تحميل كتاب تطبيقات تعلم الآلة والكاء الاصطناعي باستخدام بايثون Machine Learning Application using python

The biggest risk in developing machine learning applications is that it is a trial-and-error- based process. Sometimes the algorithms work on certain types of datasets and sometimes they underperform on other types of datasets. There is no single answer in the best algorithm selection despite having a cheat sheet (which we discussed earlier).

In spite of this limitation in the business world, a machine learning consultant is under pressure to deliver the best solutions for the client. Even if the right algorithm is found, the hyper-parameter optimization takes a long time to determine the right parameters for the algorithms. Deep learning algorithms such as ANN and RNN require very high computational power and processing time in order to arrive at the optimum set of parameters.

There are other technical risks as well, such as requirement of high computational power through use of GPUs, and RAM for deep learning applications is something that, if not planned out properly for the production environment, can lead to derailment of the entire project. I have received many SOWs from my industrial clients who want industrial vision applications implemented on 4 GB RAM with 1,1 GHZ CPU with an ordinary GPU, and I have to inform them of the minimum computational requirements for neural network-based systems.

Another risk that I have found with most of my clients is that they have hired a team of machine learning engineers from the market; however none of them has the competence to produce a production-ready machine learning application, as they have never done this. Such technical resources are a liability, rather than an asset, for the organization. They have to then hire a consultant like me to guide them in building a machine learning algorithm end-to-end. Experienced machine learning engineers with knowledge of implementing machine learning applications are hard to find.

Data Analysis

During a particular iteration when the data was collected, Microsoft Excel software was used to record the expert’s response in a tabular format. For any given key area, a graph was made to check whether there was a consensus reached, and if the graph showed sufficiently well that there was consensus, then the iteration was stopped. So the data analysis was done manually with the help of computer software. The mapping of technology maturity and phases of technology adoption was undertaken using Excel software to create a technology map graph. This was also done with the help of Microsoft Excel.

It is possible that bias could have slipped into the study had we not made sure that the results and the responses of the experts were kept anonymous and did not affect the outcome of this study. So due care was taken to ensure that the experts were not known among themselves. As I have already mentioned, there are in the retail industry two groups of people: one whose members like technology and the other whose members do not like technology. We did not do an expert selection based on these specific criteria, so this study could very well be biased on such grounds and we have not tested for this.

Qualitative research has as its biggest limitation that of not being able to exactly quantify the outcome of the future, and this is very much applicable to our study as well. However, by using categorical variables in our questionnaires, we have tried to take the quantitative analysis of our outcome as well. Mapping of the technological adoption and understanding of the technological maturity is not something that a normal human being can do unless they have been associated with the industry; that is why we chose experts to carry out the study. However, it is possible that some of the experts may not have sufficient knowledge or exposure to the advances in artificial intelligence and machine learning. We acknowledge that this could be a limitation to the study.