lingo.lol is one of the many independent Mastodon servers you can use to participate in the fediverse.
A place for linguists, philologists, and other lovers of languages.

Server stats:

66
active users

#MachineLearning

18 posts17 participants2 posts today

I just discovered the ARC-AGI initiative and the associated test to estimate how close "AI" models are from #AGI

arcprize.org/arc-agi

While I found the initiative interesting, I'm not sure I understand what in this test really guarantees that the model is capable of some form of generalization and problem-solving.
Wouldn't it be possible for specialized pattern-matching/discovering algorithms to solve such problems?
I imagine some computer scientists, mathematicians or computational neuroscientists have already had a look at this, so would anyone knows of some articles/blogs on the topic?

Maybe @wim_v12e? Is this something you already looked at?

ARC PrizeARC Prize - What is ARC-AGI?The only AI benchmark that measures AGI progress.

The ML\LLM AI world is overcrowded✅ and it's hard to find majestic new projects❌
Here is (updating) list of thoughts, premature ideas, and research directions that I cannot pursue alone, but I believe can make radical changes
Please fight me, discuss, and ask
Disrrruption, aye?🏴‍☠️
#AI #MachineLearning #nlp #data

Now out in #TMLR:

🍇 GRAPES: Learning to Sample Graphs for Scalable Graph Neural Networks 🍇

There's lots of work on sampling subgraphs for GNNs, but relatively little on making this sampling process _adaptive_. That is, learning to select the data from the graph that is relevant for your task.

We introduce an RL-based and a GFLowNet-based sampler and show that the approach performs well on heterophilic graphs.

openreview.net/forum?id=QI0l84

I am transmitting a hiring opportunity for a good friend (@jblugagne). I have had the pleasure to know him and work with him for a long time, and warmly encourage interested people to apply. I have worked and am still working a bit on DeLTA myself, so I am biased but I think it's a very nice project :) Boosts appreciated!

> We are looking for a Senior Python Developer to join our group at the University of Oxford’s Department of Engineering Science. This is a full-time position focused on advancing our open-source computer vision software for quantitative microscopy, DeLTA. There will also be opportunities to explore commercial applications and contribute to potential spin-off efforts.

> We’re looking for someone with strong Python skills and experience in software release and management. Backgrounds in computer vision, machine learning, or microscopy are a plus.

> eng.ox.ac.uk/jobs/job-detail/?

eng.ox.ac.ukJob Detail

🖥️ **How Overconfidence in Initial Choices and Underconfidence Under Criticism Modulate Change of Mind in Large Language Models**

🔗 doi.org/10.48550/arXiv.2507.03.

arXiv logo
arXiv.orgHow Overconfidence in Initial Choices and Underconfidence Under Criticism Modulate Change of Mind in Large Language ModelsLarge language models (LLMs) exhibit strikingly conflicting behaviors: they can appear steadfastly overconfident in their initial answers whilst at the same time being prone to excessive doubt when challenged. To investigate this apparent paradox, we developed a novel experimental paradigm, exploiting the unique ability to obtain confidence estimates from LLMs without creating memory of their initial judgments -- something impossible in human participants. We show that LLMs -- Gemma 3, GPT4o and o1-preview -- exhibit a pronounced choice-supportive bias that reinforces and boosts their estimate of confidence in their answer, resulting in a marked resistance to change their mind. We further demonstrate that LLMs markedly overweight inconsistent compared to consistent advice, in a fashion that deviates qualitatively from normative Bayesian updating. Finally, we demonstrate that these two mechanisms -- a drive to maintain consistency with prior commitments and hypersensitivity to contradictory feedback -- parsimoniously capture LLM behavior in a different domain. Together, these findings furnish a mechanistic account of LLM confidence that explains both their stubbornness and excessive sensitivity to criticism.

💡 Wie lassen sich Methoden des Machine Learning sinnvoll in der Sammlungsdigitalisierung & -forschung einsetzen?

Im Interview spricht unser Kollege @mathias_zinnen über:
🔍 ML-Methoden für Text, Bild & strukturierte Daten
🛠️ Open-Source-Tools
💡 Vermittlungsangebote zu ML-Tools im Sammlungskontext
🌱 Warum ressourcenschonende ML-Ansätze wichtig sind

➡️ sammlungen.io/blog/interview-m
#MachineLearning #2D #SODaZentrum #OCR #ML

In the ISE2025 lecture today, our students learned about unsupervised learning on the example of k-Means clustering. One nice hands-on example is image colour reduction based on k-means clustering, as demonstrated in a colab notebook (based on the Python DataScience Handbook by Vanderplus)

colab notebook: colab.research.google.com/driv
Python DataScience Handbook: archive.org/details/python-dat

#ise2025 #lecture @fizise @sourisnumerique @enorouzi #datascience #machinelearning #AI @KIT_Karlsruhe

Summer ☀️ read: a new paper on model-based clustering just appeared in Computo!

