Hadi Vafaii
@hadivafaii.bsky.social
Postdoc at UC Berkeley, Redwood Center | 🧠🤖 | 🎹 | 🎾 | https://mysterioustune.com/
I'm excited to announce that we've added two more contenders to the RL Debate Series!
- Anne Collins (Professor @ucberkeleyofficial.bsky.social)
- Niels Leadholm (Research Manager @thousandbrains.org)
Things are heating up!!
See the flyer for details👇
🌎 sensorimotorai.github.io/debates/
🧠🤖🧠📈
- Anne Collins (Professor @ucberkeleyofficial.bsky.social)
- Niels Leadholm (Research Manager @thousandbrains.org)
Things are heating up!!
See the flyer for details👇
🌎 sensorimotorai.github.io/debates/
🧠🤖🧠📈
October 3, 2025 at 6:19 AM
I'm excited to announce that we've added two more contenders to the RL Debate Series!
- Anne Collins (Professor @ucberkeleyofficial.bsky.social)
- Niels Leadholm (Research Manager @thousandbrains.org)
Things are heating up!!
See the flyer for details👇
🌎 sensorimotorai.github.io/debates/
🧠🤖🧠📈
- Anne Collins (Professor @ucberkeleyofficial.bsky.social)
- Niels Leadholm (Research Manager @thousandbrains.org)
Things are heating up!!
See the flyer for details👇
🌎 sensorimotorai.github.io/debates/
🧠🤖🧠📈
Not all 'active' learning is 'reinforcement' learning.
➡️ This suggests we need to upgrade—or even replace—RL with a more general theory of active learning that's not solely reward-based
✅Which is why we are launching the RL Debates
More info: sensorimotorai.github.io/debates/
🧵[4/5]
🧠🤖🧠📈
➡️ This suggests we need to upgrade—or even replace—RL with a more general theory of active learning that's not solely reward-based
✅Which is why we are launching the RL Debates
More info: sensorimotorai.github.io/debates/
🧵[4/5]
🧠🤖🧠📈
September 17, 2025 at 4:32 PM
Not all 'active' learning is 'reinforcement' learning.
➡️ This suggests we need to upgrade—or even replace—RL with a more general theory of active learning that's not solely reward-based
✅Which is why we are launching the RL Debates
More info: sensorimotorai.github.io/debates/
🧵[4/5]
🧠🤖🧠📈
➡️ This suggests we need to upgrade—or even replace—RL with a more general theory of active learning that's not solely reward-based
✅Which is why we are launching the RL Debates
More info: sensorimotorai.github.io/debates/
🧵[4/5]
🧠🤖🧠📈
For decades, reinforcement learning (RL) has been synonymous with "active learning".
This figure from Sutton & Barto's classic textbook summarizes the standard view👇
🤔 But what if the premise of an 'external, scalar reward' is a simplification we need to move past?
🧵[3/5]
This figure from Sutton & Barto's classic textbook summarizes the standard view👇
🤔 But what if the premise of an 'external, scalar reward' is a simplification we need to move past?
🧵[3/5]
September 17, 2025 at 4:32 PM
For decades, reinforcement learning (RL) has been synonymous with "active learning".
This figure from Sutton & Barto's classic textbook summarizes the standard view👇
🤔 But what if the premise of an 'external, scalar reward' is a simplification we need to move past?
🧵[3/5]
This figure from Sutton & Barto's classic textbook summarizes the standard view👇
🤔 But what if the premise of an 'external, scalar reward' is a simplification we need to move past?
🧵[3/5]
🌍 More info on our website: sensorimotorai.github.io/debates/
💬 Join us on Slack: join.slack.com/t/sensorimot...
💡 The main motivation: recognizing how central "action" is in all things intelligence
Perception, cognition, and knowledge are fundamentally intertwined with "action"
🧵[2/5]
💬 Join us on Slack: join.slack.com/t/sensorimot...
💡 The main motivation: recognizing how central "action" is in all things intelligence
Perception, cognition, and knowledge are fundamentally intertwined with "action"
🧵[2/5]
September 17, 2025 at 4:32 PM
🌍 More info on our website: sensorimotorai.github.io/debates/
💬 Join us on Slack: join.slack.com/t/sensorimot...
💡 The main motivation: recognizing how central "action" is in all things intelligence
Perception, cognition, and knowledge are fundamentally intertwined with "action"
🧵[2/5]
💬 Join us on Slack: join.slack.com/t/sensorimot...
💡 The main motivation: recognizing how central "action" is in all things intelligence
Perception, cognition, and knowledge are fundamentally intertwined with "action"
🧵[2/5]
What drives behavior in living organisms? And how can we design artificial agents that learn interactively?
