Twitxy Saved

Gennady Roshchupkin

Verified account This account is verified. Learn more

@Roshchupkin

Assistant Professor | Computational Population Biology group leader | Erasmus MC | AI | Data Science | Genomics | Epidemiology | Medical Imaging | UMCNL
Rotterdam, The Netherlands roshchupkin.org Joined May 2010
847 Following 406 Followers 633 Posts

Latest posts 5

Pinned post

Gennady Roshchupkin Verified account @Roshchupkin
๐Ÿš€ New Podcast Launch ๐Ÿš€

๐ŸŽ™๏ธ The Lifelong Study: Insights From Rotterdam โ€“ Created using AI to explain complex scientific research in simple terms! ๐ŸŽง
open.spotify.com/show/77krIekT1โ€ฆ
This podcast breaks down cutting-edge discoveries from the Rotterdam Study, making the latest health research accessible to everyone.

๐ŸŒ Episode 1: Rotterdam Study Updated 2024 Get the latest insights from this long-term study that tracks health risks across a population, offering insights into diseases and the effects of lifestyle and environment on aging.

๐Ÿค– Episode 2: Making Sense of the Gene Code: Unlocking the Secrets of Deep Learning in Genomics Explore how AI and deep learning are revolutionizing genomics, uncovering hidden patterns in our DNA while addressing the challenge of interpretability.

๐Ÿงฌ Episode 3: Genetics Unpacked: Causality Through Mendelian Randomization Learn how researchers use genetic differences to uncover the causes of diseases like high blood pressure and diabetes through Mendelian Randomization

๐Ÿ” Episode 4: Hidden Harbingers: Uncovering the Secrets of Arterial Calcification
Discover how arterial calcification develops over 14 years, its links to cardiovascular risks, and why it matters for your health.
#AI #Genomics #RotterdamStudy #Healthcare #DeepLearning #Aging #MendelianRandomization
open.spotify.com The Lifelong Study: Insights From Rotterdam Podcast ยท roshchupkin ยท The longitudinal nature of the Rotterdam Study and the valuable insights gained from tracking a population over an extended period. www.roshchupkin.org/podcast
Gennady Roshchupkin Verified account @Roshchupkin
Google map UI deserve a special space in hell... You search something and open app in a minute, everything is gone. You search something and open in a hour to check something new, the old location is still there and you need to click five times to have a default screen...
Gennady Roshchupkin Verified account @Roshchupkin
It is the weekend, so why not think about something slightly more philosophical than benchmarks, agents and neural networks. ๐Ÿ™‚

Recently I have been thinking about a different analogy.

Suppose, just for the sake of argument, that ๐ข๐ง๐ญ๐ž๐ฅ๐ฅ๐ข๐ ๐ž๐ง๐œ๐ž ๐จ๐ซ ๐œ๐จ๐ง๐ฌ๐œ๐ข๐จ๐ฎ๐ฌ๐ง๐ž๐ฌ๐ฌ ๐ข๐ฌ ๐ง๐จ๐ญ ๐ฌ๐จ๐ฆ๐ž๐ญ๐ก๐ข๐ง๐  ๐œ๐จ๐ฆ๐ฉ๐ฅ๐ž๐ญ๐ž๐ฅ๐ฒ ๐š๐œ๐œ๐ข๐๐ž๐ง๐ญ๐š๐ฅ, but somehow a more fundamental property of the universe.

There are philosophical ideas in this general direction: that under the right conditions, consciousness or intelligence may emerge naturally, perhaps inevitably.

Now look at where AI might be heading.

Today we train huge models at enormous computational cost.

But perhaps this eventually converges toward something like a very large, general intelligence infrastructure: one enormously capable model, or a relatively small number of them, available almost as a utility.

Then what runs on our phones, laptops, robots and other devices would not necessarily be a complete intelligence of its own.

It might instead be something closer to what we currently call an ๐š๐ ๐ž๐ง๐ญ:

a local context,
a memory,
a collection of skills,
a set of goals,
and a particular way of prompting or interacting with a much larger underlying intelligence.

In that world, the intelligence is global.

The individual agent is mostly the interface through which it is expressed.

And that made me wonder about us.

What if biological intelligence worked in an analogous way?

What if each of us is not carrying an entirely separate โ€œmodelโ€ of intelligence inside our skull, but is somehow a local expression, receiver, or inference process of something more fundamental?

Our brains would then provide the equivalent of:

local memory,
context,
goals,
skills,
sensory input,
and perhaps a very complicated biological prompt.

The underlying intelligence would be something more universal.

