Computer Vision News - March 2018

● Trainers are users who are interested in using known architectures in new fields in which they are experts. Trainers’ main objective is network optimization. They focus on examining how an existing architecture learns a new model. Trainers would include, for instance, bioinformatics or machine learning engineers. ● End users are the most common type. These users most often don’t need to understand the ‘guts’ of the training process in any kind of detail at all, only how to implement an existing network for a given problem with new data. Their main interest is in analyzing the results and locating the cause of any problem when something doesn’t go as expected. End users would include, for instance, data analysts or product engineers using machine learning methodologies. At this stage the LSTMVis system is mainly aimed at Architect- and Trainer-type users. Its goal is to answer the question “ what data do different hidden states of my RNN/LSTM represent? ” The system was developed based on a series of interviews with deep learning experts, going through several versions and iterations. In the end, the following goals for the system were identified: ● Goal 1 - Formulate a hypothesis concerning a certain quality of the hidden state that the user hypothesizes the model has learned. The hypothesis is tested by displaying the initial hidden state and tracing hidden state changes over time as a function of the original input. ● Goal 2 - Refine or reject the hypothesis. The former hypothesis, if it was not definitively confirmed, may be rejected, or the user may try to refine it, if possible. ● Goal 3 - Compare models and datasets. This goal serves to enable generalization of insights concerning the hidden state of the model and analyze how different tasks and domains may influence/change the hidden states. LSTMVis includes a five tasks development for achieving these three goals: ● Task 1 enables analysis of hidden state dynamics raw data. [Goal 1] ● Task 2 allows the user to filter results by selecting a specific text pattern and a threshold. This technique allows the user to formulate hypotheses and refine them, and to separate significant signal from background noise. [Goal 1 and 2] ● Task 3 searches the hidden state and outputs matching similar text patterns based on similarity in hidden state representation. The task should return text snippets that are in fact similar. [Goal 2] We Tried for You: RNNVis 11 Tool Computer Vision News