📢 To address these, the Sensorimotor AI Journal Club is launching the "RL Debate Series"👇
w/ @elisennesh.bsky.social, @noreward4u.bsky.social, @tommasosalvatori.bsky.social
🧵[1/5]
🧠🤖🧠📈
📢 To address these, the Sensorimotor AI Journal Club is launching the "RL Debate Series"👇
w/ @elisennesh.bsky.social, @noreward4u.bsky.social, @tommasosalvatori.bsky.social
🧵[1/5]
🧠🤖🧠📈
September 17, 2025 at 4:32 PM
What drives behavior in living organisms? And how can we design artificial agents that learn interactively?
📢 To address these, the Sensorimotor AI Journal Club is launching the "RL Debate Series"👇
w/ @elisennesh.bsky.social, @noreward4u.bsky.social, @tommasosalvatori.bsky.social
🧵[1/5]
🧠🤖🧠📈
📢 To address these, the Sensorimotor AI Journal Club is launching the "RL Debate Series"👇
w/ @elisennesh.bsky.social, @noreward4u.bsky.social, @tommasosalvatori.bsky.social
🧵[1/5]
🧠🤖🧠📈
"2000s Dad Rock" is a thing 🙂
August 31, 2025 at 12:02 AM
"2000s Dad Rock" is a thing 🙂
Our first 'Sensorimotor AI Journal Club' meeting was a blast!
(...which, we might as well just call the "AI Club for Non-conformists"👇😎)
📽️ full presentation: youtube.com/watch?v=efc7...
🧵[1/4]
🧠🤖🧠📈
(...which, we might as well just call the "AI Club for Non-conformists"👇😎)
📽️ full presentation: youtube.com/watch?v=efc7...
🧵[1/4]
🧠🤖🧠📈
August 10, 2025 at 12:50 AM
Our first 'Sensorimotor AI Journal Club' meeting was a blast!
(...which, we might as well just call the "AI Club for Non-conformists"👇😎)
📽️ full presentation: youtube.com/watch?v=efc7...
🧵[1/4]
🧠🤖🧠📈
(...which, we might as well just call the "AI Club for Non-conformists"👇😎)
📽️ full presentation: youtube.com/watch?v=efc7...
🧵[1/4]
🧠🤖🧠📈
Announcing the new "Sensorimotor AI" Journal Club — please share/repost!
w/ Kaylene Stocking, Tommaso Salvatori, and @elisennesh.bsky.social
Sign up link: forms.gle/o5DXD4WMdhTg...
More details below 🧵[1/5]
🧠🤖🧠📈
w/ Kaylene Stocking, Tommaso Salvatori, and @elisennesh.bsky.social
Sign up link: forms.gle/o5DXD4WMdhTg...
More details below 🧵[1/5]
🧠🤖🧠📈
July 9, 2025 at 10:31 PM
Announcing the new "Sensorimotor AI" Journal Club — please share/repost!
w/ Kaylene Stocking, Tommaso Salvatori, and @elisennesh.bsky.social
Sign up link: forms.gle/o5DXD4WMdhTg...
More details below 🧵[1/5]
🧠🤖🧠📈
w/ Kaylene Stocking, Tommaso Salvatori, and @elisennesh.bsky.social
Sign up link: forms.gle/o5DXD4WMdhTg...
More details below 🧵[1/5]
🧠🤖🧠📈
truly ahead of his time
July 1, 2025 at 5:30 AM
truly ahead of his time
Stay tuned for Part 2, where I plan to:
1⃣ introduce Bayes' theorem, and provide a visual proof
2⃣ apply it to explain both optical illusions and social delusions
But first, you need to read Part 1 as a prerequisite 🙂 Here's the link again: mysterioustune.com/2025/06/29/p...
[5/6]🧵
🧠🤖🧠📈
1⃣ introduce Bayes' theorem, and provide a visual proof
2⃣ apply it to explain both optical illusions and social delusions
But first, you need to read Part 1 as a prerequisite 🙂 Here's the link again: mysterioustune.com/2025/06/29/p...
[5/6]🧵
🧠🤖🧠📈
June 30, 2025 at 5:30 AM
Stay tuned for Part 2, where I plan to:
1⃣ introduce Bayes' theorem, and provide a visual proof
2⃣ apply it to explain both optical illusions and social delusions
But first, you need to read Part 1 as a prerequisite 🙂 Here's the link again: mysterioustune.com/2025/06/29/p...
