Yannic Kilcher

Andreas Köpf, Yannic Kilcher, Dimitri von Rütte et al.

Aligning large language models (LLMs) with human preferences has proven to drastically improve usability and has driven rapid adoption as demonstrated by ChatGPT. Alignment techniques such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) greatly reduce the required skill and domain knowledge to effectively harness the capabilities of LLMs, increasing their accessibility and utility across various domains. However, state-of-the-art alignment techniques like RLHF rely on high-quality human feedback data, which is expensive to create and often remains proprietary. In an effort to democratize research on large-scale alignment, we release OpenAssistant Conversations, a human-generated, human-annotated assistant-style conversation corpus consisting of 161,443 messages in 35 different languages, annotated with 461,292 quality ratings, resulting in over 10,000 complete and fully annotated conversation trees. The corpus is a product of a worldwide crowd-sourcing effort involving over 13,500 volunteers. Models trained on OpenAssistant Conversations show consistent improvements on standard benchmarks over respective base models. We release our code and data under a fully permissive licence.

Dimitri von Rütte, Luca Biggio, Yannic Kilcher et al.

Generating music with deep neural networks has been an area of active research in recent years. While the quality of generated samples has been steadily increasing, most methods are only able to exert minimal control over the generated sequence, if any. We propose the self-supervised description-to-sequence task, which allows for fine-grained controllable generation on a global level. We do so by extracting high-level features about the target sequence and learning the conditional distribution of sequences given the corresponding high-level description in a sequence-to-sequence modelling setup. We train FIGARO (FIne-grained music Generation via Attention-based, RObust control) by applying description-to-sequence modelling to symbolic music. By combining learned high level features with domain knowledge, which acts as a strong inductive bias, the model achieves state-of-the-art results in controllable symbolic music generation and generalizes well beyond the training distribution.

Leonard Adolphs, Benjamin Boerschinger, Christian Buck et al.

This paper presents first successful steps in designing search agents that learn meta-strategies for iterative query refinement in information-seeking tasks. Our approach uses machine reading to guide the selection of refinement terms from aggregated search results. Agents are then empowered with simple but effective search operators to exert fine-grained and transparent control over queries and search results. We develop a novel way of generating synthetic search sessions, which leverages the power of transformer-based language models through (self-)supervised learning. We also present a reinforcement learning agent with dynamically constrained actions that learns interactive search strategies from scratch. Our search agents obtain retrieval and answer quality performance comparable to recent neural methods, using only a traditional term-based BM25 ranking function and interpretable discrete reranking and filtering actions.

Paulina Grnarova, Yannic Kilcher, Kfir Y. Levy et al.

Many applications in machine learning can be framed as minimization problems and solved efficiently using gradient-based techniques. However, recent applications of generative models, particularly GANs, have triggered interest in solving min-max games for which standard optimization techniques are often not suitable. Among known problems experienced by practitioners is the lack of convergence guarantees or convergence to a non-optimum cycle. At the heart of these problems is the min-max structure of the GAN objective which creates non-trivial dependencies between the players. We propose to address this problem by optimizing a different objective that circumvents the min-max structure using the notion of duality gap from game theory. We provide novel convergence guarantees on this objective and demonstrate why the obtained limit point solves the problem better than known techniques.

Surya Kant Sahu, Abhinav Java, Arshad Shaikh et al.

Benford's Law (BL) or the Significant Digit Law defines the probability distribution of the first digit of numerical values in a data sample. This Law is observed in many naturally occurring datasets. It can be seen as a measure of naturalness of a given distribution and finds its application in areas like anomaly and fraud detection. In this work, we address the following question: Is the distribution of the Neural Network parameters related to the network's generalization capability? To that end, we first define a metric, MLH (Model Enthalpy), that measures the closeness of a set of numbers to Benford's Law and we show empirically that it is a strong predictor of Validation Accuracy. Second, we use MLH as an alternative to Validation Accuracy for Early Stopping, removing the need for a Validation set. We provide experimental evidence that even if the optimal size of the validation set is known before-hand, the peak test accuracy attained is lower than not using a validation set at all. Finally, we investigate the connection of BL to Free Energy Principle and First Law of Thermodynamics, showing that MLH is a component of the internal energy of the learning system and optimization as an analogy to minimizing the total energy to attain equilibrium.

Benjamin Borschinger, Jordan Boyd-Graber, Christian Buck et al.

We investigate a framework for machine reading, inspired by real world information-seeking problems, where a meta question answering system interacts with a black box environment. The environment encapsulates a competitive machine reader based on BERT, providing candidate answers to questions, and possibly some context. To validate the realism of our formulation, we ask humans to play the role of a meta-answerer. With just a small snippet of text around an answer, humans can outperform the machine reader, improving recall. Similarly, a simple machine meta-answerer outperforms the environment, improving both precision and recall on the Natural Questions dataset. The system relies on joint training of answer scoring and the selection of conditioning information.

