Ramin Hasani
@ramin_m_h
Followers
3,809
Following
296
Media
87
Statuses
1,106
Explore trending content on Musk Viewer
#MerzAestheticsXหลิงออม
• 795705 Tweets
LINGORM FILL
• 771047 Tweets
GEMFOT x TSS2024
• 92790 Tweets
期日前投票
• 81753 Tweets
#プロセカ放送局
• 69617 Tweets
Daisuke
• 33476 Tweets
ドラフト
• 31673 Tweets
#トキメキ旅
• 31363 Tweets
ボボステ
• 27201 Tweets
ヴェノム
• 25546 Tweets
サクラコ
• 24111 Tweets
WahWah Kabir Guru Pura
• 23572 Tweets
Hopkins
• 21798 Tweets
ブレイバーン
• 16342 Tweets
メルトリリス
• 15505 Tweets
Eşit
• 15281 Tweets
ドナルド
• 14767 Tweets
自民衝撃の197議席
• 13183 Tweets
オリエンス
• 11682 Tweets
Pinned Tweet
we just released LFMs! go check them out. So proud of our team
@LiquidAI_
🙏🏻
Today we introduce Liquid Foundation Models (LFMs) to the world with the first series of our Language LFMs: A 1B, 3B, and a 40B model. (/n)
85
266
2K
7
10
63
In a new article published today in Nature MI
@NatMachIntell
we solved a differential equation that describes the interaction of neurons and synapses! This equation had no known closed-form solution since the year 1907!
This solution is important! 1/n
57
712
3K
Because, we can now simulate brain dynamics composed of billions of neurons and trillions of synapses with biologically realistic mechanisms! Something we could not do since today!
Moreover, inspired by this solution… 2/n
5
88
440
I am thrilled to announce the launch of
@LiquidAI_
!
is an MIT spin-off designing a new generation of foundation models that are private, flexible, and reliable.
We raised $37.6M in seed capital led by
@OSSCapital
& PagsGroup. 1/n
12
40
275
very easy to use and is available in Torch and TF!
Paper:
Code:
This is a join work of
@mlech26l
and I, together with
@xanamini
,
@LLiebenwein
Aaron Ray, Gerald Teschl, Max Tschikowski, & Daniela Rus
@MIT_CSAIL
@MIT
4/n n=4
15
45
281
We built a new powerful neural network model which is flexible, liquid, and causal and is super fast!
The model is called closed-form continuous-time (CfC) networks and could potentially become the core building block of future intelligent systems. 3/n
10
38
185
had an amazing time today with the legendary and unparalleled Joscha Bach
@Plinz
, discussing the future of AI and true AGI.
Topics discussed:
Agency
Liquid
self-organization
Evolution
brains
7
4
125
In a new article published today in Science Robotics
@SciRobotics
, we showed a class of neural nets can understand the task they are given (flight navigation)!
These are AI systems that WE understand what they do and THEY understand what they do! (1/n)
2
15
73
Announcing our new performant sequence modeling algorithm:
Liquid-S4
Liquid neural nets with S4 parametrization = SOTA in modeling sequences! 1/n
With
@mlech26l
@johnsonwang0810
@MakramChahine
@xanamini
and Daniela Rus
@MIT_CSAIL
@MIT
3
12
57
#neurips2021
Did you know that for Neural ODEs and Continuous Normalizing Flows, pruning is all you need?!
[Paper]
[Talk]
@LLiebenwein
@xanamini
@MIT_CSAIL
@MIT
0
15
55
now you know why Liquid-S4!
✨Excited to finally drop our new paper: SSMs “look like” RNNs, but we show their statefulness is an illusion🪄🐇
Current SSMs cannot express basic state tracking, but a minimal change fixes this! 👀
w/
@jowenpetty
,
@Ashish_S_AI
24
206
1K
1
5
46
I co-founded with my friends
@mlech26l
@xanamini
and Daniela Rus off of
@MIT_CSAIL
.
