ELLIS / ELIZA Social, June 3rd 2024

Date: June 3rd, 17:30
Location: Building 80 Room 00-021

The unit invites everyone interested to the next social event on June 3rd! Get ready for an engaging gathering filled with insightful discussions. A great event to connect with the AI community in Freiburg!

This time, there are two ten-minute presentations.

  1. Tom Viering, Assistant Professor at TU-Delft, who is currently visiting the ML lab, will present a talk on “Surprising Learning Curves and Where to Find Them.”
  2. Sai Prasanna will be presenting a talk on his recent paper accepted at the RLC conference, “Dreaming of Many Worlds: Learning Contextual World Models Aids Zero-Shot Generalization.”

The talks are followed by a BBQ, providing the perfect networking opportunity! In case the weather turns bad the BBQ will be substituted by a board game night at the ML lab instead.

[ELLIS Meetup] Multi-objective Differentiable Neural Architecture Search

Speaker: Arber Zela
Date: 29th April 2024

Pareto front profiling in multi-objective optimization (MOO), i.e. finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives like neural network training. Typically, in MOO neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences for the trade-off between performance and hardware metrics, and yields representative and diverse architectures across multiple devices in just one search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments with up to 19 hardware devices and 3 objectives showcase the effectiveness and scalability of our method. Finally, we show that, without extra costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only transformer space for language modelling.

[ELLIS Meetup] Efficient Hyperparameter Optimization in the Age of Deep Learning

Speaker: Josif Grabocka
Date: 2022-09-23 

Josif Grabocka from the Representation Learning Research Group at Uni Freiburg will give a talk titled “Efficient Hyperparameter Optimization in the Age of Deep Learning“.


Hyperparameter Optimization (HPO) is essential in designing state-of-the-art Machine Learning systems. Unfortunately, existing HPO techniques are not efficient enough for Deep Learning, due to the high cost of evaluating multiple hyperparameter configurations in complex Deep Learning systems. In this talk, I will present three approaches for scaling HPO for Deep Learning along the lines of transfer learning, gray-box HPO, and pipeline optimization with pre-trained deep networks. The talk will end with a demonstration of the capability of HPO to achieve state-of-the-art predictive results on tabular datasets.

[ELLIS Meetup] Prior-data Fitted Networks (PFNs)

Speaker: Samuel Müller
Date: 2022-09-09

Location: Building 101, Room 00.036

Samuel Müller from the Machine Learning Lab will give a talk titled “Prior-data Fitted Networks (PFNs): Use neural networks for 100x faster Bayesian predictions“.


Bayesian methods can be expensive and complicated to approximate with e.g. Markov Chain Monte Carlo (MCMC) or Variational Inference (VI) approaches. Prior-data Fitted Networks (PFNs) are a new, cheap and simple method to accurately approximate Bayesian predictions. I will explain how to build a PFN out of a Transformer by learning to model artificial data. I present the results from our paper that introduces PFNs, in which PFNs beat VI and MCMC for some standard tasks. As well as some more recent results with our new TabPFN, where we show that a simple PFN can replace a full AutoML tool for small datasets.

[ELLIS Meetup] Learning to solve arbitrary mobile manipulation tasks

Speaker: Daniel Honerkamp
Date: 2022-08-26


Mobile manipulation remains a critical challenge across both service and industrial settings and is a key component to visions such as household assistants. But for this it requires the combination of a wide range of capabilities, such as perception and exploration in unknown environments while controlling large, continuous action spaces for simultaneous navigation and manipulation.
In this talk I will first provide an overview of the main challenges and current benchmarks. I will then summarize and walk through the pipelines of current state-of-the-art approaches with a focus on both the high-level task learning and the low-level motion execution on robotic agents. Lastly, I will discuss potential paths forward and ways to integrate these low- and high-level components.

[ELLIS Meetup] Introduction to Vision-and-Language Navigation

Speaker: Chenguang Huang
Date: 2022-08-12


Robot navigation has been an actively researched challenge for decades. With different focuses, navigation goals can be defined in many different ways such as goal images, relative positions, semantic objects, and sound sources. Recently, with the advances in both computer vision and natural language processing methods, researchers propose to define the navigation goal with natural language instructions and define the task as Vision-and-Language Navigation. It is the very task that requires an agent to combine natural language instructions with visual perception data and navigate in an unseen environment. In this talk, I will give an overview of the Vision-and-Language Navigation tasks and introduce the current state of the research. In the end, I will discuss some future research directions.

