rolan2kn (@rolan2kn) profile picture

R&D Staff Engineer at Synopsys, Ph.D. from @dcc @uchile | Computational Topology and Geometry. #TopologicalDataAnalysis

Santiago, Metropolitana de San
Joined January 2019
Photo by rolan2kn (@rolan2kn)Photo by rolan2kn (@rolan2kn)Photo by rolan2kn (@rolan2kn)Photo by rolan2kn (@rolan2kn)
Nice! TDA features are perfect for this task: surprising that it has not been tried before, it seems so natural a posteriori. I'll look for people in Chile to replicate this kind of work, maybe we can use our pseudodistances for faster search ojs.aaai.org/index.php/AAAI/… @rolan2kn
Replying to @viditnanda
Why use TDA for this? Because an idealised version of the problem is viscerally geometric. If you forget the chemical information for a second and pretend that the target is rigid, then you are looking for a signature shape in the boundary of your ligand. 4/
2
6
327
I completely agree with this.
Young scientists regularly ask me for career advice. Academia or industry? Big company or startup? US or Europe?  Good scientists in AI disciplines are fortunate to have many choices. But choosing can be stressful. I always give the same advice. 1/10
1
54
Mel Blanc, the voice of Bugs Bunny, was in a serious car accident that put him in a coma. After many unsuccessful attempts to get him to talk, the doctor asked, "Bugs, can you hear me?" Mel responded in Bugs voice: "Whats up Doc?" They used this technique to lead him out of his coma.
597
10,767
482
108,691
24,066,988
Looking through the Dune 2 (amazing movie!) art book, I was thinking about the relationship between art and games. In the early days, you just couldn’t apply much art to a game. The fundamental display resolution and color palette were very limited, and distribution media size was an even harsher limit. Significant amounts of engineering effort and artist contortions were called on to allow just a little bit more artistry to shine through. Fine art could be done on the boxes, manuals, and advertisements, but it had little connection to the game experience. The introduction of CD-ROM made distribution size almost a non-issue, such that the cost of creating art dominated. We had a few enemy characters in Doom 2 that were basically “paint overs”, and Doom 3 could have really benefitted from having multiple variations on the enemies, but that was all we could manage with our team size. Unlike many aspects of game development, art can largely be done in parallel, and companies started building out artist armies that could fill out multiple CD-ROMs, DVDs, and Blu-ray discs. Players value art, but there is a diminishing return effect for the typical player. The hit-driven, power-law return nature of the AAA gaming business generally pushes the creation of art past what would be a natural value point, which in some ways diminishes the value of individual works of art, and leads to a lot of work being created and never used. Doom 2016 went through multiple internal reboots, discarding entire art styles and characters along the way. One of the things that stood out for me during my early VR experiments when I was using scenes from Rage was how much amazing detail was already there that I never even noticed in dozens of conventional play throughs. There is SO MUCH in modern games that barely gets looked at by most players. The modern gaming cosmetics markets are actually pretty cool from an art value standpoint — there is precise feedback on how much a given work of art is valued by players, and there are all the traditional sales tools available to try to maximize the value of the art. I haven’t worked on a team that did cosmetics, but I bet there are a lot of interesting things to learn there. It would be great if artists got bonuses when their cosmetics are hugely successful.
85
150
11
1,771
260,023
Replying to @fchollet
@fchollet claimed that "learning to reason" (robustly across problem instances) is hard because of the nature of the MLE objective. This paper proves that claim. The reason is "statistical features inherently exist in data distributions, but can hinder model generalization performance [because learning them can mask the underlying function]". H. Zhang, L. H. Li, T. Meng, K.-W. Chang, and G. Van den Broeck, “On the paradox of learning to reason from data,” in IJCAI, 2023, starai.cs.ucla.edu/papers/Zh…
Replying to @fchollet
From 2013 to 2016 I was actually quite convinced that RNNs could be trained to learn any program. After all, they're Turing-complete (or at least some of them are) and they learn a highly compressed model of the input:output mapping they're trained on (rather than mere pointwise associations). Surely they could perform symbolic program synthesis in some continuous latent program space? Nope. They do in fact learn mere pointwise associations and completely useless for program synthesis. The problem isn't with what the function space can represent -- the problem is the learning process. It's SGD.
