EdgeNeXt Splits Convs for Edge Speed

EdgeNeXt splits channels into conv plus attention paths with transposed attention, so edge chips run global models fast.

Channel splits plus transposed attention cut edge latency 2 times

Why Does This Exist?

Edge NPUs run depthwise well but large token attention poorly. Full spatial attention stalls them.

EdgeNeXt transposes attention to channels. It fits efficient architectures, rivals mobilevit latency, and uses convolutional layers splits.

Think of It Like This

Splitting checkout lanes by cash plus card

One long lane jams mixed buyers. Split to cash plus card lanes with a shared manager and both move.

Splits are lanes, transposed attention is the manager. Where it stops: too many lanes starve each, just as thin splits lose capacity.

How It Actually Works

Split $C$ into $g$ groups. Run $3 \times 3$ depthwise per group plus Split Depthwise Transpose Attention $softmax(K^T Q/\sqrt{d})V$ over $C \times C$ not $N \times N$ . Merge with $1 \times 1$ .

Worked example

$N = 3136$ , $C = 128$ . Spatial pairs $3136^2 = 9.8$ M. Channel pairs $128^2 = 16{,}384$ , about $600$ times fewer ( $9{,}834{,}496/16{,}384 = 600$ ). Latency $18$ ms to $9$ ms on Jetson runs at $79$ percent.

Watch Out For

Transposing when spatial order leads accuracy

Channel attention blurs where. Symptom: classification holds but detection sags. Fix: keep spatial blocks for dense heads and use transposed mainly for classification stages.

The Quick Version

EdgeNeXt splits to depthwise groups plus channel attention.
Transposed $128^2$ pairs replace $3136^2$ spatial pairs.
Fixture $600$ times fewer pairs at $79$ percent.
Edge latency halves on Jetson class runs.
Keep spatial attention where location leads.