[ELLIS Meetup] Robust Deep Neural Networks for Computer Vision

Speaker: Adam Kortylewski
Date: 2022-07-29


Deep learning sparked a tremendous increase in the performance of computer vision systems over the past decade. However, Deep Neural Networks (DNNs) are still far from reaching human-level performance at visual recognition tasks. The most important limitation of DNNs is that they fail to give reliable predictions in unseen or adverse viewing conditions, which would not fool a human observer, such as when objects are partially occluded, seen in an unusual pose or context, or in bad weather. This lack of robustness in DNNs is generally acknowledged, but the problem largely remains unsolved.
In this talk, I will give an overview of the principles underlying my work on building robust deep neural networks for computer vision. My working hypothesis is that vision systems need a causal 3D understanding images by following an analysis-by-synthesis approach. I will discuss a new type of neural network architecture that implements such an approach, and I will show that these generative neural network models are vastly superior to traditional models in terms of robustness, learning efficiency and because they can solve many vision tasks at once. Finally, I will give a brief outlook on current projects of mine and future research directions.

[ELLIS Meetup] Robots Learning (Through) Interactions

Speaker: Jens Kober
Date: 2022-06-24


The acquisition and self-improvement of novel motor skills is among the most important problems in robotics. I will discuss various learning techniques we developed that enable robots to have complex interactions with their environment and humans. Complexity arises from dealing with high-dimensional input data, non-linear dynamics in general and contacts in particular, multiple reference frames, and variability in objects, environments, and tasks. A human teacher is always involved in the learning process, either directly (providing data) or indirectly (designing the optimization criterion), which raises the question: How to best make use of the interactions with the human teacher to render the learning process efficient and effective? I will discuss various methods we have developed in the fields of supervised learning, imitation learning, reinforcement learning, and interactive learning. All these concepts will be illustrated with benchmark tasks and real robot experiments ranging from fun (ball-in-a-cup) to more applied (sorting products).

[ELLIS Meetup] Deep Learning 2.0

Speaker: Frank Hutter
Date: 2022-05-05


Deep Learning (DL) has been able to revolutionize learning from raw data (images, text, speech, etc) by replacing domain-specific hand-crafted features with features that are jointly learned for the particular task at hand. I propose to take deep learning to the next level, by also jointly (meta-)learning other, currently hand-crafted, elements of the deep learning pipeline: neural architectures and their initializations, training pipelines and their hyperparameters, self-supervised learning pipelines, etc.

I dub this new research direction Deep Learning 2.0 (DL 2.0), since

  • Like deep learning, it replaces previous hand-crafted solutions with learned solutions (but now on the meta-level),
  • It allows for a qualitatively new level of automated customization to the task at hand, and
  • It allows for automatically finding meta-level tradeoffs with the objectives specified by the user, such as algorithmic fairness, interpretability, calibration of uncertainty estimates, robustness, energy consumption, etc. It can thus achieve Trustworthy AI by design, breaking qualitatively new ground for deep learning.

[ELLIS Meetup] Neural Ensemble Search

Speaker: Arbër Zela
Date: 2022-04-15


Ensembles of neural networks achieve superior performance compared to standalone networks not only in terms of accuracy on in-distribution data but also on data with distributional shift alongside improved uncertainty calibration. Diversity among networks in an ensemble is believed to be key for building strong ensembles, but typical approaches only ensemble different weight vectors of a fixed architecture. Instead, we investigate neural architecture search (NAS) for explicitly constructing ensembles to exploit diversity among networks of varying architectures and to achieve robustness against distributional shift. By directly optimizing ensemble performance, our methods implicitly encourage diversity among networks, without the need to explicitly define diversity. We find that the resulting ensembles are more diverse compared to ensembles composed of a fixed architecture and are therefore also more powerful. We show significant improvements in ensemble performance on image classification tasks both for in-distribution data and during distributional shift with better uncertainty calibration.