"Interactive Debugging and Testing Support for Deep Learning" by Tianyi Zhang
Recurrent Neural Networks (RNNs) have been widely used in NLP tasks. Yet it is challenging to debug RNNs due to their inherent complexity and opaqueness. To address this challenge, we present an interactive debugger that transforms an RNN model, which is complex and unfamiliar to regular developers, back to something they are familiar with----a Finite State Machine (FSM). The FSM provides a bird’s-eye view of the internal decision-making process of the RNN model. As the model reads each word in an input sentence, it will transit between different states until it reaches the end of the sentence. If a developer clicks on a state, they can see the frequent words and phrases associated with this state. In this way, we convert those high-dimensional arrays to symbolic values that are more interpretable to programmers. Given a misclassified text, our debugger will produce a state trace with intermediate decisions made by the RNN model. Similar to how we can step through a program, we can step through the states in the trace to inspect the decision-making process of the model.
Tianyi Zhang
Assistant Professor in Computer Science
Purdue University
Presented at the "It Will Never Work in Theory" miniconf at Strange Loop 2022. https://neverworkintheory.org/events/...
------- Sponsored by: -------
Stream is the # 1 Chat API for custom messaging apps. Activate your free 30-day trial to explore Stream Chat. https://gstrm.io/tsl
Watch video "Interactive Debugging and Testing Support for Deep Learning" by Tianyi Zhang online, duration hours minute second in high quality that is uploaded to the channel Strange Loop Conference 01 November 2022. Share the link to the video on social media so that your subscribers and friends will also watch this video. This video clip has been viewed 2,211 times and liked it like visitors.
