I'm interested in applying several different types of regularization to neural nets and want to make sure I haven't learned the material incorrectly. I have successfully coded Weight Decay and Dropout, among others; in the case of Dropout it was a simple matter of setting a flag on a randomly selected set of connections; with Weight Decay it was a simple matter of multiplying the connection weights by a regularization term (typically slightly less than 1) after each learning step. I think applying L1, L2 and Tikhonov Regularization may turn out to be just as easy, once I clear up a few potential misunderstandings and lingering doubts I have after wading through some of the literature:
-
Can L1, L2 and Tikhonov calculations be appended as extra terms directly to a cost function, such as MSE? As I said above, it was easy for me to apply Weight Decay simply by applying a regularization multiplier to a whole set of weights; it would likewise be easy for me to code L1, L2 and Tikhonov by applying a new term to the existing cost functions I've already written. From the notation used in Martin Toma's reply to this post on the Data Science Forum, it appears that I can simply add a weighted sum of the absolute value of each weight (in the case of L1) or a weighted sum of each squared weight to the cost function. Nevertheless, the topic is actually logistic regression, which is only tangentially related to neural nets, plus the added terms are not explicitly depicted together with the MSE term.
-
If #1 is correct, is is possible to apply them in supervised learning situations where there are no cost functions?
-
I've gathered that L2 is a special case of Tikhonov Regularization, after reading a few research papers like Training with Noise is Equivalent to Tikhonov Regularization and Tikhonov Training of the CMAC Neural Network and of course skimming various Crossvalidated and Data Science Forum threads and the Wikipedia pages on Regularization and Tikhonov Regularization. On occasion though I've seen sources refer to them as if they were equivalent, so please correct me if L2 is a synonym for Tikhonov and not just a special case.
-
L2 is differentiable and therefore compatible with gradient descent, which is probably why it is the most popular type of Tikhonov Regularization. From the Wikipedia page, however, it appears that we can substitute other operations like difference and Fourier operators to derive other brands of Tikhonov Regularization that are not equivalent to L2. Are any of these alternate Tikhonov types applicable to neural nets? Are any of them actually used often in practice? I suspect that some of them would not be differentiable or suitable for gradient descent and therefore would be impractical.
-
If #4 is true, can someone post an example of a Tikhonov formula that is not equivalent to L2, yet would be useful in neural nets? I have a harder time following matrix notation than with the kind of standard summation operator notation used here or common neural notation used in sources like Patrick K. Simpson's old reference "Artificial Neural Systems: Foundations, Paradigms, Applications, and Implementations." Most of the explanations of Tikhonov I've seen to date, however, use matrix notation, which has made it more difficult for me to clear up my confusion.
-
There are arcane references in the literature to a "Tikhonov Matrix," which I can't seem to find any definitions of (although I've run across several unanswered questions scattered across the Internet about its meaning). An explanation of its meaning would be helpful.
I may be better off moving #6 to the Mathematics forum if it doesn't have a simple answer, but the other questions are all closely interrelated so I don't want to break them up into separate threads, if possible. Most of them may have simple yes-or-no answers, except for the request for a non-L2 version of Tikhonov Regularization. I don't necessarily need a top-to-bottom explanation of these three types of regularization or their use cases, since there are already some good threads and other resources for that; I just want to nip some possible misconceptions in the bud before I try to implement them in code. Thanks in advance.
Best Answer
If you are interested in pursuing neural networks (or other machine learning), you would be well served to spend some time on the mathematical fundamentals. I started skimming this trendy new treatise, which IMO is not too difficult mathematically (although I have seen some complain otherwise, and I was a math major, so YMMV!).
That said, I will try to address your questions as best I can at a high level.
Yes, these are all penalty methods which add terms to the cost function.
I am not sure what you are thinking here, but in my mind the supervised case is the most clear in terms of cost functions and regularization. In general you can think of the error conceptually as
$$\text{Total Error}=\left(\text{ Data Error }\right) + \left(\text{ Prior Error }\right)$$
where the first term penalizes the misfit of the model predictions vs. the training data, while the second term penalizes overfitting of the model (with the goal of lowering generalization errors).
