We thank all the reviewers for their constructive comments. We will provide details in the final draft. MCUNet shows consistent improvement across different devices (F746, H743) and tasks (classification, detection). R1: Whether the overall network topology brings major improvement. R2: Why the auto-tuning in TVM fails to work on MCUs.

Neural Information Processing Systems http://nips.cc/

We thank the reviewers for their detailed comments. We are glad to see a generally positive assessment of our work. We will report larger-scale results in the final draft. Below, we address specific reviewer comments. When using 12G GPU machines, this difference is significant.

They have demonstrated remarkable achievements across various applications, consistently delivering state-of-the-art outcomes. MEFT is the first method that is proposed to modify a PLM to its reversible variant. Another limitation of MEFT is its lower score when trained in FP16 and on a deeper model. For deeper models, we offer a practical and effective setting in Figure 7. For the reader's easy understanding, in this section, we explain MEFT For the second reversible layer, if we don't switch the order of Compared to GLUE tasks where all tasks are classification tasks and the classification heads are randomly initialized, the question-answering tasks are sequence-to-sequence tasks and need the pre-trained output layer that shares the same parameters as the word embedding layer.