AntiDote: Dynamic CNN Optimization

• AntiDote is an attention-based framework that dynamically prunes feature-map redundancies in CNNs on a per-input basis to optimize runtime efficiency.
• It employs simple average pooling to generate binary masks for channel and spatial pruning, adapting the optimization to input-specific activation patterns.
• Training-phase targeted dropout aligns with dynamic inference, achieving up to 54.5% FLOPs reduction with minimal accuracy loss on benchmarks like VGG16 and ResNet56.

AntiDote is an attention-based dynamic optimization framework for convolutional neural networks that targets runtime efficiency by pruning feature-map redundancy on a per-input basis rather than treating importance as a fixed property of weights or filters. It combines testing-phase dynamic pruning of channels and spatial columns with training-phase optimization by targeted dropout, so that the network learns to remain accurate when low-attention features are removed at inference. In the reported experiments, this training-testing co-optimization yields 37.4% to 54.5% FLOPs reduction on VGG16 and ResNet56 across CIFAR-10, CIFAR-100, and ImageNet100, with negligible or even negative accuracy drops (Yu et al., 2020).

1. Static redundancy, dynamic significance, and the AntiDote premise

The framework is motivated by a limitation of conventional compression methods for CNNs. Unstructured weight pruning, structured pruning, and filter or channel pruning typically estimate importance from model parameters or dataset-level statistics alone. In this view, a channel or filter has a fixed, static significance: once removed, it is absent for all inputs. AntiDote rejects that premise and instead treats significance as input-dependent.

The underlying observation is that CNNs exhibit per-input feature activation. For a particular image, many channels that are important on average may be weakly activated, while channels that are usually weak may become important for rare or specific patterns. This suggests that static pruning must remain conservative, because it must preserve components that are important for at least some inputs. AntiDote therefore shifts the optimization target from model-wide redundancy to per-input redundancy.

This design leads to three stated benefits. First, it can identify and remove per-input feature redundancy while considering model-input interaction. Second, it can remove redundancy in multiple dimensions, namely channel and spatial dimensions. Third, training-testing co-optimization helps preserve accuracy even when the feature pruning ratio is high. In this sense, AntiDote is best understood as feature-level, attention-guided dynamic sparsification rather than static model compression (Yu et al., 2020).

2. Attention mechanism and multi-dimensional masking

AntiDote operates on an intermediate feature map

$F \in \mathbb{R}^{C \times H \times W},$

where $C$ is the number of channels and $H,W$ are spatial dimensions. It models redundancy along two axes: channel redundancy and spatial-column redundancy.

For channel importance, AntiDote uses global average pooling: $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$ This yields one scalar per channel and functions as a lightweight activation-energy estimate for the current input.

For spatial importance, it averages across channels: $$A_{\text{spatial}}(F,h,w)=\frac{1}{C}\sum_{i=1}^{C}F_{h,w}(i).$$ This produces an $H \times W$ map that functions as a heat map of where the network is looking.

These attention values are then converted into binary masks through top-$k$ selection. For channel pruning, with retention rate $p$,

$k=\text{int}(p \cdot C),$

and only the $k$ channels with largest attention are retained. For spatial pruning,

$C$0

and only the top-$C$1 spatial positions are kept. The mechanism is deliberately simpler than SE- or CBAM-style attention modules: AntiDote uses averaging rather than extra MLPs or convolutions, and repurposes attention not to reweight features but to binarize them into masks and remove redundant features entirely.

A central property of the framework is its multi-dimension flexibility. Each layer may use only channel pruning, only spatial pruning, or both simultaneously, with layer-specific retention ratios determined by sensitivity analysis. This implies that the framework can adapt to whether redundancy is primarily channel-wise, spatial, or mixed in a given architecture or dataset (Yu et al., 2020).

3. Testing-phase dynamic optimization

At inference, AntiDote performs a layerwise dynamic pruning pipeline. Given a feature map $C$2, it first computes $C$3 and $C$4, then generates binary masks $C$5 and $C$6, and applies them as

$C$7

where broadcasting is used across the omitted dimensions.

The key consequence is that the next convolution need only process active channels and active spatial locations. For a standard convolution from $C$8 to $C$9,

$H,W$0

If only a fraction $H,W$1 of channels and $H,W$2 of spatial locations are active for a given input, the approximate cost becomes

$H,W$3

with an approximate per-layer speedup

$H,W$4

This dynamic computation differs from early exiting or dynamic depth methods. AntiDote keeps the full network depth but reduces width and spatial computation inside layers. It also differs from topology-level routing methods, because the conditional computation occurs at the feature-map level rather than by choosing among subnetworks or experts.

A common misconception is to equate the reported savings with direct wall-clock acceleration. The work primarily reports algorithmic FLOPs reduction. Because the induced sparsity is structured but input-dependent, actual latency gains depend on whether the target hardware and kernels can exploit variable active channels and spatial patterns efficiently (Yu et al., 2020).

4. Training-phase co-optimization with Targeted Dropout

The training-phase component is Targeted Dropout, abbreviated TTD. Unlike standard dropout, which randomly removes activations for regularization, TTD removes specifically low-attention channels and spatial positions. The objective is not generic regularization but alignment between training-time perturbations and test-time pruning.

