Sparse Coding
- Sparse coding is a class of unsupervised methods for learning sets of over-complete bases to represent data efficiently. The aim of sparse coding is to find a set of basis vectors such that an input vector can be represented as a linear combination of these basis vectors:
- The advantage of having an over-complete basis is that our basis vectors are better able to capture structures and patterns inherent in the input data .
- However, with an over-complete basis, the coefficients are no longer uniquely determined by the input vector. Therefore, in sparse coding, we introduce the additional criterion of sparsity to resolve the degeneracy introduced by over-completeness.
- The sparse coding cost function is defined on a set of m input vectors as:
- Although the most direct measure of sparsity is the norm, it is non-differentiable and difficult to optimize in general. In practice, common choices for the sparsity cost are the penalty and the log sparsity
- It is also possible to make the sparsity penalty arbitrary small by scaling down and scaling up by some large constant. To prevent this from happening, we will constrain to be less than some constant . The full sparse coding cost function hence is:
- One problem is that the constraint cannot be forced using simple gradient-based methods. Hence, in practice, this constraint is weakened to a "weight decay" term designed to keep the entries of small:
- Another problem is that the L1 norm is not differentiable at 0, and hence poses a problem for gradient-based methods. We will "smooth out" the L1 norm using an approximation which will allow us to use gradient descent. To "smooth out" the L1 norm, we use in place of , where is a "smoothing parameter" which can also be interpreted as a sort of "sparsity parameter" (to see this, observe that when is large compared to , the is dominated by , and taking the square root yield approximately .
- Hence, the final objective function is:
- The set of basis vectors are called "dictionary" (). is "adapted" to if it can represent it with a few basis vectors, that is, there exists a sparse vector in such that . We call the sparse code. It is illustrated as follows:
Learning
- Learning a set of basis vectors using sparse coding consists of performing two separate optimizations (i.e., alternative optimization method):
- The first being an optimization over coefficients for each training example
- The second being an optimization over basis vectors across many training examples at once.
- However, the classical optimization alternates between D and can achieve good results, but very slow.
- A significant limitation of sparse coding is that even after a set of basis vectors have been learnt, in order to "encode" a new data example, optimization must be performed to obtain the required coefficients. This significant "runtime" cost means that sparse coding is computationally expensive to implement even at test time, especially compared to typical feed-forward architectures.
Remarks
- From my view, due to the sparseness enforced in the dictionary learning (i.e., sparse code), the restored matrix is able to remove noise of original matrix, i.e., having some effect of denoising. Hence, Sparse coding could be used to denoise images.
References
- Sparse coding: http://ufldl.stanford.edu/wiki/index.php/Sparse_Coding
- Sparse coding: autoencoder interpretation: http://ufldl.stanford.edu/wiki/index.php/Sparse_Coding:_Autoencoder_Interpretation
- Sparse coding: exercise: http://ufldl.stanford.edu/wiki/index.php/Exercise:Sparse_Coding
- Sparse coding and dictionary learning for image analysis, ICCV 2010 tutorial.