An enginering and science blog for Excel automation
I don’t know where this post will lead to (started the journey 01/21/19). The commitment is to get a Neuronal network add-in for Excel… limit time is end of 2019 (nothing bad would happen if goals could be achieved earlier).
Absolutely, Excel is not the right tool to carry serious ANN -in terms of perfomance-, but, considering that all the computations can be shown in real time, and that you can mount a RAD model of every architecture, it should be considered among the better options to learn the subject from the basics. And the two final reasons where I finally arrive at: because I can do it, and because is so ubiquous that it does not limit to matematicians and IT people to play this things. If you want to make big stuff, you should consider other platforms and even other hardware… but that’s not my target and this is not your site if you have arrived here looking for other than Excel.
Going to the business, from all the architectures that can be performed under the label of ANN, I’ll try to first pursue the ConvNet as they expose most of the algebra needed to perform the whole thing, so I hope it will be easier for me to adapt to other type of architectures. The final goal is to get a tool that can be good for any ANN model (considering the size limitations).
This post should end fused with my previous post on Neuronal Network on Excel (task pending…).
Expecifically I’m looking after the following goals:
NN matrix algebra (multiplication, add/substraction, element-wise operations… with matrices/vectors). Half achieved, refactoring the MatLab translator module. Others should came from the JS translator.
NN neuron manipulation (add/delete/copy/move).
NN layer manipulation (add/delete/copy/move).
NN specific functions. Filter, Pool, stride, convolute, backpropagate,…
BMP/JPG import/export… someway achieved refactoring the ASCII Excel module.
As soon as each goal be achieved, then it will be converted to milestone.
There’s one specific problem if this is to be done under Excel Environment… limits. For example, the >2K3 column limit of 16384 can hardly allow to show/represent more than a set of (72×72, RGB) bits in vectorized form in one worksheet. I still don’t have managed a way to overcome this limit, as I consider necessary to show the data on the cells, but surely will come to one (split neurons on different worksheets, aggregate inputs in 2D shape,…).
Other’s implementations that I can base the thing on…
Not an exhaustive list, and maybe not any usefull, but here is a recopilation of other software I’ve seen to perform NN tasks:
SNNS (Stuttgart Neural Network Simulator)
IMPLEMENTATION OF AN MS EXCEL TOOL FOR BACKPROPAGATION
NEURAL NETWORK ALGORITHM IN ENVIRONMENTAL ENGINEERING
EDUCATION (Selami DEMİR & others)
Andrew Ng specialization course -videos- Deeplearning.ai, on Coursera platform. Very instructive, tips are explained in a very clear form, so anyone, even with not a high level of maths, can follow. You need to know and understand algebra. Only one handicap, IMO, is that it’s focused on TensorFlow (Google) platform.
Jeremy Howard’s courses on fast.ai platform, online courses -also on Youtube– (and here a little trick if you’re to follow this course). To use on PyTorch, that seems to have its momentum by the end of 2018 and as a true OpenSource alternative can stay for long. Differences between them spotted here.
On Medium there are a lot of interesting articles as well (I must confess that I personally get surprised of how people put a lot of creativity to achieve a goal that stands above most “expert” no-sayers):