothing more than 75% accuracy.
When I talked to my Professor about it, he said that he does not have any more data, but asked me to do random perturbation on this data to obtain more data. I have research online about random perturbation of data, but have come up short.
Can someone point me in the right direction for performing random perturbation on 270 samples of data so that I can get more data?
Also, since by doing this, I will be constructing 'fake' data, I don't see how the neural network would be any better cos isn't the point of neural nets using actual real valid data to train the network?
Thanks,
Faisal.
I think trying to fabricate more data is a bad idea: you can't create anything with higher information content than you already have, unless you know the true distribution of the data to sample from. If you did, however, you'd be able to classify with the Bayes optimal error rate, which would be impossible to beat.
What I'd be looking at instead is whether you can alter the parameters of your neural net to improve performance. The thing that immediately springs to mind with small amounts of training data is your weight regulariser (are you even using regularised weights), which can be seen as a prior on the weights if you're that way inclined. I'd also look at altering the activation functions if you're using simple linear activations, and the number of hidden nodes in addition (with so few examples, I'd use very few, or even bypass the hidden layer entirely since it's hard to learn nonlinear interactions with limited data).
While I'd not normally recommend it, you should probably use cross-validation to set these hyper-parameters given the limited size, as you're going to get unhelpful insight from a 10-20% test set size. You might hold out 10-20% for final testing, however, so as to not bias the results in your favour.
First, some general advice:
Normalize each input and output variable to [0.0, 1.0]
When using a feedforward MLP, try to use 2 or more hidden layers
Make sure your number of neurons per hidden layer is big enough, so the network is able to tackle the complexity of your data
It should always be possible to get to 100% accuracy on a training set if the complexity of your model is sufficient. But be careful, 100% training set accuracy does not necessarily mean that your model does perform well on unseen data (generalization performance).
Random perturbation of your data can improve generalization performance, if the perturbation you are adding occurs in practice (or at least similar perturbation). This works because this means teaching your network on how the data could look different but still belong to the given labels.
In the case of image classification, you could rotate, scale, noise, etc. the input image (the output stays the same, naturally). You will need to figure out what kind of perturbation could apply to your data. For some problems this is difficult or does not yield any improvement, so you need to try it out. If this does not work, it does not necessarily mean your implementation or data are broken.
The easiest way to add random noise to your data would be to apply gaussian noise.
I suppose your measures have errors associated with them (a measure without errors has almost no meaning). For each measured value M+-DeltaM you can generate a new number with N(M,DeltaM), where n is the normal distribution.
This will add new points as experimental noise from previous ones, and will add help take into account exprimental errors in the measures for the classification. I'm not sure however if it's possible to know in advance how helpful this will be !
I am working on a project for uni which requires markerless relative pose estimation. To do this I take two images and match n features in certain locations of the picture. From these points I can find vectors between these points which, when included with distance, can be used to estimate the new postition of the camera.
The project is required to be deplyoable on mobile devices so the algorithm needs to be efficient. A thought I had to make it more efficient would be to take these vectors and put them into a Neural Network which could take the vectors and output an estimation of the xyz movement vector based on the input.
The question I have is if a NN could be appropriate for this situation if sufficiently trained? and, if so, how would I calculate the number of hidden units I would need and what the best activation function would be?
Using a neural network for your application can very well work, however, I feel you will need a lot of training samples to allow the network to generalize. Of course, this also depends on the type and number of poses you're dealing with. It sounds to me that with some clever maths it might be possible to derive the movement vector directly from the input vector -- if by any chance you can come up with a way of doing that (or provide more information so others can think about it too), that would very much be preferred, as in that case you would include prior knowledge you have about the task instead of relying on the NN to learn it from data.
If you decide to go ahead with the NN approach, keep the following in mind:
Divide your data into training and validation set. This allows you to make sure that the network doesn't overfit. You train using the training set and determine the quality of a particular network using the error on the validation set. The ratio of training/validation depends on the amount of data you have. A large validation set (e.g., 50% of your data) will allow more precise conclusions about the quality of the trained network, but often you have too few data to afford this. However, in any case I would suggest to use at least 10% of your data for validation.
As to the number of hidden units, a rule of thumb is to have at least 10 training examples for each free parameter, i.e., each weight. So assuming you have a 3-layer network with 4 inputs, 10 hidden units, and 3 output units, where each hidden unit and the output units have additionally a bias weight, you would have (4+1) * 10 + (10+1) * 3 = 83 free parameters/weights. In general you should experiment with the number of hidden units and also the number of hidden layers. From my experience 4-layer networks (i.e., 2 hidden layers) work better than 3-layer network, but that depends on the problem. Since you also have the validation set, you can find out what network architecture and size works without having to fear overfitting.
For the activation function you should use some sigmoid function to allow for non-linear behavior. I like the hyperbolic tangent for its symmetry, but from my experience you can just as well use the logistic function.