[5/6]🧵
🧠🤖🧠📈
1⃣ introduce Bayes' theorem, and provide a visual proof
2⃣ apply it to explain both optical illusions and social delusions
But first, you need to read Part 1 as a prerequisite 🙂 Here's the link again: mysterioustune.com/2025/06/29/p...
[5/6]🧵
🧠🤖🧠📈
💡 Main insight: Perception ≠ Reality
🧠 Plato wrote about this back in 380 BC
✅ But now, we have the mathematics of Bayesian posterior inference to reason about this concept
If you’ve also wondered about this, you’re in good company. At least a 2400+ year old company 😉
[4/6]🧵
🧠 Plato wrote about this back in 380 BC
✅ But now, we have the mathematics of Bayesian posterior inference to reason about this concept
If you’ve also wondered about this, you’re in good company. At least a 2400+ year old company 😉
[4/6]🧵
June 30, 2025 at 5:30 AM
💡 Main insight: Perception ≠ Reality
🧠 Plato wrote about this back in 380 BC
✅ But now, we have the mathematics of Bayesian posterior inference to reason about this concept
If you’ve also wondered about this, you’re in good company. At least a 2400+ year old company 😉
[4/6]🧵
🧠 Plato wrote about this back in 380 BC
✅ But now, we have the mathematics of Bayesian posterior inference to reason about this concept
If you’ve also wondered about this, you’re in good company. At least a 2400+ year old company 😉
[4/6]🧵
3D visual perception is a standard way of motivating why we need inference to interpret the world.
💡But did you know John Locke first thought of this example way back in 1690?
Read my blog to learn about the rich history behind this idea.
[3/6]🧵
💡But did you know John Locke first thought of this example way back in 1690?
Read my blog to learn about the rich history behind this idea.
[3/6]🧵
June 30, 2025 at 5:30 AM
3D visual perception is a standard way of motivating why we need inference to interpret the world.
💡But did you know John Locke first thought of this example way back in 1690?
Read my blog to learn about the rich history behind this idea.
[3/6]🧵
💡But did you know John Locke first thought of this example way back in 1690?
Read my blog to learn about the rich history behind this idea.
[3/6]🧵
“Perception as Inference” is a century-old idea that has inspired all major theories in neuroscience 🧠, including:
✅ Sparse Coding
✅ Predictive Coding
✅ Free Energy Principle
& more!
In my new blog post, I build the intuition behind this idea from ground up 👉[1/6]🧵
🧠🤖🧠📈
✅ Sparse Coding
✅ Predictive Coding
✅ Free Energy Principle
& more!
In my new blog post, I build the intuition behind this idea from ground up 👉[1/6]🧵
🧠🤖🧠📈
June 30, 2025 at 5:30 AM
“Perception as Inference” is a century-old idea that has inspired all major theories in neuroscience 🧠, including:
✅ Sparse Coding
✅ Predictive Coding
✅ Free Energy Principle
& more!
In my new blog post, I build the intuition behind this idea from ground up 👉[1/6]🧵
🧠🤖🧠📈
✅ Sparse Coding
✅ Predictive Coding
✅ Free Energy Principle
& more!
In my new blog post, I build the intuition behind this idea from ground up 👉[1/6]🧵
🧠🤖🧠📈
✅When unrolled in time, the iP-VAE inference algorithm looks like a deep, stochastic, spiking ResNet with parameter sharing.
✅The exp nonlinearity and emergent divisive normalization likely underlie iP-VAE's effectiveness — similar to xLSTM (openreview.net/forum?id=ARA...)
🧵[11/n]
✅The exp nonlinearity and emergent divisive normalization likely underlie iP-VAE's effectiveness — similar to xLSTM (openreview.net/forum?id=ARA...)
🧵[11/n]
May 19, 2025 at 6:36 AM
✅When unrolled in time, the iP-VAE inference algorithm looks like a deep, stochastic, spiking ResNet with parameter sharing.
✅The exp nonlinearity and emergent divisive normalization likely underlie iP-VAE's effectiveness — similar to xLSTM (openreview.net/forum?id=ARA...)
🧵[11/n]
✅The exp nonlinearity and emergent divisive normalization likely underlie iP-VAE's effectiveness — similar to xLSTM (openreview.net/forum?id=ARA...)
🧵[11/n]
iP-VAE is a spiking inference model, and it reproduces well-studied cortical response properties.
For example, contrast-dependent response latency of V1 neurons (Carandini et al., 1997):
(compare to Fig. 3A here: doi.org/10.1523/jneu...)
