AI for Everyone Presentation.pptx

AI for Everyone Presidential Initiative for Artificial Intelligence and Computing Objectives of this Course 1. The meaning behind common AI terminology, including neural networks, machine learning, deep learning, and data science 2. What AI realistically can--an...

AI for Everyone Presidential Initiative for Artificial Intelligence and Computing Objectives of this Course 1. The meaning behind common AI terminology, including neural networks, machine learning, deep learning, and data science 2. What AI realistically can--and cannot--do 3. How to spot opportunities to apply AI to problems in your own organization 4. What it feels like to build machine learning and data science projects 5. How to work with an AI team and build an AI strategy in your $13 Trillion AI value creation by 2030 Source: McKinsey Global Institute A lot of the value created by AI will be outside the software industry. AI will have a huge impact on all the major industries. There are 2 types of AI ALMOST NO ANI PROGRESS Artificial Narrow Intelligence AGI Artificial General LOTS OF PROGRESS Intelligence Artificial Narrow Intelligence (ANI) These are AIs that do one thing such as: smart speaker self-driving car AI to do web search AI applications in farming or in a factory. These types of AI are one trick ponies but when you find the appropriate trick, this can be incredibly valuable. Artificial General Intelligence (AGI) That is the goal to build AI. They can do anything a human can do or maybe even be superintelligent and do even more things than any human can. Progress in ANI vs AGI The rapid progress in ANI has caused people to conclude that there's a lot of progress in AI, which is true. But that has caused people to falsely think that there might be a lot of progress in AGI as well which is leading to some irrational fears about evil clever robots coming over to take over humanity anytime now. Achieving AGI Will Take Time AGI is an exciting goal for researchers to work on, but it requires many technological breakthroughs before we get there and it may be decades or hundreds of years or even thousands of years away.  Machine Learning The rise of AI has been largely driven by one tool in AI called machine learning. You will start thinking how machine learning might be applied to your company or to your industry. Supervised Learning If the input is an audio clip, and the AI's job is to output the text transcript, then this is speech recognition. Input (A) Output (B) Application Audio Text (0/1) Speech Recognition Supervised Learning If you want to input English and have it output a different language, Chinese, Spanish, something else, then this is machine translation. Input (A) Output (B) Application English Chinese Machine Translation Supervised Learning All the large online ad platforms have a piece of AI that inputs some information about an ad, and some information about you, and tries to predict, will you click on this ad or not? Input (A) Output (B) Application Ad + User Click? Machine Info (0/1) Translation Supervised Learning If you want to build a self-driving car, one of the key pieces of AI is the AI that takes as input an image, and some information from radar, or from other sensors, and outputs the position of other cars, so your self-driving car can avoid the other cars. Input(A) Position of other Self-Driving Car Image, radar cars info Supervised Learning In Manufacturing, we take as input a picture of something you've just manufactured, such as a picture of a cell phone coming off the assembly line., and you want to output, is there a scratch, or is there a dent, or some other defects on this thing you've just manufactured? This is visual inspection which is helping manufacturers to reduce or prevent defects in the things that they're making. Input (A) Output (B) Image of a phone Defects (0/1) Visual Inspection Supervised Learning This set of AI called supervised learning, just learns input to output, or A to B mappings. On one hand, input to output, A to B it seems quite limiting. But when you find a right application scenario, this can be incredibly valuable. Why Now? Why Now? Why Now? Why Now? Why Now? Why