ICC student symposium 2013 at the Amsterdam University Colleg...
Using labelled data in classifying film excerpts into genres
What makes a movie belong to a certain genre? How do we categorise movies in genres? What brain areas are activated when watching an action sequence? What does the plateau of about 65% mean in MVPA analysis? What is the significance of MVPA and what is its future? These are just a few of the many questions raised during the Brain Lab 2012 experience.
During the Brain Lab course we tried to look for brain areas that categorised film excerpts well into their corresponding genres. We hypothesised that film elicits emotions, emotional processing, human movement processing, face perception, and theory of mind, as well as simple visual processing. We attributed different effects of film to genres and let the participants also categorise the genres themselves. However, when you think of it, what makes one movie belong to a certain genre? The four basic film genres are action, comedy, drama, and nonfiction (Visch 2007). We would naturally expect a crossover between genres. For example, The Hunger Games would be categorised in the genres action and drama, because of its romantic component. And, of course, one joke does not make a movie a comedy, but what should be done with funny moments in action films, such as in The Avengers? Therefore, I suggest focusing more on categorising events in the movie and basing the genre categorisation upon this. Rather than expecting to find certain brain activity correlating with one genre, a list of events could be compiled of things typically happening in action (explosions, chasing, gun fights), drama (divorce, death, unrequited love), comedy (people tripping, jokes, humiliation), and nonfiction (nature or city scenes, newsman talking, actions such as baking cakes or constructing cars).
A lot is possible with such an approach. First of all, we can write a protocol to categorise. For instance, if 60% of the time, action-typical events are recognised, we call the movie an action movie. With 3 chases, 5 gun fights, 2 swearing people, 2 jokes, and 2 deaths, we would state, this movie is an action movie with a certain probability. We write such a protocol for all the genres.
This was just the preparation. Imagine that you show specifically cut excerpts from movies of across different genres to the group of participants. Let them watch a gun fight, a dying character, joking characters, a weatherman talking about the weather, while you know which movie sequences are shown. This is then the labelled data. During the acquisition of the fMRI data, certain brain areas or networks will be active for excerpts, and hopefully, different for different excerpts. Now, you have clear set of brain areas for events typical of each genre. In a second experiment, you would show the participants longer movie excerpts. With the patterns of brain (de)activations acquired in the first experiment, it should be feasible to predict what events the participants have watched during the second experiment, perform some data mining and classification on the unlabelled data, and give a well-educated guess on the genre the participants were watching. This should pose give a different method of predicting movie genres as used in the Brain Lab 2012 course, and hopefully give a higher probability score.
In the case that this works out, we would have done a bit of mind reading! What is essentially different from this approach to the approach used in class is that we view genre not as clear-cut, distinct categories, but rather view films as composed of element-wise genre specific events. These elements are events and film viewing focuses more on genre specific events than overall genre perception.
What are your thoughts on this?
Visch V. 2007. Looking for genres: the effect of film figure movement on genre