During training, each convolutional layer repeats the same attention and masking procedure used at inference. The masked feature map

$H,W$5

is then passed to the next layer, and backpropagation proceeds through the masked computation graph. Repeated exposure to low-attention feature removal encourages the network to depend less on weak features and to concentrate discriminative information in high-attention channels and regions.

Because different layers tolerate pruning differently, AntiDote introduces block-wise sensitivity analysis. For each block, varying pruning ratios are tested and the resulting accuracy drop is observed. On VGG16 for CIFAR-10, an aggressive blockwise pattern $H,W$6 caused the pruned model to drop below 70% accuracy and was therefore treated as exceeding a recoverable upper bound. Training then follows a dropout ratio ascent strategy: begin with a modest ratio such as 0.1, train to convergence or plateau, increase by a small step such as $H,W$7, and continue until the target ratios are reached.

This co-optimization removes a central train-test mismatch present in many dynamic inference methods, where the model is trained assuming full computation but deployed with partial computation. AntiDote instead trains the model under the same low-attention removals it will encounter at test time, so no additional fine-tuning after pruning is required (Yu et al., 2020).

5. Empirical performance and ablation findings

The reported implementation uses PyTorch, with experiments on CIFAR-10, CIFAR-100, and ImageNet100, and on VGG16 and ResNet56. Static pruning baselines include $H,W$8-norm filter pruning, Taylor pruning, Geometric Median pruning, and Function-Oriented pruning.

The principal quantitative results are as follows.

Setting	FLOPs reduction	Final accuracy
VGG16 on CIFAR-10	53.5%	93.1%
ResNet56 on CIFAR-10	37.4%	93.2%
VGG16 on CIFAR-100, setting 1	40.4%	73.2%
VGG16 on CIFAR-100, setting 2	44.9%	72.9%
VGG16 on ImageNet100, setting 1	51.2%	79.6%
VGG16 on ImageNet100, setting 2	54.5%	79.4%

For VGG16 on CIFAR-10, the baseline is 93.3% accuracy at $H,W$9 FLOPs, while AntiDote reaches 93.1% at $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$0 FLOPs using channel pruning ratios $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$1 and no spatial pruning. For ResNet56 on CIFAR-10, the baseline is 93.0% at $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$2 FLOPs, while AntiDote uses channel pruning ratios $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$3 and spatial pruning ratios $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$4 to reach 93.2% at $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$5 FLOPs. For VGG16 on ImageNet100, the best-reported configuration uses channel pruning $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$6 and spatial pruning $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$7, reducing FLOPs from $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$8 to $$A_{\text{channel}}(F,c)=\frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_c(i,j).$$9 while increasing accuracy from 78.5% to 79.4%.

The redundancy composition analysis further clarifies where the savings arise. On VGG16 with ImageNet100, channel pruning contributes about 2.4% FLOPs reduction, while spatial pruning contributes about 52.1%, indicating that on $$A_{\text{spatial}}(F,h,w)=\frac{1}{C}\sum_{i=1}^{C}F_{h,w}(i).$$0 images most redundancy is spatial. On VGG16 with CIFAR, spatial redundancy is limited and virtually all speedup comes from channel pruning. On ResNet56 with CIFAR-10, both are important, with 18.2% reduction from channels and 19.2% from spatial locations.

Ablation results support the choice of attention as the pruning signal. Attention-based pruning outperforms random pruning, while inverse-attention pruning is catastrophic. In VGG16, pruning only 10% of top-attention channels causes about an 80% accuracy drop, indicating that the largest attention scores are indeed strong indicators of per-input importance (Yu et al., 2020).

6. Practical interpretation, limitations, and relation to adjacent methods

AntiDote sits at the intersection of static compression, dynamic inference, and attention-guided computation. Its distinctive feature is that it prunes feature maps rather than just filters or weights, and does so with simple attention statistics and explicit training support. This makes it conceptually close to spatially adaptive computation and conditional computation, but its granularity is the channel and spatial-column level inside standard CNN layers.

Its practical appeal is strongest in settings where average FLOPs, memory traffic, and energy matter more than uniform per-input latency. Mobile and edge inference, real-time analytics, and large-scale cloud inference are natural applications, because AntiDote can exploit the fact that easy inputs or background-heavy inputs activate only part of the model’s representational capacity.

The limitations are equally clear. Dynamic masks increase implementation complexity relative to static pruning. Because active computation varies by input, latency may vary as well. Real acceleration depends on runtime-aware kernels and indexing support for structured dynamic sparsity. The method also depends on sensible layerwise sensitivity tuning; poorly chosen pruning ratios can hurt convergence or produce irreversible degradation.

A further clarification concerns attention itself. AntiDote does not use attention in the sense of adding a powerful auxiliary module to improve representation quality. It uses simple averages as an inexpensive importance proxy and then converts that proxy into hard pruning decisions. This suggests that its contribution lies less in expressive attention modeling than in the redefinition of runtime optimization as per-input, multi-dimensional pruning guided by activation statistics (Yu et al., 2020).