Of course, the opposite possibility is equally interesting.

Perhaps AI will develop in the other direction.

Models may become so efficient, computation so cheap, and hardware so powerful that every device will eventually contain its own completely independent superintelligent system.

No central intelligence required.

Every phone, robot or computer would carry the whole model.

And if that becomes the dominant architecture, the analogy with biological intelligence changes completely.

Maybe that is also what we are:

billions of genuinely independent intelligent systems,
each carrying its own enormously complex model,
with no global intelligence behind us at all.

So there are two very different pictures.

๐Ž๐ง๐ž ๐ ๐ฅ๐จ๐›๐š๐ฅ ๐ข๐ง๐ญ๐ž๐ฅ๐ฅ๐ข๐ ๐ž๐ง๐œ๐ž, expressed through many local agents.

Or ๐›๐ข๐ฅ๐ฅ๐ข๐จ๐ง๐ฌ ๐จ๐Ÿ ๐ข๐ง๐๐ž๐ฉ๐ž๐ง๐๐ž๐ง๐ญ ๐ข๐ง๐ญ๐ž๐ฅ๐ฅ๐ข๐ ๐ž๐ง๐œ๐ž๐ฌ, each carrying the whole machinery locally.

What I find fascinating is that AI may give us a new language for thinking about a very old question.

Not because large language models explain consciousness โ€” they certainly do not.

But because we are now actively engineering different ways in which intelligence can be distributed:

centralized versus local,
shared versus independent,
general versus specialized,
model versus agent,
global knowledge versus individual memory.

And perhaps the architecture that eventually works best for artificial intelligence will at least give us another analogy for thinking about our own.

So here is my weekend question:

๐€๐ซ๐ž ๐ฐ๐ž ๐ž๐š๐œ๐ก ๐œ๐š๐ซ๐ซ๐ฒ๐ข๐ง๐  ๐จ๐ฎ๐ซ ๐จ๐ฐ๐ง โ€œ๐ฆ๐จ๐๐ž๐ฅโ€ ๐จ๐Ÿ ๐ข๐ง๐ญ๐ž๐ฅ๐ฅ๐ข๐ ๐ž๐ง๐œ๐ž?

Or are we, in some sense, local agents through which something much more universal is expressed?

No idea.

But perhaps a nice thing to think about over the weekend. ๐Ÿ™‚
Gennady Roshchupkin Verified account @Roshchupkin
For our Advanced Data Science and Machine Learning for Health Research summer course, weโ€™ve been developing a set of practical exercises that Iโ€™m really happy with.

The goal was to find a good balance between teaching genuinely advanced machine learning methods and keeping everything realistic within the limited time and computing capacity you normally have in a summer course.

The practicals cover things like:

๐Ÿง  Convolutional Neural Networks
๐Ÿ” Explainable AI
๐Ÿ“‰ Autoencoders and latent spaces
What I like about these exercises is that students donโ€™t just hear about these methods in lectures.
They actually train models, inspect what the networks are learning, visualize intermediate representations, explore latent spaces, and get a much more intuitive feeling for what is going on inside these architectures.

Of course, nobody becomes a deep learning expert in a few days.

But I think there is a big difference between knowing that โ€œCNNs are used for imagesโ€ or โ€œautoencoders learn latent representationsโ€ and actually building one, changing things, breaking things, visualizing the results, and seeing for yourself what happens.

And given the relatively modest computing requirements, I think this is a very nice win-win: you can explore fairly advanced methods without needing a huge GPU cluster.

We made all the practical material openly available on GitHub:

github.com/roshchupkin/Adโ€ฆ

Please use it, copy it, adapt it for your own courses, share it with students, or just play around with it yourself.

Iโ€™m very happy if these exercises find their way into other courses as well.

Teaching material becomes much more useful when it is shared and reused.

#MachineLearning #DeepLearning #AI #MedicalAI #DataScience #Education #OpenScience
Gennady Roshchupkin Verified account @Roshchupkin
For my machine learning courses, Iโ€™ve been gradually building an interactive educational portal to make some of these concepts easier to understand by actually playing with them.

One of the new pages Iโ€™m particularly proud of is this Autoencoder Explainer.

You can explore how an autoencoder compresses and reconstructs data, play with the latent space, move points around, and directly see how changes in the representation affect the reconstructed data.

I think these concepts become much more intuitive once you can interact with them rather than only looking at equations and architecture diagrams.

Try it here:
ml-health.roshchupkin.org/ae_explainer.hโ€ฆ
More interactive ML material is coming as I keep developing the course.

#MachineLearning #DeepLearning #AI #Education