Searcharxiv⌕ Search

arXiv subjects

Aaron Maiwald

Publications and source records attributed to Aaron Maiwald.

5 recordsLinked to original sources

Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety

Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.

cs.CL↗

The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models

Pretraining filtering is one of the most common safeguards for open-weight language models. Here we study its effectiveness, durability and scalability using the ESM-2 family of protein language models. We find that filtering viral proteins from the pretraining data meaningfully degrades masked language modelling, mutation effect prediction and long-range contact prediction on viral proteins, while largely preserving performance on non-viral proteins. These effects are concentrated among proteins from human-infecting viruses, and grow as we scale models from 8M to 3B parameters. As models increase in size, we find that stringent filtering becomes more effective. However, fine-tuning on sequences from the removed viral families restores the lost performance. Although the cost to recover performance increases with model scale, even for the largest models it costs about 80 USD of H100 time. This work establishes pretraining filtering as moderately effective at selectively degrading performance on viral proteins, but not durable against fine-tuning with modest computational resources and access to relevant data.

cs.LG↗

Open-weight genome language model safeguards: Assessing robustness via adversarial fine-tuning

Novel deep learning architectures are increasingly being applied to biological data, including genetic sequences. These models, referred to as genomic language models (gLMs), have demonstrated impressive predictive and generative capabilities, raising concerns that such models may also enable misuse, for instance via the generation of genomes for human-infecting viruses. These concerns have catalyzed calls for risk mitigation measures. The de facto mitigation of choice is filtering of pretraining data (i.e., removing viral genomic sequences from training datasets) in order to limit gLM performance on virus-related tasks. However, it is not currently known how robust this approach is for securing open-source models that can be fine-tuned using sensitive pathogen data. Here, we evaluate a state-of-the-art gLM, Evo 2, and perform fine-tuning using sequences from 110 harmful human-infecting viruses to assess the rescue of misuse-relevant predictive capabilities. The fine-tuned model exhibited reduced perplexity on unseen viral sequences relative to 1) the pretrained model and 2) a version fine-tuned on bacteriophage sequences. The model fine-tuned on human-infecting viruses also identified immune escape variants from SARS-CoV-2 (achieving an AUROC of 0.6), despite having no exposure to SARS-CoV-2 sequences during fine-tuning. This work demonstrates that data exclusion might be circumvented by fine-tuning approaches that can, to some degree, rescue misuse-relevant capabilities of gLMs. We highlight the need for safety frameworks for gLMs and outline further work needed on evaluations and mitigation measures to enable the safe deployment of gLMs.

cs.LG↗

Securing Dual-Use Pathogen Data of Concern

Training data is an essential input into creating competent artificial intelligence (AI) models. AI models for biology are trained on large volumes of data, including data related to biological sequences, structures, images, and functions. The type of data used to train a model is intimately tied to the capabilities it ultimately possesses--including those of biosecurity concern. For this reason, an international group of more than 100 researchers at the recent 50th anniversary Asilomar Conference endorsed data controls to prevent the use of AI for harmful applications such as bioweapons development. To help design such controls, we introduce a five-tier Biosecurity Data Level (BDL) framework for categorizing pathogen data. Each level contains specific data types, based on their expected ability to contribute to capabilities of concern when used to train AI models. For each BDL tier, we propose technical restrictions appropriate to its level of risk. Finally, we outline a novel governance framework for newly created dual-use pathogen data. In a world with widely accessible computational and coding resources, data controls may be among the most high-leverage interventions available to reduce the proliferation of concerning biological AI capabilities.

cs.AI↗

Image-based Data Representations of Time Series: A Comparative Analysis in EEG Artifact Detection

Alternative data representations are powerful tools that augment the performance of downstream models. However, there is an abundance of such representations within the machine learning toolbox, and the field lacks a comparative understanding of the suitability of each representation method. In this paper, we propose artifact detection and classification within EEG data as a testbed for profiling image-based data representations of time series data. We then evaluate eleven popular deep learning architectures on each of six commonly-used representation methods. We find that, while the choice of representation entails a choice within the tradeoff between bias and variance, certain representations are practically more effective in highlighting features which increase the signal-to-noise ratio of the data. We present our results on EEG data, and open-source our testing framework to enable future comparative analyses in this vein.

eess.SP↗