NeurIPS 2020

### Review 1

Summary and Contributions: This paper proposes a method that co-designs TinyNAS with TinyEngine, which deploys the deep neural networks on the commercial microcontroller STM32 and achieves promising results on ImageNet. The proposed MCUNet is the first one to achieve >70% ImageNet accuracy on the off-the-shelf commercial microcontroller. This is a very significative direction for research of efficient deep learning and AutoML.

Strengths: 1.Good performance. 2. The method is easy to follow and makes sense. 3. This paper proposes a two-stage NAS, TinyNAS, to first optimize the search space and then perform NAS to meet the constraints, which searches for architectures under a wide space with low cost. 4. The proposed TinyEngine introduces model-adaptive memory scheduling and computation kernel specialization which greatly improves inference efficiency and enlarges the design space in turn. TinyEngine is used for co-design with TinyNAS. As a consequence, both the architecture and accelerator are optimized. Moreover, the memory scheduling is adapted according to the overall network, not only the layer-wise optimization, which is benefit for the memory usage reduction. 5. The final deployed model achieves great performance on ImageNet and Visual&Audio Wake Words. Especially, MCUNet achieves 70.2% top-1 accuracy on ImageNet on the STM32 MCU. The result is very promising as the capacity of STM32 is very small.

Weaknesses: 1. The co-designing mechanism is the core contribution of the paper but the detailed process is not described clearly. It is suggested to provide an elaborate diagram or pseudocode to introduce the whole framework. 2. MCUNet achieves convincing results on STM32. It is meaningful to explore its generalization ability. Please demonstrate the potential/ability of MCUNet being deployed on more scenarios and tasks, e.g., more devices or tasks like object detection, semantic segmentation. 3.The method is described as a co-design scheme. But the experiments failed to highlight the improvements of the co-design scheme, when compared to those single design schemes. 4.The author introduce the tiny engine with the motivation that "TinyEngine optimizes the memory scheduling according to the overall network topology to get a better strategy." But it is unclear whether the overall network topology indeed is the main reason for the huge improvements. The authors need to add supporting results or supporting articles. 5. Some typo issues need to be checked, e.g., "to achieves" in line 14.

Correctness: The claims and method are correct. The experiments are performed comprehensively and the empirical methodology are correct.

Clarity: The co-designing details are not so clear. The other parts of the paper are described clearly and are easy to follow.

Relation to Prior Work: The relation to prior work is discussed clearly. Though many previous works attempt to optimize neural architectures on customized hardware platform. Rare of them co-design the architecture with the accelerator. This brings more probabilities for TinyML being applied in wide fields.

Reproducibility: Yes

Additional Feedback: After reading the rebuttal and comments from other reviewers, I would like to keep my original rating. Regarding some details that need to be fully illustrated in this paper. The proposed method is meaningful that may have a great impact on the TinyAI .

### Review 2

Summary and Contributions: This paper proposes a joint design of neural architecture search and inference engine for MCUs. The NAS part scales the input resolution and width multiplier to fit the model to the limited resource. A novel CDF analysis has been proposed to prune the search space and reduce the overhead of NAS. The inference engine adapts the maximum memory based on the whole network and leverages the adaptive resource for layer-wise execution.

Strengths: This paper proposes a novel CDF analysis to resolve the additional overhead of searching input resolution for MCUs. The empirical evaluations have illustrated the superior performance of the design. Authors have conducted extensive experiments to validate the performance gain of their NAS algorithm as well as the inference engine. The video in the supplementary material is a good demo for the paper.

Weaknesses: I assume that the microTVM in Figure 3 uses the auto-tuning to optimize the performance. Could authors provide more discussion about why the auto-tuning in TVM fails to works on MCUs (since TinyEngine clearly outperforms it in Figure 3)

Correctness: The claims and method looks correct. The video in the supplementary material illsurates the real deployment of MCUNet on STM32F746.

Clarity: The paper is well written.

Relation to Prior Work: This paper has address the contirbutions compared to the previous hardware-aware NAS with evaluaitons.

