Prediction Machines

• Prediction Machines talk about AI from the functional and Business perspective.
• There are 5 parts for a prediction machine.
• Prediction
• Decision Making
• Tools
• Strategy
• Society

• Introduction
• AI involves trade off-more speed,less accuracy - more autonomy less control, more data less privacy - they provide the method to identify trade off,so that reader can make a suitable decision.
• So AI is a strategic decision making process.
• Everyone has their AI moment.
• AlphaGo AI application won world's no 1 player award.
• AlphaGo application is built on deep mine.
• As artificial light became cheaper and cheaper,everyone uses the artificial light in abundance. Same way prediction is becoming cheaper and cheaper.
• When prediction becomes cheap,it increases the value of other things, like when coffee is cheap,then the value of Sugar and Cream increases or when drop in autonomous vehicle prediction decreases,the value of sensors which capture the data increases.
• Complements value go up.
• Supplements value go down.
• If prediction is accurate, it gives great potential to organization - for example if Amazon could develop more accurate prediction, it could even change the business model from shopping then shipping to shipping then shopping.
• Amazon can predict before hand how much users are going to buy.

Prediction

• Prediction takes information you have called "data" and uses it to generate information you don't have.
• This technology is called "Artificial Intelligence" because the prediction accuracy has increased,the system is becoming more effective in learning and solutions are capable of handling tasks which are requiring "Human Intelligence" earlier.
• Data is new the new oil-Prediction machines rely on data.More and better data leads to better predictions.As prediction becomes cheaper, data becomes more valuable.
• With AI data plays three roles(3 key data's for AI )
• Input Data
• Training Data
• Feedback Data
• In machine learning - Machines learn from data.More accurate predictions would need more data and data acquisition could be expensive.Hence, this is a trade off.
• Organizations need to balance between adding more data,enhance prediction accuracy,and increasing value creation.
• Where human and where computers could predict efficiently.
• known knowns- we have good data , so machines are better here.
• known unknowns - less data, humans are better than machines sometimes.
• unknown unknowns - machines cannot predict, human judgement is required, there was no earlier scenarios like that.
• unknown knowns - machine could provide wrong answer with high accuracy,humans can verify machines prediction and may be able to find the mistake.
• Prediction by Exception
• When machine is not able to predict accurately due to exceptions(like unique situation or lack of data etc).Then it can escalate to humans to make the predictions.

Decision Making

• Prediction is not a decision.Making decisions requires applying judgement to a decision.
• The anatomy of a decision Input ->{training,predictions,judgement}->action->outcome..->..feedback to training again.
• As prediction becomes better,faster,cheaper we will use of more of it to make decisions. Having better predictions raises the value of judgement, machines don't judge only humans do.As machines predict better, faster and cheaper, more decisions need to be made by humans.
• Basic decisions could be hard coded(like solving basic support questions through machines).This is difficult for uncertain situations. Hence uncertainty increases the cost of judgement's.
• Reward function engineering-Job of determining rewards to various actions based on these prediction is mentioned here as a reward function engineering.For example in self driving vehicles reward function engineering is done through hard coding judgement - once perdition is made action is immediate.
•  Generally humans used to make prediction and make judgement.As machines get better at prediction,the role of reward function engineering is increasingly important.
• Will humans be pushed out?No, machine predicts based on data.If no data or less data, humans are still needed. Humans have three types of data.
• humans senses are powerful
• humans preferences are valuable
• privacy constraints restricts data access to machines.

Anatomy of a task

• Different components of a task are as follows.

• We have data.
• On the basis of data System's make prediction.
• Humans make a judgement.
• Some standard judgments are done by system.
• This judgement results into action.
• This action will have an outcome.
• The outcome produces a feedback that goes back into prediction through training data.
• We need to figure out from this what decisions we need to hard code for example in a self driving car the decision have to be taken quickly.So we may need to hard code the decision.This is called as reversed function engineering concept.
• Various actions are taken based on the prediction and judgement.What is the reward for the action and how the reward is taken care off.

