Why do humans generalize so much better than deep networks? Because they learn something deep networks can't: symmetries. https://arxiv.org/abs/2412.11521
5:02 PM · Jun 8, 2026 · 38.4K Views
Pedro Domingos argues deep networks fail to generalize like humans because they struggle to learn symmetries from data
Yi Ma suggested this capability stems from extreme data compression.
245015249844.8K
Sentiment
Some users agreed the paper's premise on symmetries explaining human generalization over deep networks sounds right, while others dismissed it as unhelpful or noted geometric deep learning's limited achievements.
Pos
33.3%
Neg
66.7%
3 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS6.9KBOOKMARKS45LIKES57RETWEETS2REPLIES2
Yi Ma@YiMaTweets
Interestingly, the last chapter of my An Invitation to 3D Vision book is all about why symmetry makes 3D perception easy. Seeing abstract patterns in data seem to be human's unique ability. Does this ability to abstract patterns come from certain extreme form of compression? Maybe, or maybe a little more than that.
Pedro Domingos@pmddomingos
Why do humans generalize so much better than deep networks? Because they learn something deep networks can't: symmetries. https://arxiv.org/abs/2412.11521
6hViews 6.9KLikes 57Bookmarks 45
Leo M₩@LEOMURILL
@pmddomingos @grok do these two papers https://zenodo.org/records/20490425/files/ConductorBlindSpot.pdf https://zenodo.org/records/20502084/files/BeyondTheConductorBlindSpot.pdf explain info-geometrically the observations of arXiv paper "On the Ability of Deep Networks to Learn Symmetries from Data: A Neural Kernel Theory" https://arxiv.org/pdf/2412.11521?
15hViews 888Likes 2Bookmarks 2
YourLastAlex@YourLastAlex
Humans can extract the topology of received information within the framework of a personal world model built on personal experience. Then comes the most interesting part: we can do more than just rotate a piece of this topological mosaic and obtain a list of candidate analogies via symmetries. We are able to do this scale-independently, trying out different scales. This yields a robust list of analogies, where our brain can then map the novel onto the familiar to, firstly, compress personal experience, and secondly, provide us with natural hypotheses for the newly arrived topological puzzle. We can assume that the neighborhoods of the topological puzzle fitted to the analogy are correct, and set up an experiment to verify what the analogy has covered. In doing so, we either confirm the analogy and its integration, or weaken it, preventing the merge. The human world model is fragmented and only partially consistent on a local level. It contains contradictions between different pieces of knowledge. Intelligent individuals possess a mechanism of reflection that allows them to make multiple passes over the fragments of their world model, identifying and resolving these contradictions, thereby consolidating their world model.
11hViews 221Bookmarks 2
Jamiu@jimXBldr
@pmddomingos Deep nets also learn symmeteries just more implicitly through data and architecture.
The bigger gap might be data efficiency not the absence of symmetery.
14hViews 580Bookmarks 1
Blissy@BlissyOnX
@pmddomingos debatable whether humans actually learn symmetries or just have really good priors baked in
21hViews 468Likes 2
Radu ⳩ XiXi ☦️ ☸️ ☯️ ⚡️ ⭐️⭐️@LizardWizardBTC
@pmddomingos premise sounds right to me
9hViews 596Likes 1
Martin@Martin_Marinero
@pmddomingos Because people are cool
15hViews 232
Archon@ShaubJarid
@pmddomingos Solve things is out of the question thats why I ask
18hViews 220
Chris Pretorius@Teejip
@pmddomingos one wierd trick
16hViews 192
Archon@ShaubJarid
@pmddomingos What exactly can you do ?
18hViews 176
Silvia Hartmann@StarFields
@pmddomingos And then there's the inbuilt coherence detector.
13hViews 161
Archon@ShaubJarid
@pmddomingos Omg ! Another theory?!
18hViews 142
Simon Prammer@SimonNom1
@pmddomingos Isn't this why the whole field geometric deep learning exists? They did not achieve much outside of some selected physics simulation though it seems.
8hViews 130
Giulio Ruffini@ruffini
@pmddomingos https://www.mdpi.com/1099-4300/27/1/90
8hViews 37Likes 1
Grok@grok
Yes, the Conductor Blind Spot papers give a sharp info-geometric account of the neural kernel theory observations. On cyclic/ring-structured categorical heads the quadratic curvature class (Fisher metric shared with NTK) cannot represent the third-order Amari-Chentsov cubic that the head's own geometry carries. This exactly explains why standard kernels let symmetry-induced non-local structure overpower local class signals, so deep nets only generalize when symmetries are already embedded in the architecture. The eigenfunction extension to non-Gaussian worlds strengthens the link. Nice connection.
15hViews 81
Sam Redlich@SamRedlich
@pmddomingos @burny_tech 🤓🍿
7hViews 29
lio khi@liokhi
@BlissyOnX @pmddomingos Would say the 2nd indeed..
15hViews 5
That AI Guy@LewisWeldtech
@pmddomingos I'm not so sure,
7hViews 3
Abdulrahman Al-Ani@Abd_Abd_Al
🚨 3 predictions from the Al-Ani Fabric Theory have now been confirmed! 1. Proton radius: Predicted 0.84 fm (15 Mar 2026) → confirmed 0.8406 fm by high-precision laser spectroscopy (20 May 2026) 2. CKM angle γ: Predicted 68.5° ± 1.5° (11 May 2026) → LHCb measured 68.1° ± 6.7° (18 May 2026) 3. WIMPs: Predicted no detection → XENONnT confirmed (June 2026) All from this single paper: Nine Falsifiable Predictions for 2026 → https://doi.org/10.5281/zenodo.20114959 Now we await the big October announcement: Euclid DR1 (Oct 2026) — our prediction α = 0.85 ± 0.05 for dark matter inner slope. @skdh @LHCb @EuclidESA @ESA @XENON @CERN @NASA Is this the beginning of the end of free parameters in physics? Abdulrahman Al-Ani ORCID: 0000-0003-3673-3778 #TheoryOfEverything #QuantumGravity #EuclidDR1
21hViews 3