EESTech Challenge 2020 (Athens Local Round) Winners
The EESTech Challenge is a yearly competition organized by EESTEC. It aims to provide opportunities for engineering students to gain knowledge in a specific popular technical field. The competition consists of two levels: several Local Rounds organized by the local branches and the Final Round, which will take place in Krakow, Poland, where the winners of the local rounds will compete with each other.
The Athens Local Round took place from 11 AM May 30 to 08 PM May 31, 2020, remotely for all teams, and was hosted at Kaggle, the well-known Machine Learning competition platform. The topic of the 2020 EESTech challenge was "Human-Computer Interaction - Chatbots" with a particular aim to Assistive Technologies.
The competition's setting
A person needing assistance is at home. He issues simple speech commands such as "turn on the lights," or "turn off the heating at the bathroom." A chatbot - personal assistant is supposed to understand these commands and map them to specific intents. Multiple phrasings can map to the same intent. For example, "It is too loud" and "Turn the volume down" both map to the intent "volume, down."
A Machine Learning approach
The audio files of the speech command competition's dataset have different durations. Neural Networks usually require inputs of fixed size. The organizers transformed the audio files to fixed-size images with Mel Spectrograms. The following is an example of a Mel Spectrogram of a voice command.
Following this transformation, a research team can treat the problem as a Computer Vision task, employing ConvNet architectures, but since Mel Spectrograms represent time sequences, state-of-the-art models also use Recurrent Neural Networks such as LSTMs.
Thirty students from NTUA, AUEB, and NKUA formed ten programming teams and participated in the contest trying to solve the Challenge.
This year's winning team is "TsiliCoffeeShop," consisting of Alexandros Kirkitsos, Konstantinos Mavrogiannis, Marios Mertzanidis, students of the School of ECE-NTUA. The first-place winners used Xception, a powerful 36-layers deep CNN with residual connections between layers and 22,855,952 parameters, pretrained on ImageNet.
Second-place winners "NLP (Name Lost Permanently)" implemented a small-footprint model with bi-directional LSTMs on top of a dense attention layer.
Third-place winners "theCage" based their model on a medium-sized ConvNet, with two bi-directional GRU layers before the classification head (a CRNN).
The top teams achieved an accuracy score of about 85% in less than 36 hours, only 10% shy of the most performing systems in the literature.
The winning team will participate in the Final Round in Krakow, Poland, this autumn.
Local Round Organizers
The Artificial Intelligence and Learning Systems Laboratory (AILS Lab) of the ECE-NTUA is proud to organize an EESTech Local Challenge round for the second time in a few years. Dr. Georgios Siolas was responsible for creating the local round Challenge and for providing academic supervision and technical assistance to the participants during the competition.
The organizers congratulate all participating teams.