🧵[9/n]
For example, contrast-dependent response latency of V1 neurons (Carandini et al., 1997):
(compare to Fig. 3A here: doi.org/10.1523/jneu...)
🧵[9/n]
May 19, 2025 at 6:36 AM
iP-VAE is a spiking inference model, and it reproduces well-studied cortical response properties.
For example, contrast-dependent response latency of V1 neurons (Carandini et al., 1997):
(compare to Fig. 3A here: doi.org/10.1523/jneu...)
🧵[9/n]
For example, contrast-dependent response latency of V1 neurons (Carandini et al., 1997):
(compare to Fig. 3A here: doi.org/10.1523/jneu...)
🧵[9/n]
We did a massive hyperparameter sweep and found that:
✅iP-VAE and LCA (a classic sparse coding algorithm) find the best overall reconstruction-sparsity trade-off
✅All iterative VAEs outperform their standard amortized counterparts (despite using 25x fewer parameters)
🧵[8/n]
✅iP-VAE and LCA (a classic sparse coding algorithm) find the best overall reconstruction-sparsity trade-off
✅All iterative VAEs outperform their standard amortized counterparts (despite using 25x fewer parameters)
🧵[8/n]
May 19, 2025 at 6:36 AM
We did a massive hyperparameter sweep and found that:
✅iP-VAE and LCA (a classic sparse coding algorithm) find the best overall reconstruction-sparsity trade-off
✅All iterative VAEs outperform their standard amortized counterparts (despite using 25x fewer parameters)
🧵[8/n]
✅iP-VAE and LCA (a classic sparse coding algorithm) find the best overall reconstruction-sparsity trade-off
✅All iterative VAEs outperform their standard amortized counterparts (despite using 25x fewer parameters)
🧵[8/n]
We also compare to other models like a Gaussian version (closely related to predictive coding networks) etc.
✅All models converge beyond their training regime (T_train = 16, T_test = 1000)
✅iP-VAE finds the best compromise between reconstruction fidelity and sparsity
🧵[7/n]
✅All models converge beyond their training regime (T_train = 16, T_test = 1000)
✅iP-VAE finds the best compromise between reconstruction fidelity and sparsity
🧵[7/n]
May 19, 2025 at 6:36 AM
We also compare to other models like a Gaussian version (closely related to predictive coding networks) etc.
✅All models converge beyond their training regime (T_train = 16, T_test = 1000)
✅iP-VAE finds the best compromise between reconstruction fidelity and sparsity
🧵[7/n]
✅All models converge beyond their training regime (T_train = 16, T_test = 1000)
✅iP-VAE finds the best compromise between reconstruction fidelity and sparsity
🧵[7/n]
If you follow those prescriptions, you end up with this canonical circuit model that computational neuroscientists have been studying since 1970s.
Let me emphasize:
✅ We "derived" this. Just from F minimization. Without putting any of those terms in there by hand.
🧵[6/n]
Let me emphasize:
✅ We "derived" this. Just from F minimization. Without putting any of those terms in there by hand.
🧵[6/n]
May 19, 2025 at 6:36 AM
If you follow those prescriptions, you end up with this canonical circuit model that computational neuroscientists have been studying since 1970s.
Let me emphasize:
✅ We "derived" this. Just from F minimization. Without putting any of those terms in there by hand.
🧵[6/n]
Let me emphasize:
✅ We "derived" this. Just from F minimization. Without putting any of those terms in there by hand.
🧵[6/n]
But free energy (F) minimization is not enough.
We need specific "prescriptions" to guide top-down algorithm development. Otherwise, we risk falling into this trap:
P.S. I recently learned this is called #Bayesplaining 🙂
🧵[4/n]
We need specific "prescriptions" to guide top-down algorithm development. Otherwise, we risk falling into this trap:
P.S. I recently learned this is called #Bayesplaining 🙂
🧵[4/n]
May 19, 2025 at 6:36 AM
But free energy (F) minimization is not enough.
We need specific "prescriptions" to guide top-down algorithm development. Otherwise, we risk falling into this trap:
P.S. I recently learned this is called #Bayesplaining 🙂
🧵[4/n]
We need specific "prescriptions" to guide top-down algorithm development. Otherwise, we risk falling into this trap:
P.S. I recently learned this is called #Bayesplaining 🙂
🧵[4/n]
This paper is half synthesis/review, half new algorithm design.
We start by recognizing the free energy (F) principle as our best candidate for a unified theory of biological and artificial intelligence.
Here's why👇
🧵[3/n]
We start by recognizing the free energy (F) principle as our best candidate for a unified theory of biological and artificial intelligence.
