Ralf Herbrich

Richard Kurle, Ralf Herbrich, Tim Januschowski et al. · amazon-science

Variational Bayesian posterior inference often requires simplifying approximations such as mean-field parametrisation to ensure tractability. However, prior work has associated the variational mean-field approximation for Bayesian neural networks with underfitting in the case of small datasets or large model sizes. In this work, we show that invariances in the likelihood function of over-parametrised models contribute to this phenomenon because these invariances complicate the structure of the posterior by introducing discrete and/or continuous modes which cannot be well approximated by Gaussian mean-field distributions. In particular, we show that the mean-field approximation has an additional gap in the evidence lower bound compared to a purpose-built posterior that takes into account the known invariances. Importantly, this invariance gap is not constant; it vanishes as the approximation reverts to the prior. We proceed by first considering translation invariances in a linear model with a single data point in detail. We show that, while the true posterior can be constructed from a mean-field parametrisation, this is achieved only if the objective function takes into account the invariance gap. Then, we transfer our analysis of the linear model to neural networks. Our analysis provides a framework for future work to explore solutions to the invariance problem.

Paul Mattes, Rainer Schlosser, Ralf Herbrich

One of the biggest challenges to modern deep reinforcement learning (DRL) algorithms is sample efficiency. Many approaches learn a world model in order to train an agent entirely in imagination, eliminating the need for direct environment interaction during training. However, these methods often suffer from either a lack of imagination accuracy, exploration capabilities, or runtime efficiency. We propose Hieros, a hierarchical policy that learns time abstracted world representations and imagines trajectories at multiple time scales in latent space. Hieros uses an S5 layer-based world model, which predicts next world states in parallel during training and iteratively during environment interaction. Due to the special properties of S5 layers, our method can train in parallel and predict next world states iteratively during imagination. This allows for more efficient training than RNN-based world models and more efficient imagination than Transformer-based world models. We show that our approach outperforms the state of the art in terms of mean and median normalized human score on the Atari 100k benchmark, and that our proposed world model is able to predict complex dynamics very accurately. We also show that Hieros displays superior exploration capabilities compared to existing approaches.

Ralf Herbrich, Rajeev Rastogi, Roland Vollgraf

We present CRISP (COVID-19 Risk Score Prediction), a probabilistic graphical model for COVID-19 infection spread through a population based on the SEIR model where we assume access to (1) mutual contacts between pairs of individuals across time across various channels (e.g., Bluetooth contact traces), as well as (2) test outcomes at given times for infection, exposure and immunity tests. Our micro-level model keeps track of the infection state for each individual at every point in time, ranging from susceptible, exposed, infectious to recovered. We develop both a Monte Carlo EM as well as a message passing algorithm to infer contact-channel specific infection transmission probabilities. Our Monte Carlo algorithm uses Gibbs sampling to draw samples of the latent infection status of each individual over the entire time period of analysis, given the latent infection status of all contacts and test outcome data. Experimental results with simulated data demonstrate our CRISP model can be parametrized by the reproduction factor $R_0$ and exhibits population-level infectiousness and recovery time series similar to those of the classical SEIR model. However, due to the individual contact data, this model allows fine grained control and inference for a wide range of COVID-19 mitigation and suppression policy measures. Moreover, the block-Gibbs sampling algorithm is able to support efficient testing in a test-trace-isolate approach to contain COVID-19 infection spread. To the best of our knowledge, this is the first model with efficient inference for COVID-19 infection spread based on individual-level contact data; most epidemic models are macro-level models that reason over entire populations. The implementation of CRISP is available in Python and C++ at https://github.com/zalandoresearch/CRISP.

Nicolas Alder, Kai Ebert, Ralf Herbrich et al.

This paper critically examines the AI Act's provisions on climate-related transparency, highlighting significant gaps and challenges in its implementation. We identify key shortcomings, including the exclusion of energy consumption during AI inference, the lack of coverage for indirect greenhouse gas emissions from AI applications, and the lack of standard reporting methodology. The paper proposes a novel interpretation to bring inference-related energy use back within the Act's scope and advocates for public access to climate-related disclosures to foster market accountability and public scrutiny. Cumulative server level energy reporting is recommended as the most suitable method. We also suggests broader policy changes, including sustainability risk assessments and renewable energy targets, to better address AI's environmental impact.

