Replies: 4 comments 3 replies
-
The visualization in "Factorizing Activations of a Single Layer" is stunning. Interpretation of the roles of activation factors makes sense. The layers are clusterings of neurons by activation value. One contribution of this work is to map neuron activation inside transformer-based language models; other work has looked at that for other deep neural networks. As a reader, I am not familiar with the foundations of this work - how Transformers work, how BERT works - so understanding the references is essential to understanding how the system implemented here actually works. The references here look really good and I appreciate the way the are presented. The explanation is quite parsimonious for a reader like me, but it's fair to say I am not the intended reader for this particular paper. It's clear there's much more to come. I am familiar with LIME and SHAP and the idea of an interpretable surrogate model. Is it fair to say that Ecco builds a surrogate view by interrogating a trained transformer model to build these clusters of activation values? How are input tokens mapped to to factors? You are looking at activation values for that specfic input? Are the clusters stable across many inputs? Input saliency reminds me of statistical sensitivity analysis. I am interested in policy simulation through Structural Causal Models with stochastic inputs. SCMs can be implemented as trained NN. Naively, transformers look like a generative, multi stage models. So, thats's something for me to learn more about. I wonder, though, if you see any potential application of this work to explaining the actions of agents in an RL setting, eg what factors of a policy or a sensed environment are "salient" in terms generating an observed system state? I could not render the notebooks via the web. Maybe that's a colab thing. It would be helpful if pdf renderings of the notebooks were available. |
Beta Was this translation helpful? Give feedback.
-
Hi Jay,
in one of your notebooks. Is there some documentation on how I can add my custom bert based model to ecco to visualize it. |
Beta Was this translation helpful? Give feedback.
-
Hi, I want to experiment with Input Saliency Maps, but the notebook is not accessible now. Can you help? |
Beta Was this translation helpful? Give feedback.
-
Do you have an idea for the reason why in NMF there always is a factor of neurons that intently focus on the first token in the sequence, and only on that token? |
Beta Was this translation helpful? Give feedback.
-
I would love your feedback on the article, the series, and on the library!
Beta Was this translation helpful? Give feedback.
All reactions