AI Hardware & Software Stack
Which statement about pruning neural networks is correct?
- AIt removes unimportant weights by zeroing them; unstructured pruning, which zeros scattered weights, is hard to accelerate in hardware
- BIt adds weights to always improve accuracy
- CIt reduces the training data
Show answer and explanation
Answer: A. It removes unimportant weights by zeroing them; unstructured pruning, which zeros scattered weights, is hard to accelerate in hardware
Pruning zeros out weights with little influence, such as those with small values, to reduce computation and memory. But when zeros are scattered, tracking where they are costs effort, and matrix units and memory operate on contiguous chunks, so the real speedup is small. Structured pruning, which zeros whole channels or layers or follows a fixed pattern, is much easier to accelerate in hardware.
- A. Correct. Structured pruning zeros weights in fixed patterns.
- B. It reduces weights.
- C. It removes weights, not data.