Neural Information Processing Systems http://nips.cc/

Neural Information Processing Systems http://nips.cc/

We propose a new technique, Singular V ector Canonical Correlation Analysis (SVCCA), a tool for quickly comparing two representations in a way that is both invariant to affine transform (allowing comparison between different layers and networks) and fast to compute (allowing more comparisons to be calculated than with previous methods). We deploy this tool to measure the intrinsic dimensionality of layers, showing in some cases needless over-parameterization; to probe learning dynamics throughout training, finding that networks converge to final representations from the bottom up; to show where class-specific information in networks is formed; and to suggest new training regimes that simultaneously save computation and overfit less.

We find that the first hidden layers are general, and that they learn generalized coordinates over the input domain. Deeper layers are successively more specific.

Neural Information Processing Systems http://nips.cc/

There is some knowledge from anatomy (Harris, et al, 2019).

Existing machine learning methods for molecular (e.g., gene) embeddings are restricted to specific tasks or data modalities, limiting their effectiveness within narrow domains. As a result, they fail to capture the full breadth of gene functions and interactions across diverse biological contexts. In this study, we have systematically evaluated knowledge representations of biomolecules across multiple dimensions representing a task-agnostic manner spanning three major data sources, including omics experimental data, literature-derived text data, and knowledge graph-based representations. To distinguish between meaningful biological signals from chance correlations, we devised an adjusted variant of Singular Vector Canonical Correlation Analysis (SVCCA) that quantifies signal redundancy and complementarity across different data modalities and sources. These analyses reveal that existing embeddings capture largely non-overlapping molecular signals, highlighting the value of embedding integration. Building on this insight, we propose Platform for Representation and Integration of mul-timodal Molecular Embeddings (PRISME), a machine learning based workflow using an autoencoder to integrate these heterogeneous embeddings into a unified multimodal representation. We validated this approach across various benchmark tasks, where PRISME demonstrated consistent performance, and outperformed individual embedding methods in missing value imputations. This new framework supports comprehensive modeling of biomolecules, advancing the development of robust, broadly applicable multimodal embeddings optimized for downstream biomedical machine learning applications.

The transferabilities of taskonomy have a practical value (they're constructed and are shown to reduce the need for supervision through transfer learning), but Taskonomy's method is computationally expensive. So, the gold standard is duplication of taskonomy's affinity matrix, but with less complexity. Therefore I see the comparison between the transferability matrix by attribution maps and taskonomy's (fig 4) valid and what the main point is. But I don't understand why/how SVCCA vs attribution map's similarity matrix comparisons (figure 3) are useful. What exactly is the value of SVCCA based similarity matrix? Why isn't figure 3 comparing between attribution map's matrix and Taskonomy's affinity matrix (after being made symmetric)?

We introduce a novel index, the Distribution of Cosine Similarity (DOCS), for quantitatively assessing the similarity between weight matrices in Large Language Models (LLMs), aiming to facilitate the analysis of their complex architectures. Leveraging DOCS, our analysis uncovers intriguing patterns in the latest opensource LLMs: adjacent layers frequently exhibit high weight similarity and tend to form clusters, suggesting depth-wise functional specialization. Additionally, we prove that DOCS is theoretically effective in quantifying similarity for orthogonal matrices, a crucial aspect given the prevalence of orthogonal initializations in LLMs. This research contributes to a deeper understanding of LLM architecture and behavior, offering tools with potential implications for developing more efficient and interpretable models. Large Language Models (LLMs), built on transformer architectures (Vaswani et al., 2017), have ushered in a new era in natural language processing (Brown et al., 2020). These complex models have demonstrated remarkable capabilities, but understanding their underlying mechanisms remains a challenge. Similarity analysis techniques (Raghu et al., 2017; Morcos et al., 2018; Kornblith et al., 2019) offer a promising approach for gaining insights into the learned representations and computational processes within these models. By analyzing the weights themselves, we aim to uncover deeper insights into the model's structure and functionality that are not apparent from representations alone. For further discussion, see Appendix E. 1. Focus on Representation, Not Weights: Similar representations across layers do not necessarily imply similar weight matrices. This discrepancy arises from the use of residual connections in transformer architectures (He et al., 2016), which create shortcuts that allow information to bypass layer transformations. Mathematically, a residual connection is represented as y = F(x, W) + x, (1) where x is the layer's input, W represents the weight matrices, F is the transformation function (including the feedforward network and attention), and y is the layer's output.