Unimelb 墨尔本大学
NEUR30006 Real and Artificial Neural Networks

Project

A project is an explora0on of an issue or concept or phenomenon we have encountered in our study of ANNs. Crucially, the project is constrained by conver0ng or interpre0ng the topic into a specific, testable proposal (o@en expressed as a hypothesis). O@en the idea is tested in a very simplified or essen0al form. We are not trying to make func0onal ML applica0ons for real world problems – projects might explore the mechanism, limita0ons, possibili0es, and exploitable features of ANNs. Alterna0vely, ANNs might be applied to different problems (i.e. different data sets) to gain some insight into the nature of the data. Projects don't have to solve problems (or even work as intended) but should explore issues and allow some conclusions about them. This is why construc0ng an interes0ng, valid, and testable hypothesis is a good assurance that you have a good project.

The other thing you will need is a suitable data set. If you’re not using MNIST data, there are many other datasets available (see Kaggle site for example), - in some cases it’s possible to make your own data (with your python skills!).

Most student projects fall into being either about the way the ANNs works, or an inves0ga0on into what is present in the data (what features, or proper0es of the data set make it learnable etc.). Some projects are looking for analogies with biological systems, or how ANNs can be used with different classifica0on problems.

The project must be based on the standard ANN model that we have been studying. You can compare with other models if you know about them, but you can't do a project based within a different programming environment (such as TensorFlow).

Unless you have an excep0on approved by your tutor, your project must:

• Use the mnist dataset (or some variant of it you implement) OR

• use the standard ANN (i.e. Tariq’s - or some variant of it that you implement) OR

• both of the above.

You can use advanced python libraries as appropriate, but, fundamentally, you need to understand what the parts of your program are doing (and how): I want to assess (among other things) how well you grasp how an ANN implements learning and how this is achieved in Python code.

Examples of the 0tles of projects that students have done in the past:

• Can an ANN learn to count?

• What are hidden nodes looking for?

• What can’t the ANN learn and why?

• How good is a simple ANN at learning textures?

• Do ANN and t-SNE make the same errors with MNSIT classificaFon?

• Is learning more sensiFve to degraded images or degraded networks?

• What, if any, informaFon in the training data is irrelevant for learning?

• Can neural networks accurately idenFfy the number of dots in an image?

• Can the weight matrices for good and poor learners be disFnguished visually?

• How adaptable is an ANN in response to a foreign input in the MNIST dataset?

• How would learning MNIST test images be improved by expanding the training data?

• Are criFcal periods in learning inherent to, or an extension to, the neural network model?

• How important is the degree of randomisaFon in the training data for learning performance?

• Can simple changes to MNIST training and test data, such as luminance and contrast, affect learning?

• ArFficial Neural Networks That Learn Curve PaMerns

• The Effect of Training Sequence on Network Performance

• The Inner DepicFons of Numbers in an ArFficial Neural Net (ANN)

• Compression of MNIST datasets using autoencoder pre- processing

• Image DistorFon and its Effect on the Accuracy of Neural Networks

• A novel data pooling method to increase learning speed in the MNIST ANN

• An exploraFon of regularisaFon methods with the aim of modelling neurological deficits

• InvesFgaFon of the recogniFon of four geometric shapes via a mulFlayer neural network

• CorrupFng the test: how arFficial neural networks respond to increasingly distorted data

• The interacFons between model depth (number of hidden layers) and acFvaFon funcFon choice

• Capability of ArFficial Neural Networks in DisFnguishing Between LeS and Right HandwriMen Digits

The report size, of 3,000 words, is a guide to the scope of the report. It may be that your report can be well-described in less than 3000 words, or it is possible that 3000 words isn’t enough to provided all of the informa0on that needs to be conveyed. In such cases there is no need to be concerned about exceeding or not mee0ng the word limit. However, reports that are overly wordy, discursive, repe00ve, or contain extraneous informa0on, could a^ract mark penal0es – regardless of the number word used. Similarly, reports that are below the recommended limit and seem to omit informa0on that would be germane to explaining aspects of the study would be penalised.