What I learned about Machine Learning?

The word Machine Learning, Artificial Intelligence added to a project always captures everyone's attention. Abbreviations like ML, DL, AI seem to captivate people into believing that most of human's workload can be automated and that the computer is special or smarter into thinking than us humans. Yes, it may be true in some aspects but most aspects it is not. Here is why. We first started the project believing that we just have to pass the data into a Machine Learning model and it will do the rest. For our case, was to detect anomalies in the data. We thought it would be really easy to just get some code from Github and just feed the data. Oh how were we wrong. From our first presentation into the final semester we just believed a Machine Learning algorithm called Recurrent Neural Networks (RNN) would solve our issue. It has been used extensively for Natural Language Processing (NLP) that is useful for google translate or providing automated or generated subtitles for videos such as in YouTube. RNNs are also used for predicting stock prices and since it can predict stock prices it can also predict how our BACnet Data would look like in the next 3 to 4 hours. However, we later learned that Deep Learning algorithms or most Neural Networks need to be fed a lot of data and with the Covid-19 problems occurring we had only obtained 2 hours worth of pure BACnet Data and so we really had to back to the drawing board. After our first presentation, it was our lowest point when our professor mentioned in class that ML is not just a fancy term and that there are so many algorithms to choose from. He was definitely right.

We moved from RNN to Random Forests all the way knowing that we do not have anomalous data and therefore we do not have labelled data. This means that we have to opt for unsupervised learning algorithms. It just never clicked until we found cluster analysis such as K-means and K-modes to be very useful. We learned these basic algorithms in our Machine Learning class ECEN 489 but we thought it would be too simple and easy to do. Eventually, we just tried it and the results we got seemed somewhat good in terms of what we were looking for - to predict anomalies. During this time we had been sending and receiving email with a Machine Learning expert Dr. Wahid Ferdous and when we stumbled amongst K-means and K-modes we felt that we felt that we could see the gold at the end of the rainbow. However, we knew that we were still at the start of the rainbow. The journey of sending and receiving emails with the Machine Learning expert and throught tons of medium articles, github code and reading we finally got there. We learned that the machine is not as smart as we thought. It is what We inform the machine to do is what is 'special' and considered AI.

Additionally, you cannot just feed collected data into a Machine Learning model. We had to preprocess the data so much in order to feed into the model. Preprocessing is one of the most vital tasks we have learned to overcome and do in terms of Machine Learning. I believe it is mostly about preprocessing. All the years of studying with math, probability, counting starts from preprocessing. From here you have to predict what kind of data you want as an outcome, what should the machine learning model generate and do. These are all the things that should be considered. We never knew that K-modes and categorical data existed until our Machine Learning expert mentioned it to us. All of this resulted from nothing until we researched and learned so much. The machine learning model is nothing special it produced another csv file in terms of the centroids. This is all it was used for. Just too simple. But simple turned out to be one of the best ways to detect anomalies in the data. Once the K-modes and K-means were done everything was set up and we never knew that Machine Learning does not have to be complex or crazy just to sound cool but it has to make a difference, it has to contribute to something. For us, it did but maybe for you it may not.

Thank you to all my teammates Sofian and Rahul for making this happen. 3 heads are always better than 1!