Here's why👇
🧵[3/n]
May 19, 2025 at 6:36 AM
This paper is half synthesis/review, half new algorithm design.
We start by recognizing the free energy (F) principle as our best candidate for a unified theory of biological and artificial intelligence.
Here's why👇
🧵[3/n]
We start by recognizing the free energy (F) principle as our best candidate for a unified theory of biological and artificial intelligence.
Here's why👇
🧵[3/n]
Elegant theoretical derivations are exclusive to physics. Right?? Wrong!
In a new preprint, we:
✅ "Derive" a spiking recurrent network from variational principles
✅ Show it does amazing things like out-of-distribution generalization
👉[1/n]🧵
w/ co-lead Dekel Galor & PI @jcbyts.bsky.social
🧠🤖🧠📈
In a new preprint, we:
✅ "Derive" a spiking recurrent network from variational principles
✅ Show it does amazing things like out-of-distribution generalization
👉[1/n]🧵
w/ co-lead Dekel Galor & PI @jcbyts.bsky.social
🧠🤖🧠📈
May 19, 2025 at 6:36 AM
Elegant theoretical derivations are exclusive to physics. Right?? Wrong!
In a new preprint, we:
✅ "Derive" a spiking recurrent network from variational principles
✅ Show it does amazing things like out-of-distribution generalization
👉[1/n]🧵
w/ co-lead Dekel Galor & PI @jcbyts.bsky.social
🧠🤖🧠📈
In a new preprint, we:
✅ "Derive" a spiking recurrent network from variational principles
✅ Show it does amazing things like out-of-distribution generalization
👉[1/n]🧵
w/ co-lead Dekel Galor & PI @jcbyts.bsky.social
🧠🤖🧠📈
Thanks for sharing. I am also doing my best to spread the word 🙂 (see top left)
From: openreview.net/forum?id=ekt...
From: openreview.net/forum?id=ekt...
February 3, 2025 at 10:13 PM
Thanks for sharing. I am also doing my best to spread the word 🙂 (see top left)
From: openreview.net/forum?id=ekt...
From: openreview.net/forum?id=ekt...
I end with a call to action:
We are each gifted a finite amount of time on this earth.
How will you spend yours? Chasing fleeting distractions? Or contributing to humanity's deepest, longest ongoing quest—minimizing our collective KL divergence?
Your choice.
🧵[15/n]
🧠🤖🧠📈 #AI
We are each gifted a finite amount of time on this earth.
How will you spend yours? Chasing fleeting distractions? Or contributing to humanity's deepest, longest ongoing quest—minimizing our collective KL divergence?
Your choice.
🧵[15/n]
🧠🤖🧠📈 #AI
January 14, 2025 at 7:53 AM
I end with a call to action:
We are each gifted a finite amount of time on this earth.
How will you spend yours? Chasing fleeting distractions? Or contributing to humanity's deepest, longest ongoing quest—minimizing our collective KL divergence?
Your choice.
🧵[15/n]
🧠🤖🧠📈 #AI
We are each gifted a finite amount of time on this earth.
How will you spend yours? Chasing fleeting distractions? Or contributing to humanity's deepest, longest ongoing quest—minimizing our collective KL divergence?
Your choice.
🧵[15/n]
🧠🤖🧠📈 #AI
In the season finale (Part X), I will go all-in on speculation: p_brain is the all-encompassing object, swallowing everything. Even physics.
In that sense, all of science and philosophy merge into a single pursuit:
➡️ Humanity's collective KL minimization.
🧵[14/n]
In that sense, all of science and philosophy merge into a single pursuit:
➡️ Humanity's collective KL minimization.
🧵[14/n]
January 14, 2025 at 7:53 AM
In the season finale (Part X), I will go all-in on speculation: p_brain is the all-encompassing object, swallowing everything. Even physics.
In that sense, all of science and philosophy merge into a single pursuit:
➡️ Humanity's collective KL minimization.
🧵[14/n]
In that sense, all of science and philosophy merge into a single pursuit:
➡️ Humanity's collective KL minimization.
🧵[14/n]
But hold on a second. We’re not done yet.
It turns out, I need 10 more parts to finish the full story arc. Here’s what I have in mind next:
🧵[13/n]
It turns out, I need 10 more parts to finish the full story arc. Here’s what I have in mind next:
🧵[13/n]
January 14, 2025 at 7:53 AM
But hold on a second. We’re not done yet.
It turns out, I need 10 more parts to finish the full story arc. Here’s what I have in mind next:
🧵[13/n]
It turns out, I need 10 more parts to finish the full story arc. Here’s what I have in mind next:
🧵[13/n]