For a problem where features $x$ are used to predict data $y$ via a function $f(x,w)$ parameterized by weights $w$, the above cost function will typically be realized mathematically by something like
$$E[w]=\|\,f(x,w)-y\,\|_p^p + \|\Lambda w\|_q^q$$
Here the terms have the same interpretation as above, with the errors measured by "$L_p$ norms": The data misfit usually has $p=2$, which corresponds to MSE, while the regularization term may use $q=2$ or $q=1$. (The $\Lambda$ I will get to below.)
The latter corresponds to absolute(-value) errors rather than square errors, and is commonly used to promote sparsity in the solution vector $w$ (i.e. many 0 weights, effectively limiting the connectivity paths). The $L_1$ norm can be used for the data misfit term also, typically to reduce sensitivity to outliers in the data $y$. (Conceptually, predictions will target the median of $y$ rather than the mean of $y$.)
Note that in many unsupervised learning scenarios there are effectively two "machines", with each taking turns as "data" vs. "prediction" (e.g. the coder and decoder parts of an autoencoder).
The two can be used as synonyms. Some communities tend to use "$L_2$" to refer to the special case where the Tikhonov matrix $\Lambda$ is simply a scalar $\lambda$. Terminology varies quite a bit.
This is not correct. Tikhonov regularization always uses the $L_2$ norm, so is always a differentiable $L_2$ regularization. The matrix $\Lambda$ does not impact this (it is constant, it does not matter if it is a scalar, a diagonal covariance, a finite difference operator, a Fourier transform, etc.).
For differentiability, the comparison is typically between $L_2$ and $L_1$. You can think of it like $$(x^2)'=2x \quad \text{ vs. } \quad |x|'=\mathrm{sgn}(x)$$ i.e. the $L_2$ norm has a continuous derivative while the $L_1$ norm has a discontinuous derivative.
This difference is not as big as you might imagine in terms of optimize-ability. For example the popular ReLU activation function also has a discontinuous derivative, i.e. $$\max(0,x)'=(x>0)$$ More importantly, the $L_1$ regularizer is convex, so helps suppres local minima in the composite cost function. (Note I am avoiding technical issues around convex optimization & differentiability, such as "subgradients", as these are not essential to the point.)
This was answered above: No, because all Tikhonov regularizers use the $L_2$ norm. (Wether they are useful for NN or not!)
This is simply the matrix $\Lambda$. There are many forms that it can take, depending on the goals of the regularization:
In the simplest case it is simply a constant ($\lambda>0$). This penalizes large weights (i.e. promotes $w_i^2\approx 0$ on average), which can be useful to prevent over-fitting.
In the next simplest case, $\Lambda$ is diagonal, which allows per-weight regularization (i.e. $\lambda_iw_i^2\approx 0$). For example the regularization might vary with level in a deep network.
Many other forms are possible, so I will end with one example of a sparse but non-diagonal $\Lambda$ that is common: A finite difference operator. This only makes sense when the weights $w$ are arranged in some definite spatial pattern. In this case $\Lambda w\approx \nabla w$, so the regularization promotes smoothness (i.e. $\nabla w\approx 0$). This is common in image processing (e.g. tomography), but could conceivably be applied in some types of neural-network architectures as well (e.g. ConvNets).
In the last case, note that this type of "Tikhonov" matrix $\Lambda$ can really be used in $L_1$ regularization as well. A great example I did once was with the smoothing example in an image processing context: I had a least squares ($L_2$) cost function set up to do denoising, and by literally just putting the same matrices into a "robust least squares" solver* I got an edge-preserving smoother "for free"! (*Essentially this just changed the $p$'s and $q$'s in my first equation from 2 to 1.)
I hope this answer has helped you to understand regularization better! I would be happy to expand on any part that is still not clear.