Reproducibility: Yes

### Review 3

Summary and Contributions: To run deep learning on MCU effectively, the paper proposes MCUNet which contains a neural architecture search method (TinyNAS) and a lightweight inference library (TinyEngine). TinyNAS optimizes the search space automatically to fit the tiny resource constraints and then performing the one-shot neural architecture search in the optimized space. TinyEngine uses a code generator-based compilation method to eliminate memory overhead in MCU. It is exciting to read this paper. By TinyNAS & TinyEngine co-design, MCUNet obtains SOTA results on ImageNet classification on MCU, and visual&audio wake words tasks.

Strengths: 1. The paper is of great significance - reaching exciting results on a MCU promotes the development of AIoT. 2. The neural architecture and inference engine co-design idea is novel. By deep experimental analysis, the proposed tricks are proven to be effective. 3. The empirical results are excellent - large improvements are obtained over TensorFlowLite micro and CMSIS-NN. 4. Extensive evaluations on different tasks. It has proven to be effective for large-scale visual recognition and practical applications, and visual&audio wake words tasks.

Weaknesses: A minor issue: How does the TinyEngine scheduling affect TinyNAS is not clear enough. I suggest adding a flowchart to detail the co-design process.

Correctness: Yes, they are correct.

Clarity: Yes, well-enough.

Relation to Prior Work: Yes, prior works are clearly discussed.

Reproducibility: No

### Review 4

Summary and Contributions: The authors tackle the problem of tiny ML, with the objective of efficiently deploying deep learning techniques on micro-controllers. The main contributions are: - TinyNAS, a two stage neural architecture search algorithm that first automatically designs a search space adapted to the target platform and then apply NAS techniques to select the best architecture in this space. - TinyEngine, an inference library increasing the speed and reducing the memory required to evaluate a model on a microcontroller unit.

Strengths: The TinyEngine results are impressive, clearly showing its memory and speed improvements over existing libraries using the same model on the same platform. It results in the ability to fit larger models on the device and therefore improves the overall performance (both latency and accuracy) for application in highly constrained environments. The method used to define the search space of tinyNAS is simple, effective and allows to search for architectures that will be very efficient on the specific microcontroller unit targeted. It is, to my knowledge, the first approach that automatically defines the search space to make the most of the resources that will be available at inference time.

Weaknesses: The small amount of space dedicated to the description of the actual NAS procedure is concerning to me. As it is one of the main contributions, I think it should be presented in more detail in the main paper, with a thorough comparison to the other budget-aware NAS methods allowing to assess its performance and make sure that the proposed comparisons are fair. Another concern is the lack of details about the experimental protocol (e.g. detailed description of the datasets, preprocessing used, training times of the different approaches, $m$ used for each experiment) that would make the results difficult to replicate.

Correctness: Is TinyNAS a new neural architecture search algorithm or a smart way of defining efficient search spaces? If it is an actual NAS algorithm, explaining it and running more experiments to compare the effectiveness of the proposed search algorithm (*using the same search spaces*) to other recent methods would be necessary to validate the importance of this contribution. If it is an algorithm allowing to define device-specific search spaces, I think it should be presented as such. Clearly stating that it is not a NAS algorithm and putting the emphasis on the TinyEngine.

Clarity: The paper is overall well written and pleasing to read.

Relation to Prior Work: While inference libraries are not my area of expertise, I think that the TinyEngine is compared to the relevant high-performing inference libraries designed to be used on a microcontroller. The difference between these libraries being explained in section 3.2 and empirically validated in section 4. The relation of the TinyNas algorithm to prior work in neural architecture search is less clear. In the introduction (l.40), it is stated that other approaches focus on applications where memory and storage are abundant and only optimize for FLOPs or latency. However, there exist some NAS approaches that propose to directly optimize the memory footprint of the learned architecture [1, 2]. I think the paper would benefit from a clear comparison of the actual search algorithm (not only the search space design) and other recent NAS methods. [1] Gordon et al. "MorphNet: Fast & Simple Resource-Constrained Structure Learning of DeepNetworks." CVPR. 2018 [2] Veniat et al. "Learning time/memory-efficient deep architectures with budgeted super networks." CVPR. 2018

Reproducibility: No

Additional Feedback: I think that the authors should put more emphasis on the TinyEngine which is an important contribution by itself. Putting the TinyNAS part as an independent work (or secondary contribution) allowed by the improvements provided by the TinyEngine. Rebuttal: The author's feedback addresses and resolves some of the weaknesses I pointed out in my initial review. Given the new results and the additional explanations about their model, I update my evaluation to better recognize the novelty of this work and its potential impact.