AI Explained - Machine Learning
Written By: Tad Witkowicz firstname.lastname@example.org
In AI Tutorial Part 1 we introduced the model of AI and showed that Machine Learning is one branch of the AI tree. Practical business applications however are a mix of the various components and as a result Machine Learning and AI terms have become interchangeable. Essentially all AI is Machine Learning because the computers are processing data and get better at it (they learn) as the time goes on.
Machine learning became viable with the advent of very fast computers and certain hardware such as GPUs (Graphical Processing Unit) which enable fast processing of data in parallel. The vast amount of data (Big Data) that the Internet is delivering today makes autonomous learning (Machine Learning) possible.
How it works.
A good way to understand how a computer learns is to look at how Facebook learns what type of content such as news feeds and advertisements it should show you to make your visit to their site productive for you (you see what you like) and for Facebook to get paid for the ads you click on.
While Facebook keeps their exact algorithms secret several articles appeared on Internet that describe the basic principles of how the algorithm works and here are the basics.
The table below illustrates Facebooks algorithm model.
The first column list actions that you may perform while using Facebook. There may be more but this set is sufficient to illustrate the point we're making about Machine Learning. The second column represents the percentage of time you performed these actions when visiting Facebook. The frequency of these actions expressed in % of visits is what Facebook collects on you when you interact with each piece of content they show you.
The second column is simply a representation of the percentages as a fraction. The third column represents points that Facebook may assign to each activity. Generally they represent the value of your activity to Facebook. Therefore, a long video watched means you were interested in the video and that is very important for Facebook to know. For one, this means the contend must've been interesting to you and second, you watched most or all of it which represents an advertising opportunity for them. Similarly Facebook assigns points to other activities that represent more or less value to your activity. Facebook Share is awarded high points as this means you spread the word which again means you liked it and you did Facebook a big favor by spreading the word to your friends. Of note is the Hide activity which they assign a large negative number as this means you did not like the content at all and don't want to see it. Facebook also doesn't want to waste resources on content that doesn't make them money. By multiplying the Fraction and Points each of your activities is assigned a score. The total score represents the total value of the type of content Facebook presented to you. Since this can be done for every piece of content that appeared on your screen, Facebook can maintain a list of scores for all the content you saw.
Facebook content includes news feeds and advertisements. They assigns various labels to the news feeds and vendors who wish to place ads on Facebook can assign approximately 15 label types to their ads such as product type, brand, price, group, gender, size, color, etc... The process of assigning labels to content is also driven by AI algorithms such as clustering and deep learning which will be discussed in other tutorials.
Having observed you in the recent past, Facebook algorithm “knows” what content type scored the highest when presented to you. Given this they look for similar ads and news feeds (i.e. content with similar labels) and display them to you because their algorithm has determined that you like this type of content. For example if you watched videos with cats, liked them, added comments and shared them with friends the Facebook algorithm will calculate a high score for this content and thus learn that you like cats. It will serve more cat content and cat related ads to you while constantly calculating the scores just to make sure your prior action was not a one time event. As a result they “know” and track your preferences. They also learn whom you shared with and assume these people also may like cats. This information can be used to identify cat people and check their scores on cat content to make sure they indeed like cats.
As Facebook serves new ads and content they continuously update your activity (column 1 in the table) so that they are current with your latest likes and dislikes. Even if you Hide some type content because you don't want to see this type, Facebook will occasionally serve this type of content against your wishes presumably to check if you may have changed your preferences.
By monitoring your reaction to news feeds Facebook learns your politics, sport preferences, wealth or even sexual orientation. By learning whom you share with they can learn who else is likely to have similar characteristics.
This is pure Machine Learning. At no point a human being is involved. The Facebook computers perform all the tasks automatically. It should be noted that statements of the type “Facebook knows” are misleading somewhat. It is their computers that “know” not the company employees.
Examples of applications.
There are many AI application we encounter every day such as Google Map, Uber, Facebook, LinkedIn, Siri, Alexa, Airport Security or Spam detection without realizing it. They combine many aspects of AI and are a result of hundreds of millions if not billions of dollars worth of research that big high-tech companies poured into developing these products.
Businesses who are not high-tech AI developers look for addressing specific business issues by using a variety of methods including AI.
One good example of AI addressing a specific problem is Unilever's use of AI in the recruiting process. Here is how it worked in a nutshell:
- Received 250,000 resumes from new graduates for their Leadership Program.
- Each applicant played 12 on-line neuroscience games which are based on tests used by scientist in cognitive neuroscience research. The games test 90 cognitive, social and emotional traits such as ability to focus, risk taking, memory, ability to read emotional cues, etc....
- The candidates were also interviewed by AI computers where they answered questions. During that time language processing and machine vision algorithms evaluated the answers and body language of each candidate. This reduced the number of candidates to 3,500.
- The 3,500 candidates were invited for interviews by humans and 800 were hired.
- The process saved 70,000 hours of recruiting time had it been done using the standard approach of resume screening, phone and in-person interviews.
While there are no results on how effective this process was in terms of quality or turnover of hired employees it addresses at least one problem of dealing with the large quantities of unqualified resumes which accompany any recruiting activity.
