Summary: we need an energy-aware theory of computation, and rate-distortion theory with Poisson latents is a good start.
But this is only the beginning, so please reach out if this sparked a thought/idea!
Here's the preprint again:
📜 arXiv: arxiv.org/abs/2602.13421
[11/11] 🧵
🧠🤖🧠📈
The RDT framing: we can take this theory and apply it anything that can be expressed as:
🔹optimization of a loss function
🔹approximating conditional distributions with some variational q
=> an energy-aware approach to modeling perception, cognition, and action?
[10/11] 🧵
🧠🤖🧠📈
We can also motivate these results within Rate-Distortion Theory (RDT).
In RDT, you optimize a generic loss function subject to coding rate constraints.
Our contribution is that if you use Poisson latents, then:
💡 coding rate budget ~ energy budget
[9/11] 🧵
We tested this empirically in a beta-VAE setting, and found that changing the reconstruction/KL trade-off results in a systematic different in the sparsity and metabolic cost of inference.
We saw this only in Poisson but not in Gaussian VAEs (~100% of Fristonian active inference models).
[8/11] 🧵
Conclusions so far:
1⃣ Poisson KL --> firing rate (metabolic cost)
2⃣ KL in general --> information rate (coding cost)
1⃣ + 2⃣ => Poisson KL couples an abstract information theoretic concept (coding rate) to a concrete biophysical quantity (firing rate)
[7/11] 🧵
Now let's try to interpret this result.
First, we need to let go of a narrow way of thinking about KL divergence.
In ML, most people think KL = regularization. But it's much deeper than that:
KL = 'unique' measure of info gain (Hobson 1969: link.springer.com/article/10.1...)
[6/11] 🧵
What causes sparsity of representations in Poisson VAEs? The answer lies in the Poisson KL term (coding rate) in the free energy objective.
Poisson KL becomes proportional to the prior rate variable, encouraging lower neural firing rates (= lower energy consumption).
[5/11] 🧵
In both ML and theoretical neuroscience, 99% of the time people default to Gaussian latents.
Previously, we showed that swapping Gaussian with Poisson results in a brain-like (spiking) generative model that reproduces sparse coding as a special case.
openreview.net/forum?id=ekt...
[4/11] 🧵
Free energy minimization is interesting, because it unifies various theories in neuroscience with machine learning models 👇
("Control/RL as inference" is also free energy minimization)
Different distributional and optimization choices yield different architectures.
[3/11] 🧵
📜 arXiv: arxiv.org/abs/2602.13421
Our starting point is the Variational Free Energy equation, which is just negative ELBO (F = -ELBO)
[2/11] 🧵
The "decoupling of information and energy" is a major point of divergence between biological and artificial computers.
Energy consumption is the biggest bottleneck in scaling AI.
To address this, we need an "energy-aware theory of computation." This work is an attempt to that end.
[1/11] 🧵
🧠🤖🧠📈
"Center for Advanced Hindsight" lmao south park predicted this in back in 2010
Enough #Bayesplaining. It's time to start "deriving."
We're unifying physics and neuroscience from first principles (yes, really).
Come argue with me tomorrow @ 11 AM.
@neuripsconf.bsky.social
🧠🤖🧠📈
Here's the poster info again:
📅 Wed, Dec 3, 11 AM — 2 PM
📍 Exhibit Hall C,D,E #500
🔗 neurips.cc/virtual/2025...
See you in sunny San Diego! 🌞
🧵[4/4]
🧠🤖🧠📈
✅ What's new in the paper:
We substantially improved the framing and clarity of our contributions (genuinely grateful to our amazing reviewers and their useful feedback 🙌).
+ Several new experiments (highlighted below), including scaling iP-VAE to complex color image datasets like CelebA.
🧵[3/4]
Final versions of paper & code:
📜 Paper: openreview.net/forum?id=573...
💻 Code: github.com/hadivafaii/I...
✅ What's new in the code:
We added a stand-alone Colab notebook implementation — a great starting point if you want to understand the iP-VAE model and train your own.
🧵[2/4]
I will be at @neuripsconf.bsky.social Dec 2-8 to present our "Brain-like Variational Inference" paper 👇
Let's connect if you enjoy first-principles thinking and brain-inspired AI 🧠 🤖
Poster info:
📅 Wed, Dec 3, 11 AM — 2 PM
📍Exhibit Hall C,D,E #500
🔗 neurips.cc/virtual/2025...
🧵[1/4]
🧠🤖🧠📈
This one is an all time favorite of mine:
direct.mit.edu/neco/article...
(tho maybe not that relevant to the "rescue" thing)
Thanks @hadivafaii.bsky.social for the invitation! x.com/hadivafaii/s...
🔗 Meeting summary: sensorimotorai.github.io/2025/11/20/d...
📜 Read the paper: www.pnas.org/doi/10.1073/...
💬 Join the conversation on Slack: join.slack.com/t/sensorimot...
🧠🤖🧠📈
💡 The killer insight: biological "quirks" are actually optimal control solutions:
🔷 STDP = solving the LQR cost (learning causal impact from feedback)
🔷 Noise = stabilizing "persistence of excitation"
🔷 ReLU = optimal policy to cross unstable fixed points
Conclusion:
✅ Agency is cellular.
🧠🤖🧠📈
Stop treating neurons as passive processors... They are active controllers! 🧠 🎮
In our latest Journal Club session, Thelonious Cooper presented "The neuron as a direct data-driven controller" by Moore et al., with special commentary from Mitya Chklovskii.
🎥 www.youtube.com/watch?v=0P7k...
🧠🤖🧠📈
NeurIPS 2025:
bsky.app/profile/hadi...
Additional links:
📝 Meeting summary: sensorimotorai.github.io/2025/11/13/a...
📅 Future meetings: sensorimotorai.github.io/schedule/
💬 Join the conversation on Slack: join.slack.com/t/sensorimot...
🧠🤖🧠📈
RL Debates 5: Anne "not everything is RL" Collins
Anne delivered an amazing synthesis of her extensive work on how working memory shapes reward-based learning in humans.
📜 Read the paper: nature.com/articles/s41...
📽️ Watch the full presentation: www.youtube.com/watch?v=eEqZ...
🧠🤖🧠📈
Additional links:
📝 Meeting summary: sensorimotorai.github.io/2025/novembe...
📅 Future meetings: sensorimotorai.github.io/schedule/
💬 Join the conversation on Slack: join.slack.com/t/sensorimot...
🧠🤖🧠📈
RL Debates 4: Adam "I literally measured value in the brain" Lowet
Adam's talk covered a lot of ground — from his recent work on distributional RL (nature.com/articles/s41...) to a broader discussion of RL & the brain.
📽️ Watch the full meeting here: www.youtube.com/watch?v=Xe7B...
🧠🤖🧠📈
Cover of the book “the brain abstracted” by M Chirimuuta
A passage from chapter 1 including the quote
Picked this one up for reading over the holidays.
“The thesis of this book is that the dominant ideas that have shaped #neuroscience are best understood as attempts to simplify the brain.” 🧠
…which is itself a simplification 😂
