Sensitivity as a Complexity Measure for Sequence Classification Tasks

نویسندگان

چکیده

Abstract We introduce a theoretical framework for understanding and predicting the complexity of sequence classification tasks, using novel extension theory Boolean function sensitivity. The sensitivity function, given distribution over input sequences, quantifies number disjoint subsets that can each be individually changed to change output. argue standard methods are biased towards learning low-sensitivity functions, so tasks requiring high more difficult. To end, we show analytically simple lexical classifiers only express functions bounded sensitivity, empirically easier learn LSTMs. then estimate on 15 NLP finding is higher challenging collected in GLUE than text predicts performance both vanilla BiLSTMs without pretrained contextualized embeddings. Within task, which inputs hard such models. Our results suggest success massively contextual representations stems part because they provide from information extracted by decoders.

برای دانلود رایگان متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

Wire length as a circuit complexity measure

We introduce wire length as a salient complexity measure for analyzing the circuit complexity of sensory processing in biological neural systems. This new complexity measure is applied in this paper to two basic computational problems that arise in translationand scale-invariant pattern recognition, and hence appear to be useful as benchmark problems for sensory processing. We present new circu...

متن کامل

Information Measure as Time Complexity

The concept of Shannon entropy of random variables was generalized to measurable functions in general, and to simple functions with finite values in particular. It is shown that the information measure of a function is related to the time complexity of search problems concerning the function in question. Formally, given a Turing reduction from a search problem y x f  ) ( to another search prob...

متن کامل

Bidirectional Sequence Classification for Tagging Tasks with Guided Learning

In this paper we present a series of adaptations of the Guided Learning framework to solve different tagging tasks. The specificity of the proposed system lies in its ability to learn the order of inference together with the parameters of the local classifier instead of forcing it into a pre-defined order (left-to-right). The training algorithm is based on the Perceptron Algorithm. We apply the...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

ژورنال

عنوان ژورنال: Transactions of the Association for Computational Linguistics

سال: 2021

ISSN: ['2307-387X']

DOI: https://doi.org/10.1162/tacl_a_00403