Thomas C. Bachlechner

Thomas C. Bachlechner, Mario Martone, Marjorie Schillo

Accurately evaluating the similarity of object vector embeddings is of critical importance for natural language processing, information retrieval and classification tasks. Popular similarity scores (e.g cosine similarity) are based on pairs of embedding vectors and disregard the distribution of the ensemble from which objects are drawn. Human perception of object similarity significantly depends on the context in which the objects appear. In this work we propose the $\textit{surprise score}$, an ensemble-normalized similarity metric that encapsulates the contrast effect of human perception and significantly improves the classification performance on zero- and few-shot document classification tasks. This score quantifies the surprise to find a given similarity between two elements relative to the pairwise ensemble similarities. We evaluate this metric on zero/few shot classification and clustering tasks and typically find 10-15 % better performance compared to raw cosine similarity. Our code is available at https://github.com/MeetElise/surprise-similarity.

Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao et al.

Deep networks often suffer from vanishing or exploding gradients due to inefficient signal propagation, leading to long training times or convergence difficulties. Various architecture designs, sophisticated residual-style networks, and initialization schemes have been shown to improve deep signal propagation. Recently, Pennington et al. used free probability theory to show that dynamical isometry plays an integral role in efficient deep learning. We show that the simplest architecture change of gating each residual connection using a single zero-initialized parameter satisfies initial dynamical isometry and outperforms more complex approaches. Although much simpler than its predecessors, this gate enables training thousands of fully connected layers with fast convergence and better test performance for ResNets trained on CIFAR-10. We apply this technique to language modeling and find that we can easily train 120-layer Transformers. When applied to 12 layer Transformers, it converges 56% faster on enwiki8.