Suppose you have a problem that you like to tackle with machine learning and use the resulting system in a real-life project. I like to share my simple pathway for such purpose, in order to provide a basic guide to beginners and keep these things as a reminder to myself. These rules are tricky since even-thought they are simple, it is not that trivial to remember all and suppress your instinct which likes to see a running model as soon as possible.
When we confronted any problem, initially we have numerous learning algorithms, many bytes or gigabytes of data and already established knowledge to apply some of these models to particular problems. With all these in mind, we follow a three stages procedure;
Selfies are everywhere. With different fun masks, poses and filters, it goes crazy. When we coincide with any of these selfies, we automatically give an intuitive score regarding the quality and beauty of the selfie. However, it is not really possible to describe what makes a beautiful selfie. There are some obvious attributes but they are not fully prescribed.
With the folks at 8bit.ai, we decided to develop a system which analyzes selfie images and scores them in accordance to its quality and beauty. The idea was to see whether it is possible to mimic that bizarre perceptual understanding of human with the recent advancements of AI. And if it is, then let's make a mobile application and let people use it for whatever purpose. Spoiler alert! We already developed Selfai app available on iOS and Android and we have one instagram bot @selfai_robot. You can check before reading.
This paper states the following phrase. Traditional machine learning frameworks (VC dimensions, Rademacher complexity etc.) trying to explain how learning occurs are not very explanatory for the success of deep learning models and we need more understanding looking from different perspectives.
They rely on following empirical observations;
Deep networks are able to learn any kind of train data even with white noise instances with random labels. It entails that neural networks have very good brute-force memorization capacity.
Explicit regularization techniques - dropout, weight decay, batch norm - improves model generalization but it does not mean that same network give poor generalization performance without any of these. For instance, an inception network trained without ant explicit technique has 80.38% top-5 rate where as the same network achieved 83.6% on ImageNet challange with explicit techniques.
A 2 layers network with 2n+d parameters can learn the function f with n samples in d dimensions. They provide a proof of this statement on appendix section. From the empirical stand-view, they show the network performances on MNIST and CIFAR-10 datasets with 2 layers Multi Layer Perceptron.
Above observations entails following questions and conflicts;
Traditional notion of learning suggests stronger regularization as we use more powerful models. However, large enough network model is able to memorize any kind of data even if this data is just a random noise. Also, without any further explicit regularization techniques these models are able to generalize well in natural datasets. It shows us that, conflicting to general belief, brute-force memorization is still a good learning method yielding reasonable generalization performance in test time.
Classical approaches are poorly suited to explain the success of neural networks and more investigation is imperative in order to understand what is really going on from theoretical view.
Generalization power of the networks are not really defined by the explicit techniques, instead implicit factors like learning method or the model architecture seems more effective.
Explanation of generalization is need to be redefined in order to solve the conflicts depicted above.
My take : These large models are able to learn any function (and large does not mean deep anymore) and if there is any kind of information match between the training data and the test data, they are able to generalize well as well. Maybe it might be an explanation to think this models as an ensemble of many millions of smaller models on which is controlled by the zeroing effect of activation functions. Thus, it is able to memorize any function due to its size and implicated capacity but it still generalize well due-to this ensembling effect.
A crucial problem in a real DL system design is to capture test data distribution with the trained model which only sees the training data distribution. Therefore, it is always important to find a good data splitting scheme which at least gives the right measures to such divergence.
It is always a waste to spend all your time for fine-tunning your model on the measure of validation data taken from training data only. Because, when you deploy the model, it undergoes new instances sampled from dynamically shifting data distribution. If you have a chance to see some samples from this dynamic environment, use that to test your model on these real instances and keep your model more coherent and don't mislead your training flow.
That being said, on the above figure, the second row depicts the right way to choose your data split. And the third row shows the smoothed version which is suggested in practice.
Above figure shows common machine learning problems in relation to different components of your work flow. It is really important to understand what is really said here and what these problems explain.
Bias is the quality of your model on training data. If it predicts wrong on training, it has a "Bias" problem. If you have a good performance on training data but not on validation data, it yields "Variance" problem. If performance differs for validation data taken from training set and test set, it is "Train - Test mismatch". If performance suffers due to distribution shift on test time, it is "Overfitting".
Bias requires better architecture and longer training. Variance needs more data and regularization. Train - Test mismatch needs more training data from distribution similar to your test data. Overfitting needs regularization, more data, and data synthesis effort.
Above chart shows a salient way of conducting DL system evolution. Follow these decisions with empirical evidences and don't skip any of these in order not to be disappointed in the end. (I said it with many disappointments 🙂 )
When we see that train, validation errors are close enough to human level performance, it means more variance problem and we need to collect more data similar to test portion and hurdle more data synthesis work. Train and validation errors far from human level performance is the sign of bias problem, requires larger models and more training time. Keep in mind that, human performance is not the limit of what your model is theoretically capable of.
Disclaimer: Figures are taken from https://kevinzakka.github.io/2016/09/26/applying-deep-learning/ which summarizes Andrew Ng's talk.
A successful AI agent should communicate. It is all about language. It should understand and explain itself in words in order to communicate us. All of these spark with the "meaning" of words which the atomic part of human-wise communication. This is one of the fundamental problems of Natural Language Processing (NLP).
