Explainable Inference on Sequential Data via Memory-Tracking

Authors: Biagio La Rosa, Roberto Capobianco, Daniele Nardi

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach (1) on a modified T-maze, which is a non-Markovian discrete control task evaluating an algorithm s ability to correlate events far apart in history, and (2) on the Story Cloze Test, which is a commonsense reasoning framework for evaluating story understanding that requires a system to choose the correct ending to a four-sentence story. Our results show that we are able to explain agent s decisions in (1) and to reconstruct the most relevant sentences used by the network to select the story ending in (2).
Researcher Affiliation Collaboration 1Sapienza University of Rome, Italy 2Sony AI
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks. It includes a system architecture sketch in Figure 1, but no step-by-step algorithms.
Open Source Code No The paper does not provide any statements or links indicating the availability of open-source code for the described methodology.
Open Datasets Yes The Story Cloze Test [Mostafazadeh et al., 2016] and Modified T-maze task [Bakker, 2002; Wierstra et al., 2007] are referenced, which are established public datasets/tasks.
Dataset Splits Yes The model is trained and tested on the official development and test dataset of the challenge, while we randomly split the former into a training and development set containing respectively 90% and 10% of the original dataset.
Hardware Specification No The paper does not specify the exact hardware components (e.g., CPU, GPU models, memory size) used for the experiments, only general training information like 'running the Deep Recurrent QLearning algorithm with Tensorflow 2.0'.
Software Dependencies Yes We train our network running the Deep Recurrent QLearning algorithm [Hausknecht and Stone, 2015] with Tensorflow 2.0... We finally train the network with Tensorflow 2.0 using the Adam optimizer...
Experiment Setup Yes In our experiments, we choose a fixed corridor length of N = 25 steps... The controller of the DNC is a single LSTM layer composed by 128 units, a memory of size 50, 1 read head, 1 write head, and a rate of 0.2 for the bypass dropout... using RMSProp with a learning rate of 1e 3 and a gradient clipping of 20.