robinryder.bsky.social
Mathematician at Imperial College London. Bayesian statistics, Data science, Languages, Phylogenies.
56 posts
814 followers
110 following
Regular Contributor
Active Commenter
comment in response to
post
Bienheureux ceux qui ne comprennent pas ce message.
comment in response to
post
Wait for my upcoming book, in which I will pin the whole of human history on you and other people who pin the whole of human history on people-who-pin-the-whole-of-human-history-on-one-topic.
comment in response to
post
I suggest you pin the whole of human history on people-who-pin-the-whole-of-human-history-on-one-topic.
comment in response to
post
Are you giving two talks on the same day at the same time?? You're good!
comment in response to
post
This was good fun to look at. Many thanks to lead author (and outgoing postdoc) Federico Pavone, and to our wonderful co-author Daniele Durante!
arxiv.org/abs/2502.11868
comment in response to
post
We also look at a neuroscience application, to analyse brain connectivity networks.
comment in response to
post
We applied this to find the underlying structure of the Infinito criminal network. It is surprisingly well resolved, and it recovers both known structure and interesting positions of certain members of the network.
comment in response to
post
We perform inference with a tailor-made MCMC scheme. In all cases, we get a nice reconstruction of the tree. In simulation studies, the posterior samples are much closer to the true tree than any of the competitors.
comment in response to
post
The underlying tree might be a fully resolved phylogeny, or it might be a simple hierarchical structure, such as a small number of clusters.
comment in response to
post
We observe several adjacency matrices - say, multiple measurements, or networks defined by different features - with the same underlying tree.
comment in response to
post
This is our framework: assume that nodes have evolved along a tree - formally, nodes have a latent position, which comes from a Brownian motion along a tree. Nodes which diverged recently are more likely to be linked by an edge (their latent positions are closer).
comment in response to
post
Many models of networks exist. You might know the Stochastic Block Model (SBM) or Latent Space Models.
These models help us discover underlying structure between the network nodes (eg discover cluster of nodes); they work for rather simple structures.
What if the underlying structure is a tree?
comment in response to
post
I should add: this parameter is NOT to optimize the log-likelihood instead of the likelihood of the stochastic matrix P.
Rather, it optimizes with argument log(P) instead of P; this trick makes the algorithm more stable.
comment in response to
post
I spent a couple of hours struggling because of this today: with logscale=FALSE, fitMk thought it had converged but really hadn't, and gave nonsensical results.
Will always use logscale=T from now on.
comment in response to
post
Je trouve des textes qui réglementent les thèses trop longues, mais rien qui interdise de soutenir sa thèse en avance.
comment in response to
post
Tu peux en trouver d'autres en choisissant des dates d'inscription et de soutenance rapprochées dans l'URL theses.fr/resultats?q=...
comment in response to
post
Sidy Fall, thèse de géographie à Nantes d'octobre 2021 à avril 2022 (!) sur "Le processus de privatisation des espaces et ressources maritimes au Sénégal"
theses.fr/s389306
comment in response to
post
Victor Viquesnel, thèse d'archéologie à Toulouse 2 de novembre 2021 à novembre 2023 sur "Dynamiques économique et culturelles en Gaule du Nord : les céramiques de Briga durant l'antiquité"
theses.fr/s171257
comment in response to
post
Par exemple :
Damien Guedon, thèse d'histoire de la musique à l'EPHE d'août 2021 à décembre 2023 sur "La vie musicale à Épinal entre le 1er août 1873 et le 2 août 1914 : l'exemple du théâtre lyrique."
theses.fr/s304129
comment in response to
post
Une recherche sur theses.fr donne plusieurs exemples de doctorats "courts".
Les données CSV contiennent la date de soutenance mais malheureusement pas la date d'inscription, mais on peut la trouver sur le site web. Il faut éditer l'URL de recherche à la main, l'interface de recherche fonctionne mal.
comment in response to
post
Bravo !
comment in response to
post
They are not independent.
(And yet something is going on, but I can't put my finger on it for now.)
comment in response to
post
My question wasn't framed well. Here is a counter-example to what I had in mind:
Consider (Zᵢ) iid from a (1/2, 1/2) mixture of a Dirac(1) and a Uniform(0,2). Take Xₙ to be the sample median (converges to 1 fast) and Yₙ to be the 10% quantile (converges to 0.4 slower).
comment in response to
post
Thanks Matti! You're right, I'm missing an assumption. I'll be back with a clearer statement.
comment in response to
post
Thinking about it some more: they won't necessarily be asyptotically independent, but it does seem that the dependence vanishes in some weaker sense.