Tools

• These tools are not technology tools but AI canvas which helps to decompose process and workflows and take a better decisions.
• Tasks are collection of decisions - decisions are based on prediction and judgement and informed by data. The decisions within a task often share these elements in common, where they differ in the action that follows.Sometimes we can automate all the decisions within a task.
• By decomposing workflows, businesses can access weather prediction machines are likely to reach well beyond that individual decision for which they have been designed.
• Decomposing decisions - AI canvas - contains prediction - judgement - actions - outcome - input - training - feedback.
• Once we identify the tasks which can be done by prediction machines, we must reconstitute tasks into jobs.The jobs may change due to the automation and humans involved may need to be re-skilled.
• Automation that eliminates the human from a task does not necessarily eliminate them from a job.

Components of an AI Canvas

• Prediction
• What do you need to know to make a decision?
• Judgement
• How do you value different outcomes and errors?
• Action
• What are you trying?
• Outcome
• What are your metrics for task success?
• Input
• What data you need to run the prediction algorithm?
• Training
• What data do you need to train the prediction algorithm?
• Feedback
• How can you use the outcomes to improve the algorithm?

Strategy

• How AI can change business strategy.
• Three aspects
• Strategic dilemma or trade off must exist.
• The problem can be resolved by reducing uncertainty
• Companies require a prediction machine that can reduce uncertainty enough to change the balance in the strategic dilemma.
• Strategy at the end leads to value creation for business.
• Google has more than 1000 AI development projects in flight.
• Better predictions enable managers to take decisions that are closer to organization objectives.
• Strategy is also about creating value, that better prediction creates.
• Data alone is not strategic advantages 
• There are three types of data
• input data,training data,feedback data.
• training data used to train the machine, input powers it to predict and feedback is about the performance.
• once machine is well trained, training data is no longer useful.
• Prediction machine reduces uncertainty.
• AI will increase the value of complements of prediction machines.
• It is a balance act between (data,AI and humans) - and if predictions are accurate, organization can take a better decision to use partners or outsource etc.
• So being in control could be achieved even if things are outsourced, as long as predictions are reasonably good in line with business requirements.
• AI learning takes time, and large companies may find it difficult to adopt, as they don't want to take risks in AI innovation as it will be initially have inferior performance,risking their customer experience at times.
• Startups may be able to do this with their small set of customers, improve AI, make AI better and even penetrate customers of established organizations and dominate.
• Deploying AI could be major strategic decision- regarding prediction machines-when to do training and when to deploy for real work.Consider factors like tolerance of error, how important to capture user data in real world.
• Learning's often requires customers who are willing to give data, there is a balance between privacy and data availability.
• Human Learning is also important, what happens is our children use fully autonomous vehicles with least manual driving experience and if something goes wrong in autopilot mode of car, will these people be able to manually manage the situation? This requires reducing automation to ensure that human skills are maintained too.
• Liability Risk To prevent liability issues, and to avoid being discriminatory,if you discover unintentional discrimination in the output of your AI you need to fix it.Black box of AI is not an excuse to ignore potential discrimination.
• Quality Risk  AI trained on lot of data may not be able to predict what happens with fewer data.Human involvement may be required.
• Security Risk Prediction machines fed on input data,they combine this data with a model to generate predictions.Hackers may manipulate the data or the model or learn from outcome of the model and re-engineer the model or the feed the outcome to hacker's own model as input's.
• Trade off between risks involved and the implementation of AI is also key.

Society

• Once AI is better than humans at a particular task,job losses may happen quickly.
• If machines share of work continues to increase, workers income will fall and the accruing to the owners of AI will rise.
• AI gives multiple advantages if a nation
• Spends heavily
• Has more people and hence more data.
• More data accessibility due to its lack of privacy protection to citizens.