Connectionist Modelling Essay Assignment
You are asked to write an original essay on one of the following topics related to the course. Essays should not exceed 3000 words. You should read several sources beyond the course materials in your preparation. Try to limit yourself to between 3 and 5 articles/chapters. Identifying those sources in this Age of Google should not present a significant challenge. Due date for these essays is Friday, May 11th, 2018.
You may also suggest an alternative essay topic if you like, but please clear it with me before you invest any time and effort in it.
| Connectionism and Music | What efforts have been made to get networks to compose, perform, or interpret music (your choice)? What has been achieved, and where do the main challenges in the future lie? |
|---|---|
| History of Connectionism | To what extent does connectionism represent a continuation of the empirical tradition in psychology? What debts are owed to thinkers pre-1970, and what is genuinely new? |
| Lesioning networks | It is possible to model the effect of brain damage by training large networks and lesioning them (removing units/weights). Identify one or two such attempts which make overt connection between network lesioning and brain damage. (this is the vegetarian option). |
| Kohonen networks | We only covered a few basic network topologies in the course. One very influential model we skipped is Kohonen’s Self-Organizing Map (SOM). You will find tons of literature and some dinky Java applets on the topic. Learn about the basic SOM algorithm and discuss its possible relevance to cognitive modelling. Suggest where SOMs have been most fruitfully applied. You will need to restrict yourself to the basic SOM model, or risk drowning in recent variants! |
| Reinforcement learning | Reinforcement learning is a form of training which attempts to eschew the provision of an exact teacher, but which instead provides occasional feedback in the simple form of "that was good" or "that was bad". Describe the principles of reinforcement learning, in particular you should describe the algorithm known as Q-learning, and describe some areas in which reinforcement learning has proven to be a useful approach. What kind of tasks would reinforcement learning be a better choice for than supervised backpropagation? |
| Deep Learning and Neural Representation | Deep learning is in vogue. Many of the achievements of deep learning are inspired by the constructive processes Hubel and Wiesel first conjectured to underlie vision. Explore this connection, with one eye on history (do deep learning researchers make this explicit?) and the other on content (what do these algorithms do? What varieties of Deep Learning exist?) |