Conor Heins
@conorheins
Followers
1K
Following
3K
Statuses
758
Working on probabilistic and bayesian machine learning at @VERSESAI. Interested in probabilistic ML, collective behavior, multi-agent systems, active inference
Berlin, Germany
Joined January 2014
1⃣ Excited to share new research from the Machine Learning Foundations Lab at @VERSESAI! "Gradient-free variational learning with conditional mixture networks" 📰Paper: 💻Code:
1
4
31
o3-mini-high took 19 seconds to write a .py file to simulate the Vicsek model and animate the results. Not surprising given all the other results being shared on X with bouncing ball examples + more complex physics, but still cool :)
0
0
6
RT @PNASNews: Researchers challenged longhorn crazy ants and humans with the same task: maneuvering a T-shaped object through two consecuti…
0
605
0
RT @CouckeNicolas: How can agents with simple bio-inspired brains make decisions together? This is what we simulate in our new arXiv prepri…
0
9
0
RT @helloVERSES: Agents and Robots and Inference oh my! VERSES researchers are boots on the ground this week at @NeurIPSConf, the largest…
0
14
0
RT @MarinaVPap: Excited (& a bit terrified) to officially share our #R #package with the world 🤩 swaRmverse provides a pipeline to analyze…
0
82
0
RT @icouzin: Delighted and honored to have been selected for the 2024 @voxdotcom #FuturePerfect50 list! @maxplanckp…
0
10
0
RT @neuroprinciples: 1/14 Happy to share our brief and sweet results on "signatures of criticality in efficient coding networks" now in @PN…
0
61
0
@harvie_zhang so our per-update cost is higher, but we converge much faster (in fewer) steps and the updates are local to layers (no need to store gradients through the compute graph)
0
0
0
RT @lancelotdacosta: Thrilled to share a theory of generalised coordinates for stochastic differential equations 🥳…
0
17
0
@harvie_zhang Yes exactly - you could put the same sorts of priors we use in the paper (matrix normal wisharts) over the linear layers' parameters and then do closed-form conjugate updates to them :)
2
0
0
@harvie_zhang ...their parameters that boils down to summing together the natural parameters of the (conjugate) prior over each experts parameters, and the sufficient statistics of the linear layer's inputs/outputs
1
0
0
5⃣ Takeaways CAVI-CMN's speed and ability to quantify uncertainty make it ideal for high-stakes, real-time applications where data is limited and noisy. Future work will explore scaling CAVI to bigger, deeper networks and more complex datasets.
0
0
3
RT @icouzin: We developed ‘dynamical social clamp’ (akin to ‘patch clamp’ in neuro)—using VR—allowing realtime bidirectional feedback betwe…
0
35
0
RT @iwai_ws: 💢5th International Conference on #ActiveInference (IWAI2024) 📷 ⏰When: September 9-11, 2024 Where: Oxfo…
0
11
0