9
64
4
364
103,517
Excited to share our latest work on a game-changing method for comparing persistence diagrams! 🚀 Our paper, 'A Class of Topological Pseudodistances for Fast Comparison of Persistence Diagrams,' was accepted at AAAI'24. #AAAI24 #TopologicalDataAnalysis 1/n
1
2
1
403
We introduce a class of pseudodistances called Extended Topological Pseudodistances (ETD)s a faster approximation to Sliced and classical Wasserstein distances while being computationally lighter and close to Persistence Statistics at the lower complexity extreme. 2/n
1
1
62
Our distance is also parallel friendly, making it suitable for heavy workloads with large sets of dense persistence diagrams that were unthinkable before. A join work with @MirceaSci See the paper at arxiv.org/abs/2402.14489 and code at github.com/rolan2kn/aaai2024…. n/n
1
32
Hello #AAAI24, going to present joint work with @rolan2kn on a new class of efficient paeudometrics for comparing Persistence Diagrams! Are any #TDA colleagues around ?
2
4
313
Topological Data Analysis as you've never seen it before: @prof_g is bringing my lecture notes (people.maths.ox.ac.uk/nanda/…) to life in a new animated video series (youtube.com/playlist?list=PL…)! Here are the videos which have been released so far... enjoy! And tell your friends.
6
106
8
453
63,396
rolan2kn retweeted
15
458
55
1,475
304,660
And for a concise introduction and an exciting literature review using this unifying framework: 👉Check out the survey "Architectures of Topological Deep Learning" led by @mathildepapillo with @naturecomputes @HajijMustafa and I. arxiv.org/abs/2304.10031 @bioshape_lab
3
12
831
As a child, Gauss famously (& probably apocryphally) annoyed his teacher by coming up with a clever workaround for summing the numbers 1 through 100. Here's a super cool formula & exposition by my friend @hrldcpr generalizing Gauss' trick to N dimensions: x.st/visual-sum-of-cubes/
9
155
8
662
Here's a neat trick to impress your friends Let's say you have some curve with a random shape, possibly even self-intersecting. Can you measure its length? This isn't just a parlor trick -- it has many practical applications. For example, the curve could be a strand of DNA 1/n
11
139
24
896
Want to venture into DL+even more exotic math (algebraic topology, differential geometry, representation theory, and differential equations)? Check the Italian GDL summer school with @pimdehaan @maurice_weiler @crisbodnar @Francesco_dgv and yours truly: sci.unich.it/geodeep2022/#sc…
📢 New & improved material to dive into geometric deep learning! 💠🕸️ We (@mmbronstein @joanbruna @TacoCohen) delivered our Master's course on GDL @AIMS_Next once again & we make all materials publicly available! geometricdeeplearning.com/le… See thread 🧵 for gems 💎 & dragons 🐉!
3
33
159
We're excited to release #GiottoDeep, a new open-source Python library for deep learning and topological data analysis! lnkd.in/eiTtpdCH #tda #pytorch #deeplearning #python #dataanalysis
20
73
There is no better way to start day 2 of #logml22 than with a keynote from @haharrington!🤩 What a wonderful & comprehensive survey on all of the recent applications of topological data analysis to biology!🤯🧬 #tda #geometry #machinelearning #summerschool
3
1
21
Usually, Ph.D. success is measured by the number of papers/citations/problems solved. There is a better way: the **skills** you acquire
6
31
4
264
🔓Special Issue "Information Geometry and Optimal Transport" is now complete including the Editors' introduction "When optimal transport meets information geometry" (32 pages). ✨Free to read or download✨ all articles for a limited time: 👉 link.springer.com/collection…
1
51
196