Marco Bonici
LightNews LightNews
banner
marcobonici.bsky.social
Marco Bonici
@marcobonici.bsky.social
350 followers 250 following 42 posts
PostDoc at the Waterloo Center for Astrophysics. Working on Cosmic Voids and Large Scale Structure of the Universe. JuliaLang enthusiast. Improv Theatre and StandUp comedy lover.
Posts Replies Media Videos
marcobonici.bsky.social
It's that time of the year again.
A meme offering a critique of a growing trend in software development. The image specifically targets individuals who, instead of optimizing their code, intentionally do not spend time trying to write efficient implementations. They then highlight the poor performance of this code, subsequently replace it with a neural network solution, and claim significant performance improvements—thereby presenting themselves as highly competent.
April 22, 2025 at 4:28 AM
marcobonici.bsky.social
We tested Effort.jl on the PT-challenge & BOSS datasets. Results? An agreement at the level of MonteCarlo noise and the analysis runnign on a laptop in a bit more than hour, without analytical marginalization! We also used the brand new MCHMC sampler, which overperformed the classical NUTS!
Triangle plot showing the standard pybird chains alongside those obtained using Effort.jl, for the BOSS analysis, focusing on cosmological parameters. The lower triangular section compares pybird and Effort.jl to validate the precision of our emulator and likelihood implementation. In the lower triangular part we focus on the comparison with the pybird chains, to ensure the correctness of our results, while in the upper part of the plot we compare the two gradient based samplers employed in this study.
January 9, 2025 at 3:05 PM
marcobonici.bsky.social
The results are extremely good: we found an expression that can give us the D(z) as a funciton of redshift and cosmological parameters with an accuracy better than 0.1% for most cases. Using the SymbolicRegression.jl D(z) gives almost the same results than the ones found when using the ODE!
95% distribution of the percentage errors, comparing the ODE and the symbolic regression rescaling. As can be seen, the two approaches give almost indistinguishable results, with the sole exception of the monopole emulator trained with 60,000 samples, as in that case the growth factor symbolic emulator error is not negligible compared to the ODE rescaled emulator error.
January 9, 2025 at 3:05 PM
marcobonici.bsky.social
We compared the emulator performance:
- with no preprocessing
- with an As rescaling
- with a D(z) rescaling
- with an As*D(z) rescaling
We also checked the impact of a bigger training dataset.
The rescaling strategies are more impactful than using a bigger training dataset!
95% distribution of the percentage errors. We show the impact of the different rescaling schemes employed. The dot-dashed and solid lines refers respectively to a training dataset with 20,000 and 60,000 elements. As can be seen, a bigger dataset helps in enhancing the performance, but the rescaling is even more impactful.
January 9, 2025 at 3:05 PM
marcobonici.bsky.social
Here's mine:
February 4, 2024 at 2:15 PM
marcobonici.bsky.social
I asked Dall-E to make a portrait of me writing on a blackboard, based on a photograph I gave him. Maybe is not super accurate, but my hair look fabulous.
Also, writing with two hands simultaneously? I am a master🤭
October 3, 2023 at 1:07 AM