Launched this week
Coincidex
Transfer-aware, replay-free continual learning for PyTorch
14 followers
Transfer-aware, replay-free continual learning for PyTorch
14 followers
Most continual-learning methods either forget (fine-tuning) or isolate and forfeit transfer (fixed masks, adapters). Coincidex does both: similar tasks share sub-networks for forward transfer, dissimilar ones stay isolated so old skills survive — with no replay buffer. A tiny, drop-in PyTorch layer (torch + numpy, CPU-friendly) that learns task similarity from context. It's also honest: reproducible benchmarks and a clear map of where it works and where it doesn't. Apache-2.0 license.
Dropped it into a small continual-learning setup and was surprised how little code it took — basically one layer swap. Love that it ships without a replay buffer and admits where it falls short instead of overpromising.
@esilape7l That was the exact goal! A single layer swap and honest limits. Really appreciate you trying it out.
finally someone tackling continual learning without the replay buffer baggage, kind of refreshing actually. the task similarity layer is tiny but the no replay + honest benchmarks angle is what got me, though i'd want to see it on tougher task sequences before fully buying in
@berkeseres Appreciate that! We wanted to be completely transparent about those limits. Looking forward to testing it on much rougher task sequences next.
dropped it into a small continual vision setup and it just worked without me hand-tuning masks. the honesty about failure modes is refreshing, more libraries should do that.
@fahri1585951 Awesome to hear it worked out of the box! Zero manual tuning and honest limits were exactly what we aimed for.
one thing that would make this much more useful for me — a small visualisation of the learned task-similarity matrix as it evolves across tasks. even just a heatmap saved per checkpoint would help me trust the routing decisions and debug when transfer isnt kicking in. could be optional and opt-in so it doesnt slow anything down.
@furkan578657 Love this idea. An opt-in heatmap would make trusting the routing decisions way easier. Officially adding this to our to-do list. Thanks for taking the time to check it out!
Dropped it into a small continual-learning test and watched old task accuracy hold up without any replay buffer, which is honestly impressive. The honest part is what got me, actually seeing the failure modes mapped out saved me a ton of guesswork.
@oglukebabc72648 That’s awesome to hear! We mapped out the limits because nobody likes guessing why a model broke. Glad the no-buffer routing held up for your test. Thanks for checking it out!
the context-driven similarity learning without a replay buffer is genuinely clever, that alone usually forces a tradeoff you avoided. appreciate the honest breakdown of where it falls short too, that kind of transparency is rare.
@nazmiyeabike Thank you! We figured honest limits are always better than hype. Glad the no-buffer routing made sense to you.