HW/SW Collaborative Techniques for Accelerating TinyML Inference Time at No Cost
Bibliographic record
Abstract
With the unprecedented boom in TinyML development, optimizing Artificial Intelligence (AI) inference on resource-constrained microcontrollers (M CU s) is of paramount importance. Most of the existing works focus on peak memory or computation reduction. The tasks are partitioned in the patch-based or device-based during the execution. However, it comes with a price of the latency and communication overhead. In this paper, we propose several techniques to accelerate the Convolutional Neural Networks (CNN s) inference process. These techniques are both architecture- and application-aware. From the application perspective, 1) we maximize computation reuse through instruction reordering, 2) fuse several linear layers together to improve computation patterns, and 3) enable memory reuse of intermediate buffers for improving memory behavior. From the architecture perspective, we propose techniques that take into account knowledge about underlying architecture of the MCU including 1) cache-aware and 2) multi-core parallelism-aware techniques. Those solutions only require the general MCUs features thus demonstrating board generalization across various networks and devices. These techniques come at no additional cost. It improve the inference latency without any compromise of the model accuracy or the model size. Our evaluation on a use-case from the health-care domain with real-data set for four CNNs - LeNet, AlexNet, ResNet20, and SqueezeNet - show that we achieve up to 71 % reduction in inference latency.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.004 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.014 | 0.004 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".