"meaning" is described as "the idea that is represented by a word, phrase, etc. How about representing the meaning of a word in a computer. The first attempt is to use some kind of hardly curated taxonomies such as WordNet. However such hand made structures not flexible enough, need human labor to elaborate and do not have semantic relations between words other then the carved rules. It is not what we expect from a real AI agent.
Then NLP research focused to use number vectors to symbolize words. The first use is to donate words with discrete (one-hot) representations. That is, if we assume a vocabulary with 1K words then we create a 1K length 0 vector with only one 1 representing the target word. Continue reading Why do we need better word representations ?→
As we witness the golden age of AI underpinned by deep learning, there are many different tools and frameworks continuously proposed. Sometimes it is even hard to catch up what is going on. You choose one over another then you see a new library and you go for it. However, it seems the exact choice is oblivious to anyone.
According to me, libraries are measured by flexibility and run-time trade-off. If you go with a library which is really easy to use, it is slow as much as that. If the library is fast, then it does not serve that much flexibility or it is so specialized to a particular type of models like Convolutional NNs.
After all the tears and blood dropped through years of experience in deep learning, I decide to share my own intuition and opinion about the common deep learning libraries so that these might help you to choose the right one for your own sake .
This paper is an interesting work which tries to explain similarities and differences between representation learned by different networks in the same architecture.
To the extend of their experiments, they train 4 different AlexNet and compare the units of these networks by correlation and mutual information analysis.
They asks following question;
Can we find one to one matching of units between network , showing that these units are sensitive to similar or the same commonalities on the image?
Is the one to one matching stays the same by different similarity measures? They first use correlation then mutual information to confirm the findings.
Is a representation learned by a network is a rotated version of the other, to the extend that one to one matching is not possible between networks?
Is clustering plausible for grouping units in different networks?
Answers to these questions are as follows;
It is possible to find good matching units with really high correlation values but there are some units learning unique representation that are not replicated by the others. The degree of representational divergence between networks goes higher with the number of layers. Hence, we see large correlations by conv1 layers and it the value decreases toward conv5 and it is minimum by conv4 layer.
They first analyze layers by the correlation values among units. Then they measure the overlap with the mutual information and the results are confirming each other..
To see the differences between learned representation, they use a very smart trick. They approximate representations learned by a layer of a network by the another network using the same layer. A sparse approximation is performed using LASSO. The result indicating that some units are approximated well with 1 or 2 units of the other network but remaining set of units require almost 4 counterpart units for good approximation. It shows that some units having good one to one matching has local codes learned and other units have slight distributed codes approximated by multiple counterpart units.
They also run a hierarchical clustering in order to group similar units successfully.
For details please refer to the paper.
My discussion: We see that different networks learn similar representations with some level of accompanying uniqueness. It is intriguing to see that, after this paper, these are the unique representations causing performance differences between networks and whether the effect is improving or worsening. Additionally, maybe we might combine these differences at the end to improve network performances by some set of smart tricks.
One deficit of the paper is that they do not experiment deep networks which are the real deal of the time. As we see from the results, as the layers go deeper, different abstractions exhumed by different networks. I believe this is more harsh by deeper architectures such as Inception or VGG kind.
One another curious thing is to study Residual netwrosk. The intuition of Residual networks to pass the already learned representation to upper layers and adding more to residual channel if something useful learned by the next layer. That idea shows some promise that two residual networks might be more similar compared to two Inception networks. Moreover, we can compare different layers inside a single Residual Network to see at what level the representation stays the same.
This work proposes yet another way to initialize your network, namely LUV (Layer-sequential Unit-variance) targeting especially deep networks. The idea relies on lately served Orthogonal initialization and fine-tuning the weights by the data to have variance of 1 for each layer output.
The scheme follows three stages;
Initialize weights by unit variance Gaussian
Find components of these weights using SVD
Replace the weights with these components
By using minibatches of data, try to rescale weights to have variance of 1 for each layer. This iterative procedure is described as below pseudo code.
In order to describe the code in words, for each iteration we give a new mini-batch and compute the output variance. We compare the computed variance by the threshold we defined as to the target variance 1. If number of iterations is below the maximum number iterations or the difference is above we rescale the layer weights by the squared variance of the minibatch. After initializing this layer go on to the next layer.
In essence, what this method does. First, we start with a normal Gaussian initialization which we know that it is not enough for deep networks. Orthogonalization stage, decorrelates the weights so that each unit of the layer starts to learn from particularly different point in the space. At the final stage, LUV iterations rescale the weights and keep the back and forth propagated signals close to a useful variance against vanishing or exploding gradient problem , similar to Batch Normalization but without computational load. Nevertheless, as also they points, LUV is not interchangeable with BN for especially large datasets like ImageNet. Still, I'd like to see a comparison with LUV vs BN but it is not done or not written to paper (Edit by the Author: Figure 3 on the paper has CIFAR comparison of BN and LUV and ImageNet results are posted on https://github.com/ducha-aiki/caffenet-benchmark).
The good side of this method is it works, for at least for my experiments made on ImageNet with different architectures. It is also not too much hurdle to code, if you already have Orthogonal initialization on the hand. Even, if you don't have it, you can start with a Gaussian initialization scheme and skip Orthogonalization stage and directly use LUV iterations. It still works with slight decrease of performance.