IgorMezic Profile Banner
Igor Mezic Profile
Igor Mezic

@IgorMezic

Followers
893
Following
233
Statuses
257

Distinguished Professor of Engineering at UCSB, Pioneering Third-Wave AI, co-Founder of AIMDyn, co-Founder, CTO & Chief Scientist at MixMode.

Santa Barbara, CA
Joined September 2018
Don't wanna be here? Send us removal request.
@IgorMezic
Igor Mezic
17 days
A nice application of #Koopmanoperator theory to galactic dynamics here:
1
17
91
@IgorMezic
Igor Mezic
26 days
Last May, we held a workshop in Otranto, Italy, connecting Koopman operator theory and climate science. Here is a writeup on it #AI #Koopmanoperator #ML #dynamicalsystems #chaostheory
0
6
29
@IgorMezic
Igor Mezic
1 month
RT @wredman4: ๐Ÿšจ new paper out in Chaos "Koopman Learning with Episodic Memory"! Really excited about this work and,โ€ฆ
0
14
0
@IgorMezic
Igor Mezic
1 month
I got into science in no small part due to reading James Gleick's "#Chaos" with its wonderful depictions of the discovery process. #Lorenzattractor was one of those great discoveries. It is an example of a physical system exhibiting chaotic motion and sensitivity to initial conditions, and thus lack of predictability despite the simplicity of its description. We only saw the proof of its "chaoticity" in 2004 But the attractor still holds secrets. For example, with #Koopmantheory we discovered complex-valued observables that have much better #predictability and less sensitivity to initial conditions than typical for this system. One of those observables - discovered first by spectral projection method in - whose contour plot from the paper is shown below as a) - looks like a sinusoid on a Mobius strip. It was found with a different method in . The b) figure below - from that paper - shows the plot of the evolution of the "regular" observable compared with the sinusoid, to indicate its regularity. In contrast, part c) shows the typical chaotic evolution of time series of observables on the Lorenz attractor.
Tweet media one
2
25
189
@IgorMezic
Igor Mezic
3 months
AI for dynamical systems: the architecture that incorporates 1) a block of DNN for embeddings (feature engineering) and 2) a block of Koopman Operator based predictor/controller is emerging as a promising one in terms of accuracy and lean computation. Kolmogorov Arnold Networks (KAN's) are theoretically interesting and offer computational advantages in this context. Here is an interesting paper utilizing these structures #AI #ML #KoopmanOperator @Aimdyn_Inc
0
8
67
@IgorMezic
Igor Mezic
3 months
Koopman Operator Theory provides a framework for #metalearning of #neuralnetworks . Our paper - accepted as #NeurIPS highlight - with @wredman4 et al. provides methodology for comparing neural network training protocols. #AI #ML #KoopmanOperator #DynamicalSystems
4
34
141
@IgorMezic
Igor Mezic
3 months
RT @wredman4: Our paper "Identifying Equivalent Training Dynamics" was accepted as a spotlight (top ~5%) ๐Ÿ’ก to NeurIPS! A real dream team (bโ€ฆ
0
13
0
@IgorMezic
Igor Mezic
3 months
RT @wredman4: A big thanks to the great team!: @hopfbifurcator @s_acosta_4 @ninamiolane Newest version of the paper:
0
5
0
@IgorMezic
Igor Mezic
3 months
0
0
1
@IgorMezic
Igor Mezic
4 months
Nice recognition by the USA Today that #Koopmanoperator based AI systems provide a framework for real time diagnostics and reasoning for network security threats! Congrats to everyone at @MixModeAI ! #AI #ML
@MixModeAI
MixMode
4 months
๐Ÿš€ Big news! MixMode has been recognized by @USATODAY as one of theย Top 10 AI Companies to Watch in 2024!ย ๐ŸŽ‰
Tweet media one
0
5
19
@IgorMezic
Igor Mezic
4 months
AI omnipresent these days but most of the news we hear are about LLM's and generative AI for visuals. There is much more use for #AI , some in areas such as human ability improvement. Here is one that can be applied to rehabilitation from hand injuries #KoopmanOperator #ML
0
1
13
@IgorMezic
Igor Mezic
5 months
USA News and World report issued their latest rankings . It is interesting that UCSB, is 6th (!!!!) in the nation measuring Faculty Research Impact (second of all public schools)! By my count in front of it are only MIT, Stanford, Harvard, Berkeley and Princeton...
Tweet media one
0
0
6
@IgorMezic
Igor Mezic
5 months
For Koopman modelers: I added to my Youtube channel the series of videos on the PyKMD software. #AI #KoopmanOperator #ML #DynamicalSystems
1
10
36
@IgorMezic
Igor Mezic
5 months
The need for a paradigm shift in building AI models is clear. #AI #Koopman #ML
0
0
11
@IgorMezic
Igor Mezic
5 months
My July lecture at the Fields Institute on "The Operator is the Model" is at #AI #ML #dynamicalsystems #KoopmanOperator #theoperatoristhemodel
0
11
46
@IgorMezic
Igor Mezic
5 months
Given my training in guitar, I feel there is a full circle achieved with this paper that considers dynamics of strings of musical instruments using Koopman operator-based machine learning ๐Ÿ˜€. Added bonus: comps to a number of other methods, all outperformed by the Koopman spectral method. #dynamicalsystems #theoperatoristhemodel #AI #ML #Koopmanoperator
0
2
14
@IgorMezic
Igor Mezic
5 months
One of the reasons AI systems need an enormous amount of training and associated energy is that they do not feature a layered design that separates perception (detection), reasoning and communication. @MixModeAI 's AI -featuring @Aimdyn_Inc's Koopman operator based methodology for network security - features such design, enabling it to train lean, personalized, real-time adaptive AI models on customers networks. #AI #ML #leanAI #KoopmanOperator #DynamicalSystems
@MixModeAI
MixMode
6 months
๐Ÿšจ Is your security team overwhelmed by false positives? MixMode's #AI-powered platform uses a 3-layered approach to deliver prioritized alerts, helping your analysts make informed decisions faster.
0
1
9
@IgorMezic
Igor Mezic
6 months
@ylecun @ylecun I fully agree. MPC for nonlinear models where non-convexity can be a factor can be convexified as in BTW it does not have to be MPC, sometimes LQR suffices: #AI #ML #OptimalControl #Koopman #DynamicalSystems
0
0
9
@IgorMezic
Igor Mezic
6 months
I found it interesting that the "quadratic" (or "square root") nature of QM probabilities actually has a classical correspondence in dynamical systems theory , see Remark 2 in This renders the Bohr interpretation as natural if one insists on computing classical probabilities via unitary operators. #quantumphysics
2
0
2