The reasons to also use feature selection are (1) efficiency [stochastic gradient takes time proportional to number of non-zero features in the input], (2) accuracy [we've found doing info gain and then fitting has better 0/1 accuracy than relying on L_1 to do both {yes, I know that opens another can of worms on not optimizing the evaluation metric directly}], and (3) not all of our feature-based classifiers are regression based [we want to use this for perceptrons, for K-nearest neighbors, etc.].

Sometimes it is easy to overfit if we try too hard when discretizing the features before applying IG (for example, allowing too many bins). Another option that can work surprisingly well is to consider a single best-split-point approach, that is, to restrict ourselves to binary splits. One looks for the min-entropy split value for each feature by sorting the instances by that feature and measuring the entropy at the interesting points. Interesting points are those where there is a change of class, as it is obvious those are the only where the maximum IG can be found.

I briefly looked into the same problem a few years ago and tried out a number of different techniques. The most appealing to me were the entropy-based dynamic methods which essentially selected a number of bins and their thresholds based on a kind of minimum description length principle.

I can’t remember the exact method I used but a quick search using terms like “entropy” and “discretization” found the following papers which look vaguely familiar: Evaluating the performance of cost-based discretization versus entropy- and error-based discretization and Error-based and entropy-based discretization of continuous features. The categories of discretization in the conclusions of the latter paper are useful.

I also stumbled across this survey which may have some more pointers: Discretization techniques: a recent survey.

Hopefully some of that may help.