PhD work

My PhD focus on studying a problem called styles or personas, using deep neural networks. You can find the whole manuscript here. The papers published during this period can be found here.

But what is a style anyway?

Long story short, whenever you do something, there are two components: what you want to do (task), and how you do it (style). 

An example when you say the word "seriously?". Depending on the manner you say it, it will carry different meaning (sarcastic or surprise for example). One task, two different styles, two different results in this case.

But in the previous examples, we know the task. We can describe it clearly. But is this always the case? In simple cases, probably yeah, but if you have a complex sequence of behaviour (lets say two people interacting with each other), the set of tasks are diluted and unclear.

In such realistic scenario, there has to be a heuristic first to divide the sequence into a set of tasks, and then explore the styles of one of those tasks (or at least some of them). We can consider those as local styles.

There could be also some global styles (high tone of voice for example, defensive or offensive attitude,....etc).

We will not handle these cases during the PhD though. We will always work within a well-defined task.

What am  I doing exactly in the PhD?

My PhD focus on studying styles, given that the task is well-known, using deep learning approach. In particular, we were concerned with the following questions:

1. Propose a framework to study styles

2. What kind of benchmarks are suitable for comparison?

3. How to extract styles?

4. If we learn some styles, can we leverage them to accelerate the learning of new styles? (i.e., transfer learning of styles?)

After many over-complicated or over-simplified approaches, we settled on the use of Deep Autoencoders as the basis to study styles

Propose a framework to study styles

If you know the styles in advance, and you have labelled data for these styles (i.e., you have explicit knowledge of the styles), congrats, you are with the classification framework! You can have explicit evaluation of the styles as well (via metrics like accuracy and F1-score for example).

• • I assumed styles to be categorical here (ex: say clockwise vs counter clockwise in drawing letter X), but they could be continuous as well.

But what if you don't know (as it is usually the case), then how do you study styles? Explicit study and evaluation are out of the question in this case.

The assumptions in our work

• A notion of a task is known beforehand.

A word about tasks and style (perspective)

Identifying the set of relevant tasks and styles is contextual; it depends on the context you are looking at the problem.

The way I think about it is that: the problem is fixed, but can't be observed. Depending on the context, we can see an aspect of the problem (styles + task).

One can imagine this as having a light bulb directed to a some geometrical body. We can only observe the shadow. The shadow tells us a bit about the body, but not the complete fact. Depending on the direction of the light bulb, the shadow will change, showing us a different aspect for the problem. Stretch your imagination a bit and imagine the shadow has colours as well, depending on the direction of the light bulb, its disthance, and some aspects of the problem. On can imagine in this case that a task in this case will be the geometrical shape of the shadow (a container), and the style is the colouring of the shadow.

Lets take handwriting as an example: the problem is handwriting. Letter "E" is an example of handwriting (a shadow resulting from the light bulb being at some angle). The shadow colours that f it in the resulting geometric shape are the styles of letter E (also, mention that the bulb not only determine the letter, but also the context -- are we looking at a picture or the movement sequence of the pen).

Food for thought

• Instead of trying to model the problem itself (which we don't have access it), can we model the light bulb possible movement instead? Then, can we compile all the resulting shadows from lots of light movement, in order to get to see the actual problem?

• Problems can have problems inside them, depending on the level of granularity (another context) that you look at the problem: consider drawing as a problem. Handwriting can be an aspect of drawing, and maybe we can study handwriting as a whole. We can also look at handwriting as the problem, and the individual letters as aspects of handwriting. We can look at letter A as a problem, the sequence of micro-movement that construct letter A as the aspects, and so on.

  • Matter can be divide into pieces, and the pieces into smaller pieces, ...etc, tell we reach the atoms, then we can divide the atoms into internal components, and so on

• The genius really is to define the the limits of what you are looking for, and develop solutions that fit that limit.