8
$\begingroup$

Maxout networks were a simple yet brilliant idea of Goodfellow et al. from 2013 to max feature maps to get a universal approximator of convex activations. The design was tailored for use in conjunction with dropout (then recently introduced) and resulted of course in state-of-the-art results on benchmarks like CIFAR-10 and SVHN.

Five years later, dropout is definitely still in the game, but what about maxout? The paper is still widely cited in recent papers according to Google Scholar, but it seems barely any are actually using the technique.

So is maxout a thing of the past, and if so, why — what made it a top performer in 2013 but not in 2018?

$\endgroup$
0

2 Answers 2

3
$\begingroup$

Basically, if you read the full paper (especially, the abstract and the section 7), you find that the main accomplishment remains a marginal contribution on top of dropout.

If you see the empirical results on Table 5 (of the page 5) of the maxout's original paper, you find that the misclassification rate is only very, very slightly lower than that of dropouts. (2.47 % instead of 2.78%)

That could explain the relatively lower interest in the work.

$\endgroup$
2
  • $\begingroup$ Thanks for your answer. Cutting the error from 2.78% down to 2.47% is actually a 11% decrease in the error rate. Not a game changer but still interesting and definitely not tiny. $\endgroup$
    – user209974
    Commented Mar 18, 2019 at 10:47
  • $\begingroup$ Fair point. If that is of interest to your specific application, it certainly is worth investigating. $\endgroup$ Commented Mar 18, 2019 at 15:03
0
$\begingroup$

A few years later, stumbling on my own question, I feel I could share my view on that point.

I think it boils down eventually to a question of efficiency. There is nothing wrong per se with maxout; it is a distant cousin of ReLU that is in theory more capable. But the crux of the matter is that to produce a single scalar output, maxout consumes $k$ neurons, $k\geq 2$. The minimum price to pay to replace a standard activation function like ReLU with maxout while keeping the latent dimension is doubling the width of the scalar product. It is rather steep.

Though, a nice property of maxout is that it tends to keep the signal very much alive, instead of killing it like ReLU does. That property might prove to be important enough for maxout to flourish in some niche market despite its inherent wastefulness.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .