korrents

What Michael Carbin thinks about pruning

@michael-carbin · 5 positions · 0 changes of mind

Associate professor at MIT, where he leads the Programming Systems Group, and co-author of the lottery ticket hypothesis.

Everything they publish, on ppll ↗

Michael Carbin did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

5 dated positions, 2018, in their own words. Our reading of what Michael Carbin has said — not written or endorsed by them.

  1. Our readingPruning cuts the parameters of a trained network by over 90% without costing accuracy, so inference gets smaller and faster for free.

    Their wordsNeural network pruning techniques can reduce the parameter counts of trained networks by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy.

    ↗Frankle & Carbin, "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" (arXiv)arxiv.org 2nd of 10 in this piece

    neural networks

    Interesting

  3. Our readingThe saving pruning offers is only ever at inference, because the sparse architectures it produces are hard to train from the start.

    Their wordsHowever, contemporary experience is that the sparse architectures produced by pruning are difficult to train from the start, which would similarly improve training performance.

    ↗Frankle & Carbin, "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" (arXiv)arxiv.org 4th of 10 in this piece

    Interesting

  5. Our readingPruning does not merely shrink a network; it uncovers subnetworks whose original initialization was what made them trainable at all.

    Their wordsWe find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively.

    ↗Frankle & Carbin, "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" (arXiv)arxiv.org 6th of 10 in this piece

    Interesting

  7. Our readingA winning ticket wins on its initial weights: the connections it keeps started at values that happen to make training work.

    Their wordsThe winning tickets we find have won the initialization lottery: their connections have initial weights that make training particularly effective.

    ↗Frankle & Carbin, "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" (arXiv)arxiv.org 8th of 10 in this piece

    Interesting

  9. Our readingA small enough winning ticket learns faster than the network it was cut out of, and ends up more accurate than it.

    Their wordsAbove this size, the winning tickets that we find learn faster than the original network and reach higher test accuracy.

    ↗Frankle & Carbin, "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" (arXiv)arxiv.org 10th of 10 in this piece

    Interesting