[HN Gopher] Direct initialization of transformers using larger p...
___________________________________________________________________
Direct initialization of transformers using larger pretrained ones
Author : PaulHoule
Score : 24 points
Date : 2023-12-22 18:54 UTC (4 hours ago)
(HTM) web link (arxiv.org)
(TXT) w3m dump (arxiv.org)
| jacobsimon wrote:
| This might sound like nonsense to someone who actually
| understands neural networks, but I had a similar idea recently
| that perhaps we could use LLMs or diffusion models to generate
| the weights for a new network, e.g. if you think of a layer of a
| network as an image where each pixel represents the strength of a
| connection, then perhaps you could generate all the initial
| weights for a specialized network from a more general/powerful
| network. Curious if that's been tried or if there are fundamental
| limitations that would prevent it.
| turtles3 wrote:
| I believe what you're looking for is Hypernetworks
|
| https://arxiv.org/abs/1609.09106
| jacobsimon wrote:
| Very relevant thanks!
| londons_explore wrote:
| I would like to know if this approach is actually transferring
| knowledge, or if random initialization is just not the optimal
| starting place - for example, maybe having a different
| distribution of weights is better.
|
| A good test would be to apply this technique, but then train on a
| totally different kind of data - for example take a text LLM,
| apply this technique, but then train on audio/music data and see
| if this technique reduces training time over random
| initialization.
| peterfirefly wrote:
| Sounds like a good test.
|
| We need something like this for "non-neural" things like
| transformers and normalization layers:
|
| https://proceedings.mlr.press/v157/skorski21a/skorski21a.pdf
| DarmokJalad1701 wrote:
| From Page 7:
|
| > In the case of GPT-2 training, random initialization requires
| 64x10^9 tokens to reach a perplexity of 12, while weight
| subcloning accomplishes this in just 64x10^9 tokens, again
| demonstrating a 4x training speedup
|
| ... what?
| mirekrusin wrote:
| Looks like typo, from chart it seems to be around 20x10^9 for
| weight subcloning (doesn't look like 16x10^9 <<which would be
| 4x>> either).
| mirekrusin wrote:
| Instead of pruning least contributing neurons wouldn't it be
| better to ie. nominate next lowest contributing neuron or most
| similar one to merge them into one or redistribute its impact
| onto others?
___________________________________________________________________
(page generated 2023-12-22 23:00 UTC)