We have always been fascinated by natural learning systems and the ways they inspire us design better algorithms! We finally did it and will share with the world 2/n
2
4
45
Thank you
@QuantaMagazine
& Steve Nadis for featuring such a great article about our work on “liquid neural networks”
It describes the journey
@mlech26l
, Daniela Rus & I took
@MIT
@MIT_CSAIL
to develop liquid networks and what’s next
0
11
42
today, I want to share the core values that shape our culture at Liquid. here we go:
no-bullshit meritocracy,
burn the playbook,
proactive execution and purposeful ownership,
be white-box explainable and
let's grow together.
Allow me to elaborate:
-------------
A CULTURE
4
8
40
hmmm world’s best 🧐🙃
5
6
43
We are so grateful of the trust and support of our investors
@JosephJacks_
, Steve Pagliuca,
@JimBreyer
, Tom Preston Werner,
@naval
, Safar Partners,
@Capgemini
,
@tobi
, David Siegel,
@chrisprucha
,
@SamsungNext
@PeterDiamandis
,
@DukeCapPartners
, Bob Young,
@automattic
, and many
3
3
38
0
0
38
Our new
#icml2020
paper introduces a novel class of neural networks directly taken from the nervous system of small species, to control robots in RL environments with competitive performance to deep models.
In this thread, I summarize our findings (1/n)
1
5
36
Excited to announce that our paper got accepted at
#ICRA2019
“Designing worm-inspired neural networks for interpretable robotic control”
a joint work of us at
@tuvienna
, Mathias Lechner and
@thenzinger
of
@ISTAustria
, and Manuel Zimmer of
@IMPvienna
@univienna
0
2
34
So excited to have 2 submissions accepted at
#NeurIPS2022
🥳
Congratulations to my co-authors
@LLiebenwein
@xanamini
@MLech20
Charlie Vorbach, and Daniela Rus!
- Causal Navigation by Continuous Neural nets 🦅
- Sparse Flows 🕸
1
3
33
In our new work featured on the cover of
@NatMachIntell
,
@MLech20
,
@xanamini
& I introduced “neural circuit policies” (NCPs) for interpretable autonomous control.
Paper:
Code:
pip install keras-ncp
you are ready to go! :)
1
14
32
models are still too small!
Human brain has more than 100 trillion synapses. I expect to see a big leap in performance at a 100 trillion mark!
2
0
31
@MLStreetTalk
have you heard about the double descent phenomena? that is the reason why even when you overfit you get robust generalization.
this happens in the overparameterization regime.
see this:
and my lecture on it:
2
1
30
0
1
29
If you wanna learn fundamentally what OpenAI is doing, start from here:
1
5
28
Discussing how we think about the future of AI
@LiquidAI_
, in Davos alongside
@wef
with
@felixnaser
@xanamini
Steve Pagliuca, and Daniela Rus!
1
4
27
This is the latest from the doomers!
Instead of contributing something meaningful to science, they cheerfully become dirty city rats with zero added-value to the society!
@openai
gave you the best software in the history of mankind!!! protect it instead of bullshitting
OpenAI is in a heap of trouble, and it’s not just text.
Long thread why (1/n), based on work with
@Rahll
230
701
3K
7
0
22
So proud to be partnering with
@LambdaAPI
on this release! Lambda and team have been an awesome compute and infra partner for us for a while now!
special thanks to Sam Khosroshahi,
@mitesh711
,
@stephenbalaban
, Rahul Druggal, and Paul Sebexen
A new LLM generation:
@LiquidAI_
announces foundation models built on new high-efficiency architecture. Test LFM-40B for free w/
@LambdaAPI
0
17
66
0
5
23
we're combining our expertise with
@Capgemini
to innovate in AI, pushing boundaries in domains from healthcare to finance.
this collaboration is not just about technology, but making the trusted AI market for enterprises!
Today we announce our collaboration with Capgemini to build next-generation AI solutions for enterprises. For the last months, we've been working on this together and now following Capgemini's participation in Liquid AI's successful $37.6m seed round, we are committed to
7
17
114
0
3
21
#icml
Nature has gifted us neural networks that are lottery ticket winners! Stay tuned to learn more in our accepted
#icml2020
paper “Ordinary Neural Circuits”, with
@MLech20
,
@xanamini
, Daniela Rus, and Radu Grosu.
