Definition
Plain language
A neural network design that mixes convolutions with attention for sequence tasks.
As stated in the literature
A sequence model architecture combining self-attention with convolutional augmentation, originally for speech, recombined by agents in AIRA-Design's LRA experiments.
Why it matters: It's a strong, well-known recipe that agent-driven architecture search uses to test whether systems can rediscover or improve on established designs.
For example, an architecture-search agent proposes a Conformer-style block as a candidate for a long-range sequence task and benchmarks it against pure-attention alternatives.
Heard on the show
“Conformer-style convolutional augmentation.”Episode 053 — An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script