Interestingness as an Inductive Heuristic for Future Compression Progress
Interestingness is an essential ingredient for open-ended learning and artificial general intelligence. It can be described as a heuristic assessing the pote...
Selected Research Projects
Multiple Token Divergence—Measuring and Steering In-Context Computation Density
The divergence between the prediction distributions of the full model and a computational shortcut allows us to measure and steer the computational effort of...
Measuring In-Context Computation Complexity via Hidden State Prediction
Can we measure when to what degree a sequence model or LLM is doing something interesting? We propose hidden state unpredictability as a measure, and show th...
Automatic Album Sequencing
We present a tool that automatically sequences an arbitrary collection of music tracks into multiple coherent orders. The underlying model was trained on man...
On the Distillation of Stories for Transferring Narrative Arcs in Collections of Independent Media
The act of telling stories is a fundamental part of what it means to be human. This work introduces the concept of narrative information, which we define to ...
Learning useful representations of recurrent neural network weight matrices
Recurrent Neural Networks are general-purpose computers. The program of an RNN is its weight matrix. How to learn useful representations of RNN weights that ...
Unsupervised Musical Object Discovery from Audio
Our novel MusicSlots method adapts SlotAttention to the audio domain, to achieve unsupervised music decomposition.
Mindstorms in natural language-based societies of mind
We imagine societies of mind as consisting of a collection of multimodal neural networks, including large language models, which engage in a “mindstorm” to s...
Goal-conditioned generators of deep policies
We study goal-conditioned neural nets (NNs) that learn to generate deep NN policies in form of context-specific weight matrices, similar to Fast Weight Progr...
Learning one abstract bit at a time through self-invented experiments encoded as neural networks
There are two important things in science–finding answers to given questions, and coming up with good questions. Our artificial scientists not only learn to ...
Generative Transformer-based Models of Symbolic Polyphonic Music
Master thesis on generating classical music with the help of transformers. Written at the Bosch Center for Artificial Intelligence.
Immersions - Visualizing and sonifying how an artificial ear hears music
Real-time exploration of a sound processing neural network.
Some Wavelet Visualizations
Recently I started to learn how to use d3.js, a JavaScript library for interactive data-driven visualizations. As a first little project, I decided to make i...
Wasserstein GAN and the Kantorovich-Rubinstein Duality
Derivation of the Kantorovich-Rubinstein duality for the use in Wasserstein Generative Adversarial Networks
Wavelets II - Vanishing Moments and Spectral Factorization
Explanation of the most important properties of Daubechies wavelets and the algorithm to calculate them
Wavelets I - From Filter Banks to the Dilation Equation
Derivation of the dilation and wavelet equation from an implementation of the Fast Wavelet Transform
Werden Roboter jemals so werden wie wir? (German)
Dies ist ein kurzes Essay, das ich im Frühjahr 2016 zu einer vorgegebenen Fragestellung geschrieben habe: Werden Roboter jemals so werden wie wir? Und wolle...