This kind of curve usually shows up when different loss definitions are logged (CE vs CE + auxiliary penalty). I saw something similar while building telemetry failure models (reflex system training). The shape looks like an auxiliary loss whose influence shrinks during training (KL regularization, label smoothing annealing, mixture-of-experts load balancing).
The baseline drops faster initially because it has no constraint (standard CE) so it is free to purely minimize surprise. As the penalty term shrinks toward zero, the total loss approaches CE and the curves merge (losses become very close).
At step 0 the baseline is 11 and the non-baseline is 19. CE should start at roughly the same level. So something else must have been added to the loss (L = CE + penalty). That's often seen with KL because the divergence can be large early in training when the model distribution differs significantly from the reference distribution.
@thebearablebull Are you projecting? There is literally no direct mention of a strategy to minimize sales specifically to influence XRP's price in Ripple's Q1 2024 XRP Markets Report.
2021 has been an incredible year but most still do not understand the awesome potential that #NFTs truly represent in the #openmetaverse we lay out our thoughts as to why true digital property rights are so important and impactful https://t.co/Ua9DYtIDiP
Self-Supervised Deep Learning (and semi-supervised and weakly-supervised) making inroads in medical applications.
Labeled data is often scarce in medical applications.
DeFi users don't care what smart contract platform they are using any more than they care what database technology Chase or Robinhood uses. They just want it to be fast, cheap, and secure. That's it.
Thinking new DeFi users have some sort of loyalty to Ethereum is ridiculous.
I continue to be amazed by how little of academic ML research looks how we collect and label data, given that for almost any real application this is the biggest factor in performance.
@zhusu Grayscale bought $76M ETH. Hedge funds will be able to short ETH and buy GBTC at NAV to get the premium. NAV is $36. How do you think it's going to play out? Usually buy the rumor sell the news.