The authors propose a method to trade-off "computational costs" and "model fit" when learning a Sum-Product-Network (SPNs) represented as an Arithmetic Circuit. An SPN is a compact representation of a probabilistic model over discrete random variables with finite domain. The proposed method involves an SPN learner that is restricted to binary random variables. In practice, this requires to convert continuous variables into categoricals (e.g., using binning), and categoricals into binaries. While SPNs can handle missing data, they do are typically black-box models where the structure is learned. Computational costs are defined in terms of costs per arithmetic operations, memory/caching costs, as well as costs for feature computation. In particular, the authors assume that the feature costs might be defined over groups of features, e.g., generated by a single sensor. The made choices to define costs are plausible and rather straightforward. Model fit is measured in terms of identifying a distribution's mode, i.e., predictive performance of a maximum likelihood estimate. While the authors include datasets for density estimation in the experiments, they still focus on predicting one of the attributes rather than estimating the density of the whole data. While the proposed approach can be applied to a large and important class of problems, the title and introduction slightly overstate its applicability. The method does not address a general probabilistic model, e.g., stated as a Markov network or a probabilistic program. The significance of the work is moderate. The main contribution, besides defining hardware-related costs, is a heuristic to identify a set of SPNs at the Pareto frontier of costs vs. performance. The heuristic performs a backwards feature selection, where features are greedily removed such that the ratio of accuracy and costs is maximized. The removal is performed by pruning a previously trained SPN using an existing SPN learning algorithm. While empirical experiments confirm the effectiveness of the proposed search heuristic, the method is similar to backward feature selection sequentially selecting features that least affect the performance-cost ratio. The overall scientific contribution to the field of Machine Learning is minor. The pseudo code in Algorithm 1 is, in parts, incorrect: a_{ca,j}, acc_{ca,j}, cost_{ca,j} should be outside of the S-loop; \alpha_{select} etc. should be outside of the j-loop. Also using \notin for two sets, F_{ca,j} and F_S, is imprecise; this should be F_{ca,j} \cap F_S = \varnothing. Finally, \argmax_{F \notin F_{ca}} is ambiguous as there might be many F that are not element of F_{ca} (which is, strictly speaking, a list of sets). I assume, this should read j^* = \argmax_j CF(acc_{ca,j}, cost_{ca,j}), \alpha_select = \alpha_{ca,j^*}, and F_{rm} = \F_{ca,j^*}. Overall, I find the pseudo code not very helpful in its current form. Figure is slightly too small.

-- I enjoyed reading this paper, it is well-written, and tackles an important problem. -- My main concern is around the overheads added by the proposed algorithms. First of all, it wasn't clear from reading the paper if these algorithms need to be run for each inference query. Please clarify. If so, even though the complexity analysis is given for different components, it will help if this was empirically shown too. For instance, for each of the datasets, and inference tasks, what was the breakdown of the time spent in choosing a configuration, and actually serving a prediction for these tasks? -- The key motivation of the proposed work is for the edge computing use cases, which are as noted in the paper, latency and privacy sensitive. Thus, the overhead analysis becomes even more important. -- The paper has mainly used accuracy as the performance metric, and mentioned that it could be any other application-specific performance metric. However, it isn't clear how. Applications may have their own metrics, and they could vary over time as the paper points out too. Hence, it would definitely strengthen the practical aspects of the approach to show this aspect more clearly. -- nits: --- What is column 2 in Table 2? The heading says "64-bits". Is this the actual cost computed for 64-bits for each of the operations? Please consider renaming the header for this column if so. --- Fig. 3 is microscopic :). Please consider devoting more space. === UPDATE After Reading The Author Response === I would like to thank the authors for a great response, it clarified and satisfied my questions about the overheads and performance metrics. I am happy to maintain my current score of 7.

# Summary The paper proposes a resource-aware cost metric that takes into consideration the target embedded device's properties and system-level configuration, thus introducing probabilistic reasoning on edge computing. # Quality The proposed approach has been evaluated on different datasets for classification tasks. The paper should stress that the methods works only for classification tasks and the generality of the AC as a density estimator is lost. Indeed, since the methods relies on a greedy feature selection approach, queries on removed variables cannot be answered. # Clarity The authors firstly provide how to compute the computation cost of an AC in terms of arithmetic operations and fetching parameters from off-chip memory and storing and fetching from local cache. They the introduce the sensor interfacing cost. A greedy algorithm for learning the lowest-cost/highest-accuracy AC has been proposed that iteratively tries to remove a feature taking into account the accuracy of the new obtained AC. It is not clear the relationship between the computation cost C_AC and the cost of extracting features C_F. It seems they are strongly correlated. For instance, the gain in C_AC with respect to the each removed feature could be interesting, at least empirically. # Originality The work proposes a new approach on learning lowest-cost AC classifiers with high accuracy. The related section should be extended. See for instance [1] and [2] that are cited in the paper. Differences wrt these works should be indicated. [1] Integer undirected graphical models for resource-constrained systems, 2016 [2] Automatic mapping of the sum-product network inference problem to fpga-based accelerators, 2018 # Significance The experimental evaluation of the proposed approach prove its validity. However, there is no comparison with respect to other approaches. For instance, at least a base comparison to the accuracy obtained with the tree-augmented BN used for feature selection should be reported in the table. Minors line 32: remove the ";" line 113: represents --> represent line 115: defines --> define line 120: benchmarks,customized --> benchmarks, customized