Layerwise Development of Compositional Functional Representations Across Architectures

The project uses linear probes to investigate how functional representations are distributed across layers for MLPs and transformers trained on composite function tasks. The basic idea makes sense, and is clearly relevant to AI safety. Regarding the empirical finding that transformers learn positional embeddings more slowly than later layers, the authors may be interested in theoretical work that predicts this, and recommends using a larger learning rate for the positional embedding layer: https://arxiv.org/abs/2304.02034

However, I have a number of concerns about this project.

1) While this interpretability study is relevant to AI safety, the connection to physics is limited.

2) Several of the figures do not show what is claimed in the text or caption. For example, neither Fig. 7 nor Fig. 15 show any evidence of grokking, and Fig. 7 in fact shows hardly any learning at all. This makes me doubt the validity of the claimed results.

3) At 8 pages + appendices, the report exceeds the page limit of 6 pages + appendices.

AI Safety Relevance: While interpretability is an important subfield of AI Safety, and the authors seem to make a sincere effort towards contributing to the field, they fail to make contact with existing interpretability literature and current techniques, making many of their results moot. There is also no engagement with physics.

Innovation & Originality: Training models on ground truth composite functions and then training linear probes of each composite across model layers is a reasonable first thing to do if one is interested in interpretability. However, the field has significantly progressed beyond this and thus results, if they are to be trusted, offer little insight.

Technical Rigor & Feasibility: Most of the paper's plots show with incomplete information and at times information which contradicts what's written in the text, making it difficult to evaluate the claims.

In Figure 2 it is unclear what the unit of complexity (y axis) is. It is also unclear what the x axis is, Section 4.1 and caption underneath the figure claim that it’s across layers, but the axis label reads model width and the indices make sense for model width. There is also no mention of what the ground truth function which is being learned.

In Figure 3 Lines are overlaid on top of one another, it’s extremely difficult to understand what is being claimed. The favorable interpretation is that g(x) (blue) had perfect fit at layer 0 and f(g(x)) (brown) had perfect fit at layer 2, and g(x) had near 0 fit at layer 2 (green) and f(g(x)) had no fit at layer 0 (red). There is also no mention of how many layers the model which was trained has. Without knowing how many layers the trained model had, it’s unclear to know whether the results are trivial. There is also no interesting variation across the x axis, so it's unclear why the authors chose to include it.

In Figure 4 there is no mention of which layers the linear probe accuracy measurements are being performed. Additionally, since there seems to be no interesting variation across training it's unclear why the authors chose to include this.

Furthermore, the codebase which the authors have provided is difficult to parse. There is about 50 python scripts in a single folder. Granted they did provide the names of relevant files in the readme, but the lack of internal organization of the codebase makes quick sanity checks unnecessarily difficult.

My suggestion is for the authors to adopt cleaner code and paper writing practices, as well as closer engagement with relevant mech interp literature.

This project claims that their probes show that simple functions decompose across layers in an MLP but not in a transformer. The premise – comparing architectures in this way, changing task complexity, etc. – is good for a hackathon project. However, they don’t provide implementation details like how their datasets are constructed, or which architectures and datasets correspond to each plot. Similarly, the ‘7 axes of interpretability’ (I’m unsure that these are of the same type signature) are lightly evidenced. The phase transition is unmotivated, the authors seem to eyeball a heat-map threshold where probe accuracy jumps but never quantify this, so it seems like it was thrown in for some connection to physics. Many plots seem mislabeled, which adds to confusion (I think figure 2 should be something like layer index v. function, but it says complexity v. width). Finally: AI safety framing. Understanding modularity is good, but a model’s ability to do this is not correlated with performance or safety. The goal of interpretability is not necessarily to make models that are composable, but to understand models even when they are not.

The authors run a suite of interesting experiments to look for "compositionality" in trained neural networks. Their most striking finding is that deep fully-connected MLPs appear to learn the function x->sin(x^2) compositionally, with x^2 being linearly decodable in earlier layers but not sin(x^2). It seems like the authors may not have had time to clarify some of their results - for example my summary of the x->sin(x^2) result is inferred indirectly from their prose, while the result is not clearly depicted in figure 3.