ResNeXt: Aggregated Residual Transformations
What you'll learn: How ResNeXt uses grouped convolutions to build more powerful networks without exploding computational costs.
The Core Innovation
ResNeXt extends ResNet by introducing a new dimension called cardinality — the number of parallel transformation paths within a block. Instead of making networks deeper or wider (more channels), ResNeXt splits transformations into multiple identical branches that are then aggregated.
Think of it like a team project: instead of one person doing all the work (single path) or hiring more skilled workers (wider channels), you split the task among multiple teams working in parallel, then combine their results.
Grouped Convolutions for Efficiency
ResNeXt achieves this through grouped convolutions (which you've already learned). Each path applies the same architecture but operates on different channel groups independently. For example, with cardinality C=32, your input channels are split into 32 groups, each processed separately, then concatenated.
The beauty: increasing cardinality from 1 to 32 dramatically improves accuracy while keeping computational complexity nearly constant. You're not just making the network wider — you're creating more diverse feature transformations.
The Mathematical View
A standard ResNet bottleneck has one transformation path. ResNeXt has C paths:
- Split input into C groups
- Apply identical transformations to each group (conv-BN-ReLU-conv-BN-ReLU-conv)
- Aggregate outputs by addition
- Add skip connection
The grouped convolution does this efficiently in one operation rather than literally creating C separate branches.
Key Takeaway: ResNeXt demonstrates that cardinality (the number of parallel transformation paths) is more effective than depth or width alone, and grouped convolutions make this architectural choice computationally practical.