Identification of Knowledge Neurons in Protein Language Models
Authors:
Divya Nori,
Shivali Singireddy,
Marina Ten Have
Abstract:
Neural language models have become powerful tools for learning complex representations of entities in natural language processing tasks. However, their interpretability remains a significant challenge, particularly in domains like computational biology where trust in model predictions is crucial. In this work, we aim to enhance the interpretability of protein language models, specifically the stat…
▽ More
Neural language models have become powerful tools for learning complex representations of entities in natural language processing tasks. However, their interpretability remains a significant challenge, particularly in domains like computational biology where trust in model predictions is crucial. In this work, we aim to enhance the interpretability of protein language models, specifically the state-of-the-art ESM model, by identifying and characterizing knowledge neurons - components that express understanding of key information. After fine-tuning the ESM model for the task of enzyme sequence classification, we compare two knowledge neuron selection methods that preserve a subset of neurons from the original model. The two methods, activation-based and integrated gradient-based selection, consistently outperform a random baseline. In particular, these methods show that there is a high density of knowledge neurons in the key vector prediction networks of self-attention modules. Given that key vectors specialize in understanding different features of input sequences, these knowledge neurons could capture knowledge of different enzyme sequence motifs. In the future, the types of knowledge captured by each neuron could be characterized.
△ Less
Submitted 17 December, 2023;
originally announced December 2023.
Variability as a predictor for the hard-to-soft state transition in GX 339-4
Authors:
Matteo Lucchini,
Marina Ten Have,
**gyi Wang,
Jeroen Homan,
Erin Kara,
Oluwashina Adegoke,
Riley Connors,
Thomas Dauser,
Javier Garcia,
Guglielmo Mastroserio,
Adam Ingram,
Michiel van der Klis,
Ole König,
Collin Lewin,
Labani Mallick,
Edward Nathan,
Patrick O'Neill,
Christos Panagiotou,
Joanna Piotrowska,
Phil Uttley
Abstract:
During the outbursts of black hole X-ray binaries (BHXRBs), their accretion flows transition through several states. The source luminosity rises in the hard state, dominated by non-thermal emission, before transitioning to the blackbody-dominated soft state. As the luminosity decreases, the source transitions back into the hard state and fades to quiescence. This picture does not always hold, as…
▽ More
During the outbursts of black hole X-ray binaries (BHXRBs), their accretion flows transition through several states. The source luminosity rises in the hard state, dominated by non-thermal emission, before transitioning to the blackbody-dominated soft state. As the luminosity decreases, the source transitions back into the hard state and fades to quiescence. This picture does not always hold, as $\approx$ 40$\%$ of the outbursts never leave the hard state. Identifying the physics that govern state transitions remains one of the outstanding open questions in black hole astrophysics. In this paper we present an analysis of archival RXTE data of multiple outbursts of GX 339-4. We compare the properties of the X-ray variability and time-averaged energy spectrum and demonstrate that the variability (quantified by the power spectral hue) systematically evolves $\approx$ 10-40 days ahead of the canonical state transition (quantified by a change in spectral hardness); no such evolution is found in hard state only outbursts. This indicates that the X-ray variability can be used to predict if and when the hard-to-soft state transition will occur. Finally, we find a similar behavior in ten outbursts of four additional BHXRBs with more sparse observational coverage. Based on these findings, we suggest that state transitions in BHXRBs might be driven by a change in the turbulence in the outer regions of the disk, leading to a dramatic change in variability. This change is only seen in the spectrum days to weeks later, as the fluctuations propagate inwards towards the corona.
△ Less
Submitted 11 October, 2023;
originally announced October 2023.
