AMC (AutoML for Model Compression and Acceleration on Mobile Devices)
Algorithms AIMech:AMCAccelerationAndModelCompressionOnMobileDevicesWithRL DRAFT
Aliases: AMC, AutoML for Model Compression, AMC Reinforcement Learning Channel and Fine-Grained Pruning
AMC (AutoML for Model Compression) is a reinforcement-learning-based neural network pruning framework that processes a pretrained network layer-by-layer using an 11-dimensional layer state vector s_t = (t, n, c, h, w, stride, k, FLOPs[t], reduced, rest, a_{t-1}) and a continuous-action Deep Deterministic Policy Gradient (DDPG) actor-critic agent to predict each layer's channel or fine-grained pruning sparsity ratio a_t in (0, 1], evaluates post-pruning validation accuracy without fine-tuning as a fast delegate reward, and supports both Resource-Constrained Compression (via duty-bounded action clamping W_duty = alpha*W_all - action_max*W_rest - W_reduced with R_err = -Error) and Accuracy-Guaranteed Compression (via R_FLOPs = -Error * log(FLOPs) or R_Param = -Error * log(#Param)).
Identity Notes & Concept Scope
Distinct from hand-crafted sensitivity pruning heuristics (Li et al., 2016; He et al., 2017) that ignore cross-layer dependencies, distinct from Runtime Neural Pruning (Lin et al., 2017) which selects coarse 4-action dynamic channel groups per input image without reducing static model size, and distinct from HAQ (AIMech:HAQHardwareAwareAutomatedQuantizationWithMixedPrecision) which allocates mixed-precision quantization bitwidths rather than pruning sparsity ratios.
Claims & Grounded Evidence
4 Claims
Mechanism
#amc-continuous-ddpg-state-and-zero-finetuning-delegate-reward
AMC characterizes each layer L_t by the 11-dimensional state vector s_t = (t, n, c, h, w, stride, k, FLOPs[t], reduced, rest, a_{t-1}) (Eq. 1), predicts a continuous sparsity action a_t in (0, 1] via a DDPG actor with truncated-normal exploration mu'(s_t) ~ TN(mu(s_t | theta_t^mu), sigma^2, 0, 1) (Eqs. 2-3), and evaluates validation accuracy immediately after pruning all layers without fine-tuning as a correlated, zero-retraining proxy for final fine-tuned accuracy.
Mechanism
#amc-resource-constrained-duty-clamping-and-accuracy-guaranteed-rewards
For resource-constrained compression, AMC uses reward R_err = -Error (Eq. 4) and enforces the target reduction ratio alpha via Algorithm 1 by bounding each layer's action to [W_duty / W_t, action_max] where W_duty = alpha*W_all - action_max*W_rest - W_reduced; for accuracy-guaranteed compression, AMC uses R_FLOPs = -Error * log(FLOPs) (Eq. 5) or R_Param = -Error * log(#Param) (Eq. 6).
Application
#amc-imagenet-resnet50-and-mobilenet-mobile-speedup
On ImageNet, AMC pushes ResNet-50 fine-grained compression from 3.4x to 5x without accuracy loss (automatically pruning 3x3 convolutions more aggressively than 1x1 convolutions), improves VGG-16 4x FLOPs pruning by 2.7% over hand-crafted policies, and achieves 1.81x-1.95x measured latency speedup on MobileNet on a Google Pixel 1 Android phone (and 1.53x on Titan XP GPU).
Limitation
#amc-multi-stage-iterative-pruning-for-extreme-sparsity-limitation
Under extreme single-shot fine-grained compression (e.g., >=4x sparsity), validation accuracy without retraining collapses and ceases to provide an informative reward signal, requiring a multi-stage iterative prune-and-retrain schedule (e.g., four stages at 50%, 35%, 25%, and 20% density with 30 epochs of fine-tuning per stage).
⚠️ Curated Limitations & Negative Results
- Zero-retraining validation accuracy becomes uninformative under extreme single-shot fine-grained compression ratios (>=4x), requiring multi-stage iterative AMC search and fine-tuning.
- Fine-grained unstructured pruning policies require specialized sparse-matrix hardware accelerators (such as EIE or SCNN) to translate parameter sparsity into wall-clock inference speedup.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
broad_match | AMC is an automated reinforcement-learning pruning and compression framework for deep convolutional neural networks. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r819-primary-pdf
|
Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han
|
primary paper
full_text
|
ID: https://arxiv.org/bibtex/1802.03494
sha256:2aa8c022ce47…
|
r819-arxiv-abs
|
Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/1802.03494
sha256:6420acfd7435…
|
r819-metadata-cr
|
Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han
|
primary paper
abstract_only
|
ID: https://arxiv.org/bibtex/1802.03494
sha256:d7d415cf2fd3…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T15:02:00Z
Updated: 2026-10-19T15:02:00Z
Source File:
data/curated/AMCAccelerationAndModelCompressionOnMobileDevicesWithRL.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories