iiinova — where your digital self keeps evolving

4/4 "When we want to understand dynamics, we really need to understand dynamics in different spaces," said @danielkunin.bsky.social, speaking at TTIC. These are the dynamics of the high-dimensional parameter space, the space of possible functions and the performance space.

000

3/4 "And deep learning is really about the interaction of these three components, not just one of the components. We can't study them independently of each other," said @danielkunin.bsky.social, speaking at TTIC. Video: www.youtube.com/watch?v=4eMW...

100

2/4 "We have these deep, non-linear architectures, we have high-dimensional natural datasets, and we have stochastic gradient-based optimization," said @danielkunin.bsky.social, speaking at TTIC.

100

"Learning Mechanics of Neural Networks" - Daniel Kunin, Young Researcher Series Seminar

1/4 The Dynamics of Deep Learning. "Trying to understand deep learning...comes down to understanding this interaction of three components that make the deep learning system," said @danielkunin.bsky.social, speaking at TTIC in April last year. Video: www.youtube.com/watch?v=4eMW...

youtube.com

TTIC @tticconnect.bsky.social · 25/09/2026

121

Monday, September 28 at 11:30am: TTIC Colloquium presents Osbert Bastani of UPenn with a talk titled "Scaling Sim2Real Reinforcement Learning." Please join us in Room 530, 5th floor.

000

TTIC @tticconnect.bsky.social · 25/09/2026

TTIC honored nine graduates at its annual Diploma and Awards Ceremony: Jiahao Li, Shengjie Lin, Xiao Luo, Keziah Naggita, Max Ovsiankin, Kavya Ravichandran, Pedro Savarese, Ziwei Xie, and David Yunis. Read more: buff.ly/CXDyRca

TTIC @tticconnect.bsky.social · 24/09/2026

Congratulations to Dimitar Chakarov, recipient of the 2026 Outstanding TA Award for his work in Professor Madhur Tulsiani’s autumn 2025 course, “TTIC 31150 Mathematical Toolkit.”

000

TTIC @tticconnect.bsky.social · 23/09/2026

Welcome, new TTIC students! Our incoming class brings new ideas and research in #CS, #machinelearning, #AI, and more - and we're excited to start your Ph.D. journey with you!

000

TTIC @tticconnect.bsky.social · 22/09/2026

"Learning in an Echo Chamber: Online Learning with Replay Adversary" by Daniil Dmitriev @ddmitriev.bsky.social is now available to stream on our YouTube Channel: youtu.be/55W2mi-NHxQ

"Learning in an Echo Chamber: Online Learning with Replay Adversary" – Daniil Dmitriev, LAMP

TTIC @tticconnect.bsky.social · 21/09/2026

000

TTIC is hiring Research Assistant Professors! 3-year, fully funded role — no teaching required, full research freedom, strong mentorship & collaborations. Join a top #CS research community in Chicago. Apply by Dec 1: ttic.edu/facapp #CSFaculty #CSJobs

000

TTIC @tticconnect.bsky.social · 18/09/2026

Thank you to all who participated in the Human-Machine Creativity Workshop at TTIC, in partnership with the UChicago Dept. of Visual Arts. The day featured research presentations, a panel discussion, and substantive dialogue at the intersection of art and #AI.

000

4/5 Vinker wants to enable visual communication between humans & machines. But there are challenges. For e.g., for machines to speak the language of sketches, they would have to understand the sequential nature of sketching and identify meaningful parts and sketch accordingly.

100

3/5 A sketch is a set of Bézier curves, whose parameters are optimized using a differentiable rasterizer, by minimizing a CLIP-based perceptual loss, which accounts for semantics & geometry. The fewer the number of curves the more the abstraction. Source: clipasso.github.io/clipasso/

100

2/5 As an example, Vinker spoke of their work on CLIPasso, which uses CLIP embeddings to convert an image of an object into a sketch, at varying levels of abstraction. "Abstraction is not only what we see, it's also what we know," said @yaelvinker.bsky.social. "We can draw whatever CLIP knows."

100

TTIC @tticconnect.bsky.social · 15/09/2026

1/5 "Sometimes words are not enough," said Yael Vinker, @yaelvinker.bsky.social, of @csail.mit.edu at TTIC's Young Researcher Seminar Series in Nov '25. "I want to teach machines to speak the language of sketches." This involves using visual means to communicate concepts under a set of constraints.

222

TTIC researchers Shiry Ginosar (@shiryginosar.bsky.social), Kaylene Stocking, Matt Walter, and Alison Gopnik (UC Berkeley) received a GRCDI grant to study how robots could learn the way children do — through play, exploration, and expanding physical capabilities: buff.ly/TVYMAsk

"Science of AI and AI for Science "- Noah Smith

4/4 "Use OLMo. It’s fully open. You can see the data. You’ll be contributing back to this growing ecosystem of tools for science," said Noah A. Smith of UW/@ai2.bsky.social at TTIC. Video: www.youtube.com/watch?v=hFz7...

youtube.com

000

3/4 The OLMo project is @ai2.bsky.social's response. "We crafted this project to build fully open models for reproducibility, for transparency, to make it easier for more people to use and study the models themselves," said @nlpnoah.bsky.social, at TTIC. Video: www.youtube.com/watch?v=hFz7...

110

2/4 "The reason our field has had so much success is that people used to talk openly about everything they did in the research community," said @nlpnoah.bsky.social of UW/@ai2.bsky.social. "It’s only in the past few years that exciting advances in our field have been treated as proprietary."

100

TTIC @tticconnect.bsky.social · 11/09/2026

1/4 "Trying to do AI research on proprietary models is like trying to do astronomy based on pictures of the sky printed in the newspaper"—@nlpnoah.bsky.social of UW/@ai2.bsky.social, quoting a friend during a talk at TTIC last year. His advocacy of open models strikes a stronger chord today

101

The IDEAL 2026 Annual Meeting was hosted yesterday at TTIC! Thank you to the attendees who joined us for talks, updates, and breakout sessions. View our upcoming workshops: ttic.edu/workshops

000

TTIC @tticconnect.bsky.social · 10/09/2026

How should #AI systems fairly credit creators whose work shapes their outputs? Siddharth Prasad (TTIC) presents "In-Context Credit Assignment via the Core," applying cooperative game theory to distribute value among content creators. Now streaming: buff.ly/4MFIuCW

"In-Context Credit Assignment via the Core" – Siddharth Prasad, LAMP

“In-Context Credit Assignment via the Core” Siddharth Prasad, TTIC Originally recorded on August 6, 2026, at TTIC as part of the Learning-driven Algorithms and Machine-aided Proofs (LAMP)…

TTIC @tticconnect.bsky.social · 09/09/2026

How does individual decision-making shape equilibrium and efficiency in contagion control? Sanjiv Kapoor of Illinois Tech examines this in "Contagion Control-Policy Games," the latest TTIC Colloquium. Now on YouTube: buff.ly/T5lncex

"Contagion Control-Policy Games" – Sanjiv Kapoor, TTIC Colloquium

“Contagion Control-Policy Games” Sanjiv Kapoor, Illinois Tech Originally recorded on August 31, 2026, at TTIC. In this talk, Sanjiv Kapoor explores game-theoretic approaches to controlling…

000

TTIC @tticconnect.bsky.social · 08/09/2026

TTIC researchers present their work on #computervision at ECCV 2026. Congratulations to Xiaodan Du, Shiry Ginosar, Greg Shakhnarovich, and David Yunis on their papers, posters, and talks: buff.ly/tSOK5DU

000

TTIC @tticconnect.bsky.social · 03/09/2026

“Prediction-Specific Design of Learning-Augmented Algorithms," presented by Nicolas Christianson of Johns Hopkins University for the LAMP workshop at TTIC, is now available to stream on YouTube: youtu.be/WRURJiCTltk

"Prediction-Specific Design of Learning-Augmented Algorithms" – Nicolas Christianson, LAMP Workshop

TTIC @tticconnect.bsky.social · 31/08/2026

000

Wednesday, September 2nd at 11:00am: Talks at TTIC presents Kai-Wei Chang of @csail.mit.edu with a talk titled "Towards Time-Aware Full-Duplex Spoken Dialogue Models." Please join us in Room 530, 5th floor.

000

5/5 "The goal of learning theory is to...spend a little bit of the cost learning from this data, and hopefully in doing so reduce the time of solving your linear systems by a significant amount to make it worth it," said Misha Khodak of @uwmadison.bsky.social, at TTIC.

000

4/5 When solving a simulation using discretization in physical time, the main cost at each timestep often involves solving a linear system. "[This] is generating a ton of data, the residuals...the difference between our output and the ground truth," said Misha Khodak of @uwmadison.bsky.social

100

3/5 Integrationist approaches, on the other hand, incorporate some machine learning into the PDE solver pipeline, and have the advantage of preserving the correctness guarantees provided by the solver itself, said Misha Khodak of @uwmadison.bsky.social, at TTIC.

100

2/5 In full replacement, PDE solvers are replaced with neural operators trained on simulation data or physics-informed neural networks trained to be PDE ansatzs. These work well (fast inference) but come with some disadvantages (no correctness guarantees), said Misha Khodak of @uwmadison.bsky.social

100

TTIC @tticconnect.bsky.social · 28/08/2026

1/5 “There's a lot of excitement about using ML in scientific computing,” said Misha Khodak of @uwmadison.bsky.social, at TTIC’s Learning-driven Algorithms and Machine-aided Proofs Workshop. For e.g., using ML to solve transient partial differential equations (PDEs)

100

Chicago high school students spent 6 weeks this summer building and programming robot arms to play tic-tac-toe through TTIC's Autonomous #Robotics Immersions internship. Field trips, hands-on mentorship, and real #AI research experience. Read about the experience here: buff.ly/DwtANDr

TTIC @tticconnect.bsky.social · 27/08/2026

Monday, August 31 at 11:30am: Sanjiv Kapoor of Illinois Tech with a talk titled "Contagion Control-Policy Games." Join us in Room 530, 5th floor.

000

5/5 "Clearly, there is a big disconnect between the theory of AI and how AI is currently used...It's time to come up with a theory for modern AI," said Maria-Florina (Nina) Balcan of @csdatcmu.bsky.social, at TTIC's Learning-driven Algorithms and Machine-aided Proofs Workshop.

000

4/5 Also, "in classic AI theory [for multi-agent systems], we use game theory lenses, but this has its own limitations...all learning is typically for very simple learners, based on some very, very simple dynamics," said Maria-Florina (Nina) Balcan of @csdatcmu.bsky.social at TTIC.

120

3/5 However, the "classic theory of ML is for much simpler scenarios, typically a single learner in its own world learning a relatively simple task...this is beautiful theory, but still very insufficient for the modern day," said Maria-Florina (Nina) Balcan of @csdatcmu.bsky.social at TTIC.

100