Do-it-yourself is the best learning method. To handle, to work, to treat, to try on something, finding new approaches: just by doing we gain deeper insights, develop new ideas and desire to realize them. That is part of – human – intelligence.
Homework calls us to take action. The AI example, presented by Jacob Beautemps and Dr. Philip Häusser on YouTube and implemented as a Google Colab project, shows a simple way to involve practically into AI. By extending the training data (lengthen the number series x and y), or by modifying the training repetition rate (change the variable “epochs”), we’ll get a feeling for AI: how does it work, what can we expect from it and what maybe not.
We learn that AI
- finds structures in data sets and establishes connections,
- is able to draw conclusions and make predictions,
- for doing this takes random weights for “neurons”,
- and changes them in the course of learning, namely in a way,
- that optimizes a simple “transfer function”, using a measure of statistical error,
- but finally does not provide an exact result (7 times 2 is 13.something). Because AI thinks statistically, not analytically: ... a neural network has no maps, nor compass, it doesn't know what it is doing ... as Jacob and Philip put it into words.
Now, it is not surprising anymore that results vary a little from run to run (e.g., with 1000 epochs each). We can reduce this statistical fluctuation by increasing the amount of training data (lengthening the number series x and y) or by deepening the learning process using the variable epochs.
Larger amounts of data and deeper learning, however, cost computing time: we are back to the enormous computing power required for real AI applications. For mobile AI applications (e.g., Autonomous driving), GPU capable Industrial PCs are a preferential solution, since they offer accelerated computing on graphics cards. For this, Omtec recommends Neousys GPU platforms.