Machine Learning for Managers
Now we have come to the point where discussion of Artificial Intelligence and Machine Learning (now onward we will call AI/ML) is no longer confined between experts. No matter who you are you need to know what this new technology is and how it is affecting different aspects of your life, from entertainment, shopping choices to learning and personal relationship. Since there is so much hype about AI/ML so it is very easy to get lost.
This article is targeted to managers — professional who are responsible to building and managing teams and rolling out business solutions based on AI/ML. Since in industry AI/ML is being adopted within the broad framework of software engineering so there is a lot of scope for confusion. In fact many of the software engineering practices have been forced on AI/ML projects and those are bound to create tensions and frustrations. For example DevOPS has been quite successful in software engineering but its AI/ML flavour, known as MLOPS, may not be that successful. Again the way data is considered and used in AI/ML is very different from how it is used in other data centric problems. Here are 10 important points which managers need to know when managing AI/ML projects.
- AI/ML is not a software !
For a AI/ML projects you need all the four components:
- Data — Without data no AI/ML is possible
- Computing — Most AI/ML projects, particularly when they involve deep learning, need huge computing in terms of GPU/CPU/TPU hours and memory.
- Technical Expertise — To build a real world AI/ML solution a lot of new and original work may be needed. Data may not exist in a tabular form and the approach may not be obvious. This means that a lot of experimentation may be needed.
- Business use cases — Not all the business use cases may need AI/ML so it is important to identify use cases where it can add value.
2. Build vs Buy ?
Many big tech companies such as Microsoft, Amazon, Google etc., already have pre-trained AI/ML models and plug and play solutions which you need to pay for per API call. There are pros and cons of buying/renting AI/ML solutions.
- You may not need data, computing or technical expertise when you buy. Anyone can easily integrate AI/ML solutions with products such as image recognition, Optical character recognition, speech and text solutions etc.
- Renting an AI/ML solution can be quite expansive particularly if you are integrating it in your products. So keep this factor also in mind.
- Vendor Lock in : If your AI/ML capabilities depend on a third party tool then there is a chance of vendor lock in — the tool may not be up when you need it.
- Lack of transparency — Plug and play tool work as black box and there is no transparency what goes inside the box.
- Lack of flexibility — In most cases these plug and play tools are built for general problems and customisation may be expansive.
Here is a short comparison of Buy vs Build
Buy Vs Build Comparison
3. Investment & Return on Investment (ROI)
AI/ML solutions are expansive since they need data, computing and expertise. In fact one can write the following formula for the ROI :
ROI = B * C * D * E
B = Business, C = Computing, D = Data, E = Expertise
Business : AI/ML technologies have potential to attract more business if properly projected therefore it is very important to launch different business activities (such as advertisements, publishing white papers, organising events etc.,) around AI/ML.
If you spend all the money on expertise on none on data or computing things will not work. So the best bet is to spend 25 % on each.
4. Training & Inference
In must be kept in mind there are two parts of any AI/ML exercise, named training and inference and both have completely different requirements.
Training — This data and compute intensive exercise needs very different computing environment than for inference. Note that training is very expansive (Open AI GPT3 cost millions of dollars to train !) and it is not done frequently. Training may take many days & weeks even on powerful computers and may cost huge money if done on cloud ! Once training is done we may get rid of training data and keep just the trained model. Note that it is impossible to get the training data from the trained model by doing reverse engineering.
Inference — It is the part of AI/ML solutions that is exposed to the end user and it can be lightning fast. In fact any web server can work as an inference engine which can make predictions on the basis of the trained models.
It is advisable to develop the training code/pipeline and inference engine separately. Here is a comparison of Machine Learning Training & Inference.
5. AI/ML predictions are probabilistic in nature
Note that all the AI/ML predictions are probabilistic in nature. More general an AI/ML model is better it is considered. Performance of an AI/ML model should be measured in terms of performance metrics like precision, recall, accuracy, ROC, AUC etc. Just by looking one or two predictions to judge any AI/ML solution will not be fare. We should look for out of 100 cases how many times it is predicting sensible. In short, good or bad the AI/ML solution should not be judged on the basis of a single prediction.
6. No personal data is needed to train any AI/ML model !
Note that the main principle of AI/ML is generalisation — based on some property or feature that is common to large number of instances, for example age, gender, educational qualification, income and not something that is unique, like email or mobile number. This means that it will be very odd if a machine learning model uses any private data, since by definition, we cannot have more than one instance of any personal data. This means that it is best to mask all the private data columns (in case data is tabular) before giving it for a machine learning training. Just like Machine Learning models do not use personal data, they cannot make predictions on the basis of personal data. They can use only those properties which were used for featurization in machine learning during training, such as age & income.
7. Transfer of Learning
Transfer of Learning or Tol is one of the most important concepts of machine learning. In simple words this means we use a model trained on different data than the data we have. For example an image recognition model trained to classify fruit images also can be used to classify cats & dogs image, after some fine tuning. Managers need to be aware of the pre-trained model that can be used to build a new AI/ML solutions. Particularly, in the domain of machine learning on textual data (Natural Language Processing) there are a large number of pre-trained models available with Huggingface company which can be used. Recently OpenAI has made many of its pre-trained models like ChatGPT for public to use for free & with subscription.
8. Big vs Small Machine Learning
Machine Learning is a quite broad field and has many subdivision on the basis of the objectives, data size, complexity and use case. For example, if we have labelled data, we can carry supervised machine learning which is explained in the following way.
Machine learning is all about, at least supervised machine learning, finding a map/relationship between input and output. This relationship is discovered by considering a mathematical framework (model) having adjustable parameters, such as linear regression, decision trees or neural network.
In the training phase of machine learning the parameters are adjusted in such a way that there is minimum mismatch between the desired output and actual output. Once we are satisfied with the level of mismatch we consider the model trained and use that to find the output for any given input. In general, we may come across two type of situations. In the first the output is continuous number, such as price of a product and in the second output is a class or ‘label’ like name of the object identifies in the image classification exercise. The first exercise is called regression and the second is called classification.
There are two other type of machine learning also, called, unsupervised learning and reinforcement learning. In the first case we do not have labels and in the second case regular feedback is given to model.
In some of the machine learning models like linear regression, decision trees and support vector machines we may work with small data sets also so a different approach is needed as compared to the neural networks / deep learning where we may need to deal with huge volume of data to fit million or billions of parameters. So as a manager, we find out whether you are doing ‘small machine learning’ or ‘large machine learning’.
9. Tools and Technologies
There are a very large number of tools and technologies available which may come across in your AI/ML journey. Some of these tools are essential and some are optional. In most cases essential tools are Open source and optional are proprietary. If you work in any particular ecosystem such as Microsoft Azure, Amazon AWS or Google cloud, you may feel that it is not possible to do machine learning without the tools being suggested/recommended to you. In fact there is a single reason for going to cloud and that is for more computing ! In order to sell their products any cloud vendor will try to make your machine learning pipeline dependent on as may tools as the vendor have and not what is actually required. The best way forward is to develop most of machine learning tools locally and migrate to cloud only there is something, such as large computing, that is not available locally.
10. Research and Design (R & D)
Not every organisation can afford R & D which are expansive and returns are neither certain not come easily. However, we cannot ignore the fact that no new technology comes without R & D and AI/ML is also not an exception. At present all the big players, Google, Microsoft, Amazon etc., are involved in very extensive research activity and they best of the scientists from academia in their board ! It is very important for managers to keep R & D activities in focus and do not shy away from trying out new ideas which may not fit into their past experience & understanding of doing things.
In this short article I tried to cover some of the points which I consider important for managers. In case you find anything useful here please write, share & comment and this will encourage me to add more such posts.
