In this answer to the question Is an optimization algorithm equivalent to a neural network?, the author stated that, in theory, there is some recurrent neural network that implements a given optimization algorithm.

If so, then can we optimize the optimization algorithm (neural network)?


First, you need to consider what are the "parameters" of this "optimization algorithm" that you want to "optimize". Let's take the most simple case, a SGD without momentum. The update rule for this optimizer is:

$$ w_{t+1} \leftarrow w_{t} - a \cdot \nabla_{w_{t}} J(w_t) = w_{t} - a \cdot g_t $$

where $w_t$ are the weights at iteration $t$, $J$ is the cost function, $g_t = \nabla_{w_{t}} J(w_t)$ are the gradients of the cost function w.r.t $w_t$ and $a$ is the learning rate.

An optimization algorithm accepts as its input the weights and their gradients and returns the update. So we could write the above equation as:

$$ w_{t+1} \leftarrow w_{t} - SGD(w_t, g_t) $$

The same is true for all optimization algorithms (e.g. Adam, RMSprop, etc.). Now our initial question was what are the parameters of the optimizer, which we want to optimize. In the simple case of the SGD, the sole parameter of the optimizer is the learning rate.

The question that arises at this point is can we optimize the learning rate of the optimizer during training? Or more practically, can we compute this derivative?

$$ \frac{\partial J(w_t)}{\partial a} $$

This idea was explored in this paper, where they coin this technique "hypergradient descent". I suggest you take a look.

  • $\begingroup$ In particular, one would expect an 'optimised algorithm' not to need any external parameters to be chosen. $\endgroup$ – Andre Holzner Jul 28 at 12:13

We usually optimize with respect to something. For example, you can train a neural network to locate cats in an image. This operation of locating cats in an image can be thought of as a function: given an image, a neural network can be trained to return the position of the cat in the image. In this sense, we can optimize a neural network with respect to this task.

However, if a neural network represents an optimization algorithm, then, if you change it a little bit, then it will no more be the same optimization algorithm: it might be another optimization algorithm or some other different algorithm.

For example, most optimizations algorithms that are used to train neural networks (like Adam) are a variation of gradient descent (GD). If you think that Adam performs better than GD, then you could say that Adam is an optimization of GD. So, Adam performs better than GD with respect to something. Possibly, GD also performs better than Adam with respect to something else. Of course, this is a little bit of a stretch.

  • $\begingroup$ i think when referring to optimization they are talking about one that adapts/learns, not just finding a better algorithm-- so i think its worth mentioning how there optimizers like adam are optimized to the model (they have parameters learned to help traverse the loss landscape in a more efficient mannar) $\endgroup$ – mshlis Jul 24 at 12:40
  • $\begingroup$ @mshlis I didn't want to go into the details of Adam or any other optimization algorithm. I just want to convey the idea that Adam is better than another optimization algorithm with respect to something. $\endgroup$ – nbro Jul 24 at 14:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.