Julien Jacques and Brendan Thomas Murphy publish a new method for clustering multivariate count data. The method combines feature selection and clustering, and is based on conditionally independent Poisson mixture models and Poisson generalized linear models.

On simulations, the Adjusted Rand Index (ARI) of the model with selected variables is close to the optimal ARI obtained with the true clustering variables.

The paper and accompanying R code are available at computo-journal.org/published-

Summer ☀️ read on Computo: a new publication on reservoir computing in R!

Reservoir computing is a machine learning approach that relies on mapping inputs to higher dimensional spaces through a non-linear dynamical system (the reservoir), for example using a deep recurrent neural network and training only its final layer.

In this new publication, Thomas Ferté and co-authors Kalidou Ba, Dan Dutartre, Pierrick Legrand, Vianney Jouhet, Rodolphe Thiébaut, Xavier Hinaut and Boris P. Hejblum present the reservoirnet package, which is the first implementation of reservoir computing in R (rather then the existing Python and Julia).

The article also serves as an introduction to reservoir computing and illustrates its usefulness as well as the usage of the package on several real-case applications, including forecasting the number of COVID-19 hospitalizations at Bordeaux University Hospital using public data (epidemiological statistics, weather data) and hospital-level data (hospitalizations, ICU admission, ambulance service and ER notes). On this example, the authors also show how the weights of the connection between the input and the output layers can be used to compute feature importances.

The paper and accompanying R code are available at doi.org/10.57750/arxn-6z34

reservoirnet is available at cran.r-project.org/package=res

Continued thread

I also got experience with the following (5 = a lot, 1 = a little) :

#machinelearning #ml (3) (I have implemented some ML models myself in the past, for learning purposes.)
#guix (3) (Using it for reproducible setups of projects.)
#functionalprogramming #fp (5) (Doing it in my own projects.)
#objectorientedprogramming #oop (4) (last job and past xp in my own projects.)
#CI / #CD (3) (Last job)
#make (4) (using it for my own project setups and convenience)
#testing (4) (last job, own projects)

New #TeachingMaterial available: Functional Imaging Data Analysis – From Calcium Imaging to Network Dynamics. This course covers the entire workflow from raw #imaging data to functional insights, including #SpikeInference & #PopulationAnalysis. Designed for students and for self-guided learning, with a focus on open content and reproducibility. Feel free to use and share it 🤗

🌍 fabriziomusacchio.com/blog/202

Replied in thread

@alex_p_roe Speaking of fruit fly research, you'd be amused or surprised to learn that the original U-net architecture (which today powers stable diffusion, among many other machine learning techniques) introduced in a paper by Ronneberger et al. (2015; arxiv.org/abs/1505.04597 ) was developed to perform image segmentation of fly neural tissue as imaged with electron microscopy, to reconstruct neurons and therefore map the brain connectome.

So all those "wasteful" research funding grants to fruit fly research motivated and led to the biggest discovery fueling the whole of the modern "AI" boom. One never knows where basic research will lead, it's impossible to predict. Hence basic research is not at all wasteful, on the contrary, it's essential, it's the foundation of a rich, wealthy, creative society. And also very cheap, comparatively: albert.rierol.net/tell/2016060

Search also for the returns on the human genome project, or on the humble origins of DNA sequencing, to name just two among many.

arXiv logo
arXiv.orgU-Net: Convolutional Networks for Biomedical Image SegmentationThere is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .
Continued thread

…Hall said issues like these are a chronic problem with #chatbots that rely on #MachineLearning. In 2016, #Microsoft released an #AI #chatbot named Tay on Twitter. Less than 24 hours after its release, Twitter users baited Tay into saying #racist & #antisemitic statements, including praising #Hitler. Microsoft took the chatbot down & apologized.

Tay, #Grok & other #AI chatbots with live access to the internet seemed to be incorporating real-time information, which Hall said carries more risk.

Continued thread

If we ever see a real artificial mind, some kind of LLM will probably be a small but significant component of that, but the current wave of machine learning will most likely come to a grinding halt very soon because of a lack of cheap training data. The reason why all of this is happening now is simple: The technologies behind machine learning have been around for decades, but computers weren't fast enough and didn't have enough memory for those tools to become really powerful until the early 2000s, and around the same time, the Internet went mainstream and got filled with all kinds of data that could be datamined for training sets. Now there is so much synthetic content out there that automated data mining won't work much longer, you need humans to curate and clean the training data, which makes the process slow and expensive. I expect to see another decades long AI winter after the commercial hype is over.

If you look for real intelligence, look at autonomous robots and computer game NPCs. There you can find machine learning and artificial neural networks applied to actual cognitive tasks in which an agent interacts with its environment. Those things may not even be as intelligent as a rat yet, but they are actually intelligent, unlike LLMs.

#llm#LLMs#ai