Project Ideas

Detecting Fake News

There are different categories of fake news. Sometimes they are spread to serve political reasons. Others serve just as clickbait. And then there are "honest" fake news where the people spreading them honestly believe in them. Drinking bleach to cure Covid, and other tales about the risks of vaccinations - for example making vaccinated women infertile. The goal of this project is to develop a system that detects one particular category of fake news automatically through machine learning.

For example NyQuil chicken (cooking chicken in cough medicine) started as a warning by the FDA not to try it, only to be picked up by many social media users, who took the recipe seriously.

Detecting lies

Humans are masters of deception. Be it either to “bullshit” to impress others, or to manipulate them for one’s own advantage, we lie but we also want to believe these lies. Using deep learning and AI we will be using different communication channels to detect lying

1. video analysis to detect lies from face and body postures

2. NLP analysis to detect lies in online social media and emails. This is based on earlier work with the tribefinder (politicans/journalists)

3. The happimeter to detect lies from body language

For each of the three subprojects, we will need to identfy datasets for training the ML system, which will have "ground truth" of known lying. For instance, recording poker players with Webcams and the happimeter while they play will identify body language of lying (bluffing).

Collaboration Networks of AI researchers

It has long been claimed that scientists build on previous knowledge. As Isaac Newton said "If i see farther, this is only because I am standing on the shoulders of giants". Science has defined its own language for each subfield, a psychologist talking to a biologists is like a Chinese talking to a Finn, thus increasing communication and transfer of knowledge between science of the same field, while raising barriers among different disciplines. The recent progress in AI, in particular machine learning and deep learning, takes collective intelligence one step farther, and the language spoken between AI research is not English anymore, but python and tensorflow.

The goal of this project is to create two correlated networks from scientific citations in AI, a co-author and a citation network (from e.g. Scopus, Google Scholar), by also collecting the abstracts of the papers.

Towards "True" Emotions

While machine learning has made huge progress to recognize human emotions from facial expressions and voice, it is far from clear what these emotions. There are dozens of different emotion frameworks. Even more, latest psychological research rejects the notion of "universial emotions", claiming instead that emotions might differ from one individual to another. The goal of this project is to investigate individual emotions. For instance, do all the face pictures labelled "disgusted" used to train a FER system truly express "disgust". Using a dataset created in an earlier COIN course from TV soap operas with the text of the soap as ground truth, identify the segments where face emotions have been mis-labelled, and rerun emotion recognition there, to drill down on misinterpretation. (paper 1, paper 2)

Measuring human-plant interaction

Build a software tool to more easily measure EEGs of plants (based on cheap EEGs for humans)

Currently we use a derivative of an EEG for humans (the BYB spikerbox).

This is based on earlier COINcourse projects:

https://plantsasbiosensors.vercel.app

https://kups.ub.uni-koeln.de/53756/

https://www.dropbox.com/s/u5bxqmv975xsz2l/GreenBox_COINs_FinalPaper_Group7.pdf

A similar project is already used with tomato farming: https://www.srf.ch/news/schweiz/sprache-der-pflanzen-wenn-tomaten-mit-dem-gemuesebauern-sprechen?

We currently have a large dataset of humans interacting with garden plants (plant spiker data and video recordings) through Eurythmy. The goal is to build a software tool for easier analysis and data collection, and identify correlation and causation of human-plant interaction.

Talking to Animals with AI

Build a software platform for emotion recognition of animals based on pilots done with horses, dogs, cats, and cows. The frontend is a smartphone app capturing a video stream, the video stream is sent to an AWS server for recognition. The system consists of four components:

1. Image database with labeled pictures of horses, cats, dogs, cows

2. Model recognizing emotions of these animals, based on face emotion recognition, posture recognition, and voice emotion recognition (barking, meows, mooing, whinny)

3. Server in the cloud running the model

4. Smartphone app taking the video and sending it to server

Earlier projects:

Horses https://www.mdpi.com/1999-5903/13/10/250

Dogs https://www.mdpi.com/1999-5903/14/4/97

Cats https://www.dropbox.com/s/mlgjb8090siajnh/COINs_Submission_Team01.pdf

Cows (ongoing Master’s thesis at HSLU)

Further ideas

1. findmysoulmate as a love match maker from WhatsApp data (using models that find tribes and personal values from language)

2. Measuring/curing depression with happimeter (this requires collaboration with a clinical psychologist)