Initial thoughts

Hey @iteles @nelsonic

These are just some initial thoughts. Not sure where I should be putting this, but thought I’d stick it up for you to have a look at. Let me know if you want to have this conversation somewhere else.

#### Questions:
* What is dwyl's strategic vision for its ML/DS capacity?
* What need has been expressed by clients?
* Alternatively, what potential applications have you seen in clients fields that they are not aware of?
* What is the purpose of the three different repos (`learn-data-science`, `learn-machine-learning`, `learn-artificial-intelligence`)? From my perspective, they are Russian dolls (i.e. data-science < machine-learning < artificial intelligence). It would be useful to have one place to organise higher level strategic thoughts; it’s not clear to me where that is at the moment. I guess I would default to using the `learn-data-science` repo for that, but open to suggestions.

#### Comments:
* Given there are already a wealth of technical ML/DS resources, I’d say it makes more sense for whatever resources dwyl creates to be an attempt to “fill in the gaps”, provide context for existing resources and also provide higher level conceptual overviews for non-technical (or non-mathematical) readers.
* One of the more important reasons for collating, curating and augmenting existing learning resources in the first point would be to arrange them into a learning program that could be successfully followed by existing and future dwylers.
* I would recommend the creation of a data ethics policy that spells out how we do and don’t use public and client data. Ideally the spirit of that policy would be woven throughout any learning resources we create. I think this is separate from any data security policies dwyl may have.
* Crudely speaking, the more raw data you have access to, the more you will be able to infer from it. However, if the aim is to help smaller organisations that don’t already have access to ML resources, it’s likely they will not be generating large amounts of data already. This means that hyped techniques like deep learning may not be appropriate (since they tend to require large amounts of training data). Instead more traditional statistical inference techniques are likely to be more useful.

#### First Steps:
* Develop overview of state of current fields, their uses and applications. This would be documented across all 3 current repos.
* Build up list of external learning resources in these fields
* Prioritise fields based on relevance and interest, then using learning resources:
* assess resources by consuming them with the intention of arranging them into a learning program for others.
* start developing technical demonstrations with public data.
* Will also probably want at least one resource on assessing the various languages, libraries, frameworks and tools that are available (e.g. python, Julia, Lisp, R, MATLAB; scipy, pandas, theano, tensorflow, etc).

The technical demonstrations step could go on indefinitely; that’s the point at which understanding the strategic vision will help me/us.

#### My Objectives & Interests:
For the record, I think at the moment my objective can be loosely encompassed by “to help bring machine learning and data science into the commons”. Won’t elaborate further here.

In terms of areas of interest, here’s a short list, in no particular order:
* Bayesian inference
* Estimation (MLE, Kalmann filters, etc.)
* Bayesian belief networks
* Analog & digital signal processing
* Discrete and continuous optimisation
* Genetic programming
* Neural nets & deep learning
* Factor analysis
* Complex systems, emergence & non-linear dynamics
* Pitfalls and traps in ML (e.g. overtraining, biases in training data, interpretations of results, ROC curve etc)