NIPS 2018
Sun Dec 2nd through Sat the 8th, 2018 at Palais des Congrès de Montréal
Paper ID: 4941 Submodular Maximization via Gradient Ascent: The Case of Deep Submodular Functions

### Reviewer 1

The paper proves a very interesting result: For maximizing Deep Submodular Functions (DSF) with matroid constraints, one can provide efficient algorithms that, under mild assumptions on the singleton marginals, have approximation factor better than 1-1/e (and potentially approaching 1 when the rank of the matroid becomes large). This is given in Theorem 1 which I think is the main result of the paper. The basic idea behind the algorithm is that for DSFs there is a natural concave extension, equations (4), (5) that can be maximized by projected gradient ascent (this result has been proved in [3]). The authors show in Proposition 1 that this concave extension is close to the multilinear extension, and in section 4 they show that the projected gradient ascent algorithm can be implemented efficiently (e.g. the subgradient of the concave extension can be computed efficiently due to the deep structure). The paper is written well and the results are novel. However, I have a few comments for which I would appreciate a response from the authors: 1) How is the projection step in Algorithm 1 done? E.g. what is the complexity for general matroids? Since you prefer k being large, it would be useful to provide the exact runtime of the algorithm in terms of n and k. Also, in Figure 3 (c), is the projection step considered? (I guess for partition matroid the projection step is easier than general matroids) 2) Can the authors explain how the bounds of Theorem 1 compare with 1-1/e if k is not so large? (lets assume that the modular term m_{+-} in equation (1) is zero to make the overall objective monotone). Also, it would be beneficial if the authors could explain how the bound in Theorem 1 would be in the worst case. I.e. given a fixed k and perhaps negative modular terms, how bad the bounds of Theorem 1 could be? In particular, are they better than 1/2? 3) In your experiments, I think it would be useful in include settings with negative modular terms (m_{+-} in equation (1)). Some minor comments: - Line 40: befit -> benefit - Regarding the discussion in lines 74-78, there is now an efficient and practical algorithm for maximizing the multilinear extension over matroids (see the stochastic continuous greedy algorithm by Mokhtari et al, 2017). Basically, you don't need to compute the multilinear extension at each step (unlike [6]) and only an unbiased estimator (that can be computed very efficiently) is enough to maximize the function using stochastic optimization algorithms.

### Reviewer 2

Summary: Deep submodular functions are a class of submodular functions that may be modeled using a neural network. This work focuses on maximising deep submodular functions with matroid constraints. This is done using multilinear extension that can be optimized using gradient asnet. Further, pipage rounding is used to find an approximate optimal solution to the original problem. Pros: - It is first novel attempt to maximize deep submodular funcions under matroid constraints. - The paper is well structured and well written - Theorem 1 where they prove that multilinear extensive of DSF need not be calculated to optimize the function is the major result - Theorem 3 gives an approximation due to pipage rounding negatives/clarifications: One of the main drawbacks of the paper is experimental evaluation. One of the issue with deep neural networks is the amount it takes to optimize the function in the parameter. I believe, the goal of deep submodular functions is to leverage the underlying function structure to be able to speed up the optimize algorithms yet as express as DNNs. However, this paper does not demosnstrate this. Update after authors response: I do agree with the authors that NIPS should have space for theoretical contribution and that is precisely the reason why I tend to vote towards accepting and upgrading to 7. However, I believe, a good experimental evaluation would have further strengthened the paper.

### Reviewer 3

This paper studies gradient-based methods for optimizing a deep submodular function. The starting point is a natural concave relaxation where each variable is allowed to take fractional values. The main contribution is a bound on the gap between the concave relaxation and the multilinear extension, generalizing a bound for coverage functions by Karimi et al. For large k, and depending on the parameters of the DSF, the resulting approximation ratio can improve on 1 - 1/e. Providing improved tools for maximizing DSFs is a natural problem, and the ability to improve on the approximation ratio and scalability of the greedy algorithm is an attractive property. However, the improvements (both theoretical and empirical) appear somewhat marginal in many cases. Additional information on a few points would be helpful to assess the significance of the work: First, the bounds depend on parameters of the DSF like |V_1| and w_min/max. It is claimed that these quantities are often favorable in practice. However, none of the papers cited use DSFs. What evidence is there that real problem instances are amenable to such techniques? Second, what is the value of n (the size of the ground set) for the experiments? This is important context for understanding the scalability comparison vs greedy. Third, the comparison vs greedy uses a partition matroid constraint. Is there any improvement in solution quality (compared to greedy) for a uniform matroid? This is the more common case in practice, and would help disentangle two factors that may be driving the performance improvement: (a) the concave relaxation becoming tighter with large k, and (b) continuous methods in general simply being more appropriate for general matroids. Fourth, the motivation for using the concave relaxation is that the multilinear extension might require many samples to approximate. However, this is often pessimistic because the multilinear extensions of common families such as converage functions or facility location can be computed in closed form. Do similar techniques extend to the compositions used in DSFs? Update after author response: The response helps address some of my concerns and I have raised my score by one point. The experimental evaluation is still weak overall though; this paper introduces some nice ideas and it would be good to see them tested more thoroughly.