Kevin Roth, Yannic Kilcher, Thomas Hofmann

We establish a theoretical link between adversarial training and operator norm regularization for deep neural networks. Specifically, we prove that $\ell_p$-norm constrained projected gradient ascent based adversarial training with an $\ell_q$-norm loss on the logits of clean and perturbed inputs is equivalent to data-dependent (p, q) operator norm regularization. This fundamental connection confirms the long-standing argument that a network's sensitivity to adversarial examples is tied to its spectral properties and hints at novel ways to robustify and defend against adversarial attacks. We provide extensive empirical evidence on state-of-the-art network architectures to support our theoretical results.

Kevin Roth, Yannic Kilcher, Thomas Hofmann

We investigate conditions under which test statistics exist that can reliably detect examples, which have been adversarially manipulated in a white-box attack. These statistics can be easily computed and calibrated by randomly corrupting inputs. They exploit certain anomalies that adversarial attacks introduce, in particular if they follow the paradigm of choosing perturbations optimally under p-norm constraints. Access to the log-odds is the only requirement to defend models. We justify our approach empirically, but also provide conditions under which detectability via the suggested test statistics is guaranteed to be effective. In our experiments, we show that it is even possible to correct test time predictions for adversarial attacks with high accuracy.

Yannic Kilcher, Thomas Hofmann

Black-Box attacks on machine learning models occur when an attacker, despite having no access to the inner workings of a model, can successfully craft an attack by means of model theft. The attacker will train an own substitute model that mimics the model to be attacked. The substitute can then be used to design attacks against the original model, for example by means of adversarial samples. We put ourselves in the shoes of the defender and present a method that can successfully avoid model theft by mounting a counter-attack. Specifically, to any incoming query, we slightly perturb our output label distribution in a way that makes substitute training infeasible. We demonstrate that the perturbation does not affect the ordinary use of our model, but results in an effective defense against attacks based on model theft.

Yannic Kilcher, Gary Becigneul, Thomas Hofmann

It is commonly agreed that the use of relevant invariances as a good statistical bias is important in machine-learning. However, most approaches that explicitly incorporate invariances into a model architecture only make use of very simple transformations, such as translations and rotations. Hence, there is a need for methods to model and extract richer transformations that capture much higher-level invariances. To that end, we introduce a tool allowing to parametrize the set of filters of a trained convolutional neural network with the latent space of a generative adversarial network. We then show that the method can capture highly non-linear invariances of the data by visualizing their effect in the data space.

Yannic Kilcher, Aurelien Lucchi, Thomas Hofmann

We consider the problem of training generative models with deep neural networks as generators, i.e. to map latent codes to data points. Whereas the dominant paradigm combines simple priors over codes with complex deterministic models, we argue that it might be advantageous to use more flexible code distributions. We demonstrate how these distributions can be induced directly from the data. The benefits include: more powerful generative models, better modeling of latent structure and explicit control of the degree of generalization.

Yannic Kilcher, Aurelien Lucchi, Thomas Hofmann

In implicit models, one often interpolates between sampled points in latent space. As we show in this paper, care needs to be taken to match-up the distributional assumptions on code vectors with the geometry of the interpolating paths. Otherwise, typical assumptions about the quality and semantics of in-between points may not be justified. Based on our analysis we propose to modify the prior code distribution to put significantly more probability mass closer to the origin. As a result, linear interpolation paths are not only shortest paths, but they are also guaranteed to pass through high-density regions, irrespective of the dimensionality of the latent space. Experiments on standard benchmark image datasets demonstrate clear visual improvements in the quality of the generated samples and exhibit more meaningful interpolation paths.

Yannic Kilcher, Aurélien Lucchi, Thomas Hofmann

We consider the problem of training generative models with deep neural networks as generators, i.e. to map latent codes to data points. Whereas the dominant paradigm combines simple priors over codes with complex deterministic models, we propose instead to use more flexible code distributions. These distributions are estimated non-parametrically by reversing the generator map during training. The benefits include: more powerful generative models, better modeling of latent structure and explicit control of the degree of generalization.

Gabriel Krummenacher, Brian McWilliams, Yannic Kilcher et al.

Adaptive stochastic gradient methods such as AdaGrad have gained popularity in particular for training deep neural networks. The most commonly used and studied variant maintains a diagonal matrix approximation to second order information by accumulating past gradients which are used to tune the step size adaptively. In certain situations the full-matrix variant of AdaGrad is expected to attain better performance, however in high dimensions it is computationally impractical. We present Ada-LR and RadaGrad two computationally efficient approximations to full-matrix AdaGrad based on randomized dimensionality reduction. They are able to capture dependencies between features and achieve similar performance to full-matrix AdaGrad but at a much smaller computational cost. We show that the regret of Ada-LR is close to the regret of full-matrix AdaGrad which can have an up-to exponentially smaller dependence on the dimension than the diagonal variant. Empirically, we show that Ada-LR and RadaGrad perform similarly to full-matrix AdaGrad. On the task of training convolutional neural networks as well as recurrent neural networks, RadaGrad achieves faster convergence than diagonal AdaGrad.