Thursday, August 5, 2021

Taking a journey to expand my Artificial (AI) Knowledge - My takeaways from AI For everyone

If you are in security, or even from a general perspective Artificial Intelligence (AI) is something you have sure been hearing a lot more about. AI is the future matters or as the saying goes, AI is the new electricity. Therefore, it is important we all start getting our head around this as much as possible I've already dedicated time in the past to getting a better understanding about Machine Learning (ML) and thus decided I should dedicate more time to this area of study. To help me start this new journey, I've decided to start with AI for Everyone by Andrew Ng and plan to follow a series of training over the next few months to expand my knowledge on AI. My blog here is being used as my "notepad" where I jot done things I find interesting as I take the various trainings.

This series consists of a number of videos and to ensure I can reference my materials in the future, below is just my takeaways from the videos.

Intro:
First takeaway, is I like the fact that it starts off from a purely non-technical perspective. While my fascination with AI is mostly from the technical perspective, I like how Mr. Ng shows that within the top 11 industries, retail, travel, transportation, etc., will see greater growth in AI than high tech. Actually high tech came in at number 8. I find that very interesting.

I definitely like Mr. Ng points about the necessary and unnecessary hype surrounding AI. First major takeaway for me. AI really consists of two parts:

    1.    Artificial Narrow Intelligence (ANI)
                a. Focuses on a particular tasks.
                b. One trick ponies which can be very good or very bad at the tasks.
    2.    Artificial General Intelligence (AGI)
                a. Aim is to do anything a human can do
                b. Even more things than humans can do
                c.  May require years, decades or hundred of years to make real progress.

Ensure that your AI projects are technically feasible. Also, ML is what is driving the rise of AI today.

Very funny also that Mr. Ng says by the end of this program, I will have better knowledge and better qualified than most CEOs of large companies. ;-) I hope that is true by the end. ;-)

Understanding Terminology:
    AI: Everything below is really a subset of AI. 
    ML: The ability to learn without being explicitly programmed. 
            This is a subset of the various tools which use AI, ML, Deep Learning, etc.
    Data Science: Extracting knowledge and insights from data.  
        Can be a subset cross cutting all the various AI and ML tools
    Deep Learning: Leverages a Artificial Neuron Network (ANN).
        Deep Learning and Neural Networks are used interchangeably.  
        This is a subset of ML.

On Machine Learning:
The rise of AI has largely been driven mostly by ML. The most common type of ML is supervised learning which deals with an input to output mapping. Some common ML applications are: Spam filtering, online advertising, machine translation, self driving car, visual inspection. Big driver for the growth in supervised learning can be seen from the perspective of increase in available data and computing power.

Data in ML is called a dataset. While data is important, it can also be misused while at times overhyped. Do note that is that more data is typically better but generally you still have to consider garbage in will equal to garbage out. Images, text, audio, etc., are call unstructured data. Structured data on the other hand can be considered as data which might be in a spreadsheet.

While most of the time you may hear folks talking about supervised learning, there are other ML techniques such as unsupervised learning of which clustering is the most popular mechanism. In unsupervised learning no labeling is done and the algorithm is expected to identify things of interest. 


On being great at AI:
To be a great AI company, you have to be able to do the things that AI lets you do very well. Additionally, AI companies can easily spot opportunities to leverage automation. The following is what Mr. Ng recommend for his 5-step playbook to successful AI implementation. 

    1. Gain momentum via pilot projects to learn whether AI is a good fit for your project.
    2. Ultimately, you will need to build an inhouse AI Team.
    3. Ensure the AI team as well as various business leaders get the necessary training.
    4. Develop an AI strategy
    5. Develop communication for both internal and external ensure all stakeholders are updated accordingly.

Before embarking on an AI project, ensure technical diligence is done to assess the practicality and feasibility of the project. While we might hear about only the good news about AI, there are also lots of failures. Machine Learning tends to work well with simple concepts with large data. On the other hand, ML works poorly with complex concepts from small amounts of data and on previously unseen data which is not part of the dataset.

On Deep Learning:
The simplest possible ANN has one neuron, taking a single input and producing a single output. A more complex ANN may take multiple inputs and have multiple (thousands or even tens of thousands of) neurons and layers while producing an output. 

Deep learning models make use of numerical data. 

On Building AI Projects:
First build an AI workflow, then select an AI project, then organize the data and team to execute the AI project. 

Key steps within the project is to collect the data, train the model then ultimately deploy the model. Do note that as you go through the various steps, you may more than likely optimize/tune as you go along. 

Data science projects have different workflows from ML projects. In data science projects, similar to ML projects, you first collect the data. However, in data science, you then analyze the data then suggests hypothesis/actions. 

To make the most out of an AI project, you need AI experts as well as domain experts. Domain experts being the individuals in your business who possess the knowledge about your business. These are also called cross functional teams.. When looking at ML, consider automating a tasks before automating a job. Other questions to ask are what are the main drivers for automation and what are the main pain parts of the business. Note, you can make progress without big data meaning a small dataset. Also don't build some thing for which there is also a good solution in the market.

When working with the AI team, ensure you clearly define your acceptance criteria. This is typically done in a statistical way. The AI expert should also be able to tell how much test data you want. While it is typical to have your data split between 1 training set and 1 test set, it is also possible to have 2 test sets. 

You should never expect your models to be 100% accurate. This can be for a number of reason such as not enough training data or just natural limitations within ML. Maybe mislabeled or ambiguous data. Considering the above, ensure discussions are had with the AI expert to understand what level of accuracy might be acceptable.

While some roles are defined, there are some roles which have not been clearly defined. However, some of the roles to be of are software engineer, ML engineer, data scientist, data engineer. AI Product Manager, etc.

AI Transformation:
Most important thing is to be successful rather than being valuable when selecting that initial project. Try to show returns within 6-12 months and this project could be in-house or outsource.

When building an in-house AI team, have a dedicated that that is available to support the various business units. Focus on building company wide platforms where possible as this can help multiple business units at the same time. Remember to provide broad AI training at various level of the company. While hiring AI engineers from outside, it is a better thing to build that skillset in-house.

AI impact on society

AI and Ethics is something society also have to pay close attention to.

AI can be biased or discriminate on minorities while also being vulnerable to adversarial attacks. Actually learning about some of the bias exhibited by AI was very shocking to me. I always knew of the bias especially around face recognition but some of the others caught me by surprise. It was also nice to see that the AI community is working diligently to address this bias.

Adversarial attacks is where attackers are trying to fool the AI or basically attacking the AI system. Think someone trying to evade a SPAM filter. This is more than likely going to be an arm's race between attackers and defenders.

While AI might be used by adversaries for bad, it may also be used in other adverse ways such as making of deep fakes, oppressive surveillance by regimes, generation of fake comments, etc.

When thinking about AI have a realistic view. Don't be too optimistic or too pessimistic. Consider it the Goldilock's rule. The poridge must be just right. Not too hot not too cold. The lack of AI to properly explain itself maybe a barrier to some implementations.  You may have to instead rely on the AI team for this guidance. 

That's it for this my first set of notes. Join me as I continue this journey as I now pursue the Deep Learning Specialization.


References:
AI For Everyone
AI Transformation Playbook


No comments:

Post a Comment