Romeo Sommerfeld, Christian Helms, Ralf Herbrich

Bayesian neural networks (BNNs) offer the potential for reliable uncertainty quantification and interpretability, which are critical for trustworthy AI in high-stakes domains. However, existing methods often struggle with issues such as overconfidence, hyperparameter sensitivity, and posterior collapse, leaving room for alternative approaches. In this work, we advance message passing (MP) for BNNs and present a novel framework that models the predictive posterior as a factor graph. To the best of our knowledge, our framework is the first MP method that handles convolutional neural networks and avoids double-counting training data, a limitation of previous MP methods that causes overconfidence. We evaluate our approach on CIFAR-10 with a convolutional neural network of roughly 890k parameters and find that it can compete with the SOTA baselines AdamW and IVON, even having an edge in terms of calibration. On synthetic data, we validate the uncertainty estimates and observe a strong correlation (0.9) between posterior credible intervals and its probability of covering the true data-generating function outside the training range. While our method scales to an MLP with 5.6 million parameters, further improvements are necessary to match the scale and performance of state-of-the-art variational inference methods.

Nicolas Alder, Shivam Nitin Kajale, Milin Tunsiricharoengul et al.

(Pseudo)random sampling, a costly yet widely used method in (probabilistic) machine learning and Markov Chain Monte Carlo algorithms, remains unfeasible on a truly large scale due to unmet computational requirements. We introduce an energy-efficient algorithm for uniform Float16 sampling, utilizing a room-temperature stochastic magnetic tunnel junction device to generate truly random floating-point numbers. By avoiding expensive symbolic computation and mapping physical phenomena directly to the statistical properties of the floating-point format and uniform distribution, our approach achieves a higher level of energy efficiency than the state-of-the-art Mersenne-Twister algorithm by a minimum factor of 9721 and an improvement factor of 5649 compared to the more energy-efficient PCG algorithm. Building on this sampling technique and hardware framework, we decompose arbitrary distributions into many non-overlapping approximative uniform distributions along with convolution and prior-likelihood operations, which allows us to sample from any 1D distribution without closed-form solutions. We provide measurements of the potential accumulated approximation errors, demonstrating the effectiveness of our method.

Richard Kurle, Alexej Klushyn, Ralf Herbrich

We introduce a new method for learning Bayesian neural networks, treating them as a stack of multivariate Bayesian linear regression models. The main idea is to infer the layerwise posterior exactly if we know the target outputs of each layer. We define these pseudo-targets as the layer outputs from the forward pass, updated by the backpropagated gradients of the objective function. The resulting layerwise posterior is a matrix-normal distribution with a Kronecker-factorized covariance matrix, which can be efficiently inverted. Our method extends to the stochastic mini-batch setting using an exponential moving average over natural-parameter terms, thus gradually forgetting older data. The method converges in few iterations and performs as well as or better than leading Bayesian neural network methods on various regression, classification, and out-of-distribution detection benchmarks.

Leo Kohlenberg, Leonard Horns, Frederic Sadrieh et al.

Annotating large datasets can be challenging. However, crowd-sourcing is often expensive and can lack quality, especially for non-trivial tasks. We propose a method of using LLMs as few-shot learners for annotating data in a complex natural language task where we learn a standalone model to predict usage options for products from customer reviews. We also propose a new evaluation metric for this scenario, HAMS4, that can be used to compare a set of strings with multiple reference sets. Learning a custom model offers individual control over energy efficiency and privacy measures compared to using the LLM directly for the sequence-to-sequence task. We compare this data annotation approach with other traditional methods and demonstrate how LLMs can enable considerable cost savings. We find that the quality of the resulting data exceeds the level attained by third-party vendor services and that GPT-4-generated labels even reach the level of domain experts. We make the code and generated labels publicly available.

Thore Graepel, Ralf Herbrich

Abstract We present PAC-Bayesian bounds for the generalisation error of the K-nearest-neighbour classifier (K-NN). This is achieved by casting the K-NN classifier into a kernel space framework in the limit of vanishing kernel bandwidth. We establish a relation between prior measures over the coefficients in the kernel expansion and the induced measure on the weight vectors in kernel space. Defining a sparse prior over the coefficients allows the application of a PAC-Bayesian folk theorem that leads to a generalisation bound that is a function of the number of redundant training examples: those that can be left out without changing the solution. The presented bound requires to quantify a prior belief in the sparseness of the solution and is evaluated after learning when the actual redundancy level is known. Even for small sample size (m ~ 100) the bound gives non-trivial results when both the expected sparseness and the actual redundancy are high.