Joint work of
@tuvienna
,
@ISTAustria
and
@MIT_CSAIL
@icmlconf
0
3
22
Liquid-S4 is now accepted to
#ICLR2023
🥳🥳
paper:
code:
with
@mlech26l
@johnsonwang0810
Makram Chahine,
@xanamini
and Daniela Rus
@MIT_CSAIL
@MIT
Announcing our new performant sequence modeling algorithm:
Liquid-S4
Liquid neural nets with S4 parametrization = SOTA in modeling sequences! 1/n
With
@mlech26l
@johnsonwang0810
@MakramChahine
@xanamini
and Daniela Rus
@MIT_CSAIL
@MIT
3
12
57
2
3
22
1
0
22
A couple of weeks ago I gave a TEDx talk on Liquid Neural Networks at
@TEDxMIT
The talk is now released here:
special thanks to the TEDxMIT team for putting together such an incredible event!
please feel free to check out the talk and reach out!
0
2
20
@neuralreckoning
[Author here] relax! there is no new math invented! We used standard methods + some approximation tricks & novel sharpness results, to design a new neural net model that resembles the neuron-synapse interaction. Very useful in large brain simulations and a tractable ml model!
2
0
20
3 hours ago on
@Forbes
, Daniela Rus tells you about our invention Liquid nets!
@mlech26l
, myself,
@xanamini
& Daniela at
@MIT_CSAIL
pushed the expressivity boundary of DL systems to a new regime beyond what was achievable before.
1
5
20
Liquid neural networks will bring RNNs back to life! Updates are coming soon!
for now, you can:
read about them
See details in this
#aaai
paper:
and play around with them here:
@MIT
@MIT_CSAIL
@MLech20
0
9
20
The measure of intelligence is the ability to change
-Albert Einstein
4
5
18
Enterprise AI needs solutions that are
1) very capable and reliable (we haven’t seen this yet),
2) cost, energy and carbon efficient (we have seen a bit of this), and
3) customizable (we have seen a lot of this).
In this priority order and not the other way around.
1
3
19
Last Tuesday I gave a seminar talk on Liquid Neural Nets
@MIT_CBMM
!
here is the talk’s recording:
Special thanks to the entire CBMM team, Prof. Daniela Rus and Prof. Tomaso Pogio for providing me with this great opportunity.
[video] "Liquid Neural Networks"
Ramin Hasani, MIT
We discuss the nuts and bolts of the novel continuous-time neural network models: Liquid Time-Constant (LTC) Networks. Instead of declaring a learning system's dynamics by implicit nonlinearities, LTCs ...
0
4
13
1
3
19
We got bored of developing new ml models!
so instead, we showed some crazy properties of neural networks in the infinite width limit!
2 papers accepted at
#NeurIPS2022
By our legendary phd student
@loo_noel
!
with
@xanamini
, and Daniela Rus
@MIT_CSAIL
@MIT
0
2
17
An example of our research on training neural circuit policies by reinforcement learning, on standard control test-beds!
via
@physorg_com
0
5
15
Today, I defended my
#phd
studies at home! I have been honored to be working with so many great minds at
@tuvienna
,
@MIT_CSAIL
and
@ISTAustria
6
0
15
Just had a great discussion on “interpretable recurrent neural networks” with the
@PreferredNet
team at the offices in Tokyo, thanks
@sla
for having me!
0
1
14
It was so much fun speaking with
@Jason
last week on
@twistartups
!
Thanks for having me 🙏🏻
Enjoy 👇
New Episode!
@ramin_m_h
says
@liquidAI_
began with a mission to create AI from scratch, based on biology and physics
This led to the construction of liquid neural networks
Now worm-inspired AI systems are driving vehicles and flying jets!
1
6
15
0
2
13
@ylecun
Wrong. First of all, this is not a flavor of ConvNet! This work is a variant of S4 which is much more than a ConvNet. Second, the results are weaker than recent SSMs, such as Liquid-S4: and S5: which are SOTA in this space!
2
0
13
come say hi at this year’s
#ICLR
in Vienna! A lot of us at
@LiquidAI_
gonna be around presenting some of our recent research!
1
1
12
Had a wonderful morning chat with the legendary
@YesThisIsLion
, Transformers co-inventor in Tokyo!
We discussed about:
brains
evolution
@SakanaAILabs
@LiquidAI_
what comes after Transformers
and the future of AI research!
2
2
12
@AISafetyMemes
@ylecun
Yann’s response is indeed very accurate and informative given the arguments you are making
3
0
12
A couple of weeks back I gave my first TEDx talk at
@TEDxVienna
, my talk is out now, watch it online:
0
1
11
Talking at
@TEDxVienna
was an incredible experience. Learned priceless life-lessons from many professionals, got to know the brightest individuals and shared my work with an outstanding audience. I experienced a dozen of different emotional states all at once! This was new!
2
1
11
#lifemilestone
: Going to give my first TEDx talk at
#tedxvienna
on October 20th 2018
@ Vienna, Austria
1
2
11
@hardmaru
One of the most beautiful and fundamental works addressing some flavors of this question is the Universal Law of Robustness by Bubeck and Selke NeurIPS 2021 (Best paper)
Overparametrization is *necessary* for perfect memorization & Worst-case robustness
1
0
10
@LucaAmb
@nhatnguyen913
@NatMachIntell
[Author here] relax! there is no new math invented! We used standard methods + some approximation tricks & novel sharpness results, to design a new neural net model that resembles the neuron-synapse interaction. Very useful in large brain simulations and a tractable ml model!
0
0
9
@radekosmulski
Impressed by Optuna?? Give pyhopper a try :)
@mlech26l
did anther magic to make hyperparameter optimization ridiculously easy!
Pyhopper is simple and elegant and is at least 10x faster than Optuna!
1
0
8
@carlcarrie
Hold on guys! they forgot to compare results to the state of the art time-series modeling framework: S4 (Structural State-Space Sequence models).
State Space models widely outperform Transformers in time-series modeling!
0
0
9
Will be presenting “Worm-level Control through Serarch-based Reinforcement Learning”
at the NIPS Deep RL Symposium ()
@NipsConference
0
6
9
So excited to present this work next week at
#aaai2021
A machine-learning system developed by
@MIT_CSAIL
researchers learns on the job, not just during its training phase. By continuously adapting to new data inputs, this “liquid network” could aid decision-making in medical diagnosis and autonomous driving.
1
22
69
2
2
8
Conformal Prediction is fascinating!!
Looking forward to Prof. Candès keynote talk at
#NeurIPS2022
0
2
8
My talk at TEDxBoston is out!
I spoke about generalist AI systems from ChatGPT to liquid nets, and showcased applications in finance and in robotics!
Special thanks to John Werner for organizing such an incredible event!
2
1
8
Excited to announce that our paper got accepted to
#icra2020
Stabilizing the Recurrent Neural Network Compartment of an End-to-end Robot Learning Scheme
a joint work of us
@tuvienna
, Mathias Lechner
@ISTAustria
and Daniela Rus
@MIT_CSAIL
0
1
7
“When leaders are willing to prioritize trust over performance, performance almost always follows.” Simon Sinek
0
2
7
Thank you all for the support of our work! In case you have questions about it, please drop them under this
@MIT_CSAIL
post, where I can provide answers!
Hi there! It’s Ramin Hasani
@raminmh
! Over the next 24H, I’m taking over
@MIT_CSAIL
’s Twitter! I’m a research affiliate at CSAIL. I design brain-inspired robust deep learning algorithms!
Ask me anything about sequence modeling, time series, robots, & liquid neural nets, here 👇
12
11
67
0
2
7
@Ofirlin
@allonsygamma
@YiMaTweets
I do you one better: “reducing the complexity of Neural Tangent Kernels from O(n^2) to O(n) is not significant enough”! wtaf 😐
2
1
7
“Uncertainty is the only certainty there is, and knowing how to live with insecurity is the only security” (John Allen Paulos, 1945-)
0
1
7
This year
@iclr_conf
2024 we will present the following papers with our brilliant students
@MIT
@MIT_CSAIL
&
@LiquidAI_
1) Leveraging Low-Rank and Sparse Recurrent Connectivity for Robust Closed-Loop Control
Neehal Tumma,
@loo_noel
,
@mlech26l
, RH, Daniela Rus
(Spotlight - Top
0
2
6
The AI world is all zoomed in on GPTs and their derivatives! I think liquid networks can provide a refreshing view of how to build powerful AI systems that understand what they do and can reliably get integrated into our society, without being massively overparameterized. (14/n)
2
0
6
@tianle_cai
@SebastienBubeck
Nice work!
Though you forgot to compare your results to Liquid-S4 which is the current SOTA SSM in LRA, SC and vital sign prediction. Check it out:
1
0
6
The mother of all diffusers and llms is a universal signal processing system — 🐛
1
2
6
@chrisdonahuey
you presenting Dance Dance Convolution :D 👌🏻 good job man
#icml2017
#deeplearning
#applications
1
0
6
0
7
6
We then hypothesize that if a liquid agent really understood the task of fly-to-target, it should be able to follow the target even if the target starts moving!!!
Observation 7: The answer is YES! Liquid networks complete this el matador task, zero-shot! (10/n)
1
0
6
The sky is the limit for the applications of liquid neural networks!
This is a joint work of us
@MIT_CSAIL
@MIT
, with
@makinai_
P. Kao, A. Ray, R. Shubert,
@mlech26l
@xanamini
, and Daniela Rus.
paper:
(15/n=15)
1
1
6
@HochreiterSepp
and I will be sharing a panel to discuss the future of AI in Europe at the next
@TEDAIVienna
1
0
6
Must read work in black-box optimization!! Very well done
@RobertTLange
Check out Rob’s EvoSax library on Evolution Strategies! It’s a piece of art!
👋 Come by poster 93 in this mornings
#ICLR2023
poster session to chat about our work on Learned Evolution Strategies (LES) 🦎
📝:
0
8
116
0
0
5
F = T ∇S "Intelligence is a force, F, that acts to maximize future freedom of action or keep options open, with some strength T, with the diversity of possible accessible futures, S, up to some future time horizon. Intelligence doesn't like to get trapped".
Alex Wissner-Gross
0
1
4
Nice work on a smarter sharding method.
The solution here: The more GPUs you have the longer you can go in context length.
New paper w/
@matei_zaharia
@pabbeel
on transformers with large context size.
We propose RingAttention, which allows training sequences that are device count times longer than those of prior state-of-the-arts, without attention approximations or incurring additional overhead.
10
179
849
0
0
5
Had the opportunity to contribute to this work. Have a look at the paper which is going to be presented next week at
#rss2018
conference
MIT system lets humans control robots with brainwaves and hand gestures:
4
121
205
1
1
5
really feel honored to have shared the same stage with
@geoffreyhinton
only 10 days ago, at the
@Capgemini
Spark event in Cali!
1
0
5
We took the drone and chair to urban areas with some adversaries around to see if the networks really understood the task.
Observation 4: Liquid nets ignore irrelevant background features, find the target and complete the task at a rate beyond what the other systems do! (7/n)
1
0
5
S4 is one of the most elegant sequence modeling frameworks emerged from linear systems with structures and time/frequency transformations
#ICLR2022
Shout out to
@_albertgu
and
@krandiash
!
1
1
5
@RealAAAI
the first theoretical grounds for the verification of neural ODEs!
With Sophie Gruenbacher,
@MLech20
, Jacek Cyranka, Scott Smolka, & Radu Grosu
Joint work of
@tuvienna
,
@MIT_CSAIL
@ISTAustria
U of Warsaw, and
@stonybrooku
#aaai2021
Paper:
0
0
5
@ylecun
yes! if you (Yann LeCun) write an opinion piece, a roadmap article, or a review paper on the topic, yes, he should be cited, because you know it! 🤷🏻♂️
2
0
5
0
0
4
Let’s see some results first! on the Long Range Arena benchmark, Liquid-S4 outperforms all variants of SSMs as well as Transformers and CNN baselines with a good margin, as we see below 2/n
1
0
4