Jakob Heiss

Postdoc at UC Berkeley on Deep Learning Theory and Uncertainty Quantification
PhD student in Mathematics from ETH Zürich

[Ilia Azizi ➤NeurIPS'26 @ilia_azizi]

CLEAR is available as an open-source Python package, so feel free to try it:

🐍 pip install clear-uq
💻 https://t.co/2hEuGjzOj8
📄 https://t.co/LK0Cl6qsCM
🔗 https://t.co/bjbQMc0cit

Joint work with my great co-authors @JurajBodik , @JakobHeiss, and @bbiinnyyuu.

(4/4)

[Ilia Azizi ➤NeurIPS'26 @ilia_azizi]

On 17 real-world datasets with ensembles + quantile regression:
- 28.3% tighter intervals vs aleatoric-only baselines
- 17.5% tighter vs epistemic-only baselines
- Top method on 15 of 17 datasets

Similar gains with Deep Ensembles & Quantile NNs confirm generalizability.

(3/4)

[Ilia Azizi ➤NeurIPS'26 @ilia_azizi]

CLEAR wraps around any base model with just two calibration parameters, no retraining.

It determines how much each uncertainty source contributes. Compatible with any pair of estimators (ensembles + quantile regression), using ideas from conformal prediction.

(2/4)

[Ilia Azizi ➤NeurIPS'26 @ilia_azizi]

ICLR 2026! 🎉

Our paper "CLEAR: Calibrated Learning for Epistemic and Aleatoric Risk", accepted @iclr_conf.

Where does prediction uncertainty come from? Aleatoric (measurement noise) or epistemic (limited data)? Most methods only address one. CLEAR addresses both.

(1/4)

[BIRS @BIRS_Math]

This week at BIRS: Uncertainty Quantification in Neural Network Models https://t.co/eh9ODZ3RyD

[Jakob Heiss @JakobHeiss]

In our #aaai2024 paper "Machine Learning-powered Combinatorial Clock Auction" we modify this mechanism to ask more practical demand queries instead of value queries.
https://t.co/08UdZiMJ3x
Ermis Soumalias @JWeissteiner @SvenSeuken
@UZH_ch @ETH_en @ETH_AI_Center @ERC_Research

[Jakob Heiss @JakobHeiss]

Our #AAAI2023 paper "Bayesian Optimization-Based Combinatorial Assignment"
https://t.co/aPYiaMTOOB
video: summary: https://t.co/bpgM40asLu
@JWeissteiner @julien_siems @SvenSeuken
@UZH_ch @ETH_en @ETH_AI_Center @ERC_Research

[Jakob Heiss @JakobHeiss]

The Path-Dependent Neural Jump ODE (PD-NJ-ODE) is a #deeplearning model for learning optimal forecasts given irregularly sampled time series of incomplete past observations. We extend this method to deal with noise and statistical dependencies.

[Jakob Heiss @JakobHeiss]

Our new #TMLR paper on forecasting irregularly observed time series: "Extending Path-Dependent NJ-ODEs to Noisy Observations and a Dependent Observation Framework"
https://t.co/BraFsp8P7w
video summary: https://t.co/d0BwwUtVw3
William Andersson, @FloKrach, @JosefTeichmann
@ETH

[Jakob Heiss @JakobHeiss]

We also extended these results to Deep ReLU NNs where we train all parameters in all layers: https://t.co/8rPrEZmcbD

[Jakob Heiss @JakobHeiss]

Recently we've updated my first paper https://t.co/A0jGZb9MnR to provide more intuition on the "psychology" of wide random feature models.

[Jakob Heiss @JakobHeiss]

We extended these results to multi-dimensional inputs: https://t.co/zlDaFnDCDd

[Bertrand Charpentier @Bertrand_Charp]

Jakob Heiss will present his ICML paper "NOMU - Neural Optimization-based Model Uncertainty" at the Uncertainty in AI reading group today at 5:30pm (Berlin time).

https://t.co/WhU4wqjAmf

Co-author: @JWeissteiner, Hanna Wutte, @SvenSeuken, Josef Teichmann

[Sven Seuken @SvenSeuken]

If you are interested in preference learning in combinatorial domains, you might like our two new #IJCAI2022 papers: 👇

[Sven Seuken @SvenSeuken]

1. "Monotone-Value Neural Networks: Exploiting Preference Monotonicity in Combinatorial Assignment":
https://t.co/V3VL6Ak3ke

2. "Fourier Analysis-based Iterative Combinatorial Auctions": https://t.co/0t7alf8b9s

[Sven Seuken @SvenSeuken]

If you are at #IJCAI2022, you can see both talks, today, in the Multi-Agent-Systems session, from 3:30pm-3:42pm.

Joint work with: @JWeissteiner, @JakobHeiss, Chris Wendler, Ben Lubin, Julien Siems, and Markus Püschel.

@UZH_ch @ETH_en @ETH_AI_Center @ERC_Research

[Sven Seuken @SvenSeuken]

If you are at @icml2022, come to our spotlight talk: Thursday, 10:30-12:00 (Talk: 11:00-11:05), Session 8, Track 10, Room 310.

Presentation: https://t.co/EQ8JQH4JUS

Paper: https://t.co/s818sbtliA

Joint work with: Jakob Weissteiner, Hanna Wutte, Jakob Heiss, and Josef Teichmann

[Jakob Heiss @JakobHeiss]

neue Version vom Profizocker Kampfsimulator: http://t.co/sxHS7zxL

[Jakob Heiss @JakobHeiss]

@niKnight Natürlich! Tab-Panorma ist so geil! Jetzt muss ich doch nicht auf Opera umsteigen.

[Jakob Heiss @JakobHeiss]

Hab schon über 1200 Visits/Tag auf meinem DSO-Kampfsimulator: http://is.gd/0hcvZ4