Every aspect of a business is likely to be affected or disrupted by Machine Learning/AI. Product Recommendation, Sales Process Automation, Fraud Detection, Inventory Optimization, Risk Assessment , Sentiment Analysis, Marketing Intelligence and Optimization, Customer Service (Chatbots) are some and will be discussed in other tutorials.
Key considerations if you're thinking of deploying this in your business.
- Should focus on a narrow problem of the type mentioned above. The narrower the better.
- AI needs data to train its algorithms and then to supply them with fresh data to perform decisions and to continue learning. A bare minimum is 1,000 items but on average count on 10,000 – 100,000. Hard problems can exceed 1,000,000. In general complex problems are complex because there are many variables and it is hard to determine which variable will be most important to a given situation. Predicting which political candidate will win is complex for obvious reasons but predicting which product is likely to be bought by a customer who is interested in baseball is much easier.
- It will take experimentation so plan on a pilot. Not much point in committing to more until the experiment is done and results look promising.
- AI solutions are not like buying a CRM or Cyber Security software. Typically it is not an off the shelf product that you install. It comes as a set of tools, great tools for sure but still tools. If you're not handy the best plumbing tool set will not help you repair your plumbing.
- AI takes a lot of training data. Do you have it or can get it?
- What concrete measurements you will use to determine its ROI potential? Increasing revenue if you're considering a product recommender, greater customer handling capacity if you're planning a chatbot for your service desk, lower production costs?
Key questions you should be asking of vendors.
The list of questions depends on your industry and application of AI you envision for your business. We provide more specific questions in tutorials covering these topics. Here are some that appear universally:
- Do I need AI to address my business issue or can it be done in an easier way. As stated earlier, AI is not not cure all general intelligence the type a good team of employees is. It can be much better than the best team however at narrowly defined tasks.
- How much data and in what format do I need and do I have enough for a meaningful AI outcome?
- Is there training data I can buy?
- How much time and cost to deploy a Pilot? Describe a typical deployment timeline.
- What are you providing and what are we responsible for? Buying a bunch of AI tools/software is not likely to do much unless you have access to people with AI experience. AI specialists make between $300,000 - $500,000 in Silicon Valley because there is a shortage of them.
- How much human help is needed for the system to operate. For example, Netflix movie recommender system relies on 40 human experts to categorize each new show. It implements very advanced algorithms but they cannot function without human help.
- What specific business improvement can I expect and when?
AI by the numbers.
- Most companies process only 10% - 15% of their data to extract business value 1
- 80% of merchants in China have AI tools 2
- 1.4% of merchant inventory ( nearly as much as their profits) disappears annually, 66% from theft. Several vendors offer AI solution to theft problems 3
- 128% and 89% are the incremental value Travel and Transport industries respectively can create using AI vs traditional tools4
- 88% of 500 surveyed Healthcare executives said they increased their spending on AI in 2019 5
One area where every business wants to improve is sales. It is common knowledge that in Business-to-Business (B2B) experience of sales reps is key. B2B sales is typically complex and the selling cycle is long. Navigating the sales process takes skill and knowledge. Training sales people and forecasting new business is always a challenge.
It seems that this would be a good place to use Machine Learning in order to capture the historical data, external selling training material, combine it with new inputs and come up with the next best action a sales rep can take in order to maximize the probability of success.
We identified a company called Olono who claims to :
“ …. deliver Next Best Actions that help first line managers coach reps. The result is reduced risk, increased win rates and shorter sales cycles... B2B sales is 100+ steps, processes and actions. And every step can change your buyer’s journey. For even the best reps, staying on top of it all is nearly impossible. That’s where we come in.
Working with your existing CRM investments, Olono aggregates all your sales activity data and delivers Next Best Actions that are aligned with your unique sales methodology, to drive consistent, predictable sales execution.
Through machine learning, Olono correlates people, processes and data from across your sales tools, then learns patterns of success and risks, optimizing your buyer’s journey. The more you use Olono, the more tailored your actions”
This sounds very promising but that is all the info you get from their web site. Good luck finding answers to the questions listed above or to find if their product will actually work in your environment and how. You want concrete information? Sign up for a demo and talk to their sales people. That's OK but there are over 50 companies that seem to offer similar solutions. Clearly not all would be a fit but which ones.
How to find an AI solution for your business
We came up with a better way to find AI solutions than searching the internet, wading through vendors marketing fluff and hype and dealing with persistent sales people. We created a portal (www.180find.com) where you can publish your needs anonymously and let vendors submit their best proposals.
You can still communicate with vendors, ask questions, respond to their comments but stay invisible until you're ready. It is like a dating service, you go off-line once you find something promising.
Our “dating service” works as follows:
- You describe your business needs
- List benefits you expect
- Submit it and we'll publish it anonymously
- We invite all the relevant vendors
- You communicate directly with vendors who respond
- Remain anonymous as long as you wish
- Once you find something take the conversation off -line
- It is free
We are unbiased, don't feature or promote any vendor. We ask all the relevant vendors to look at your project and come up with a proposal.
We track news, reports and tutorials on AI deployment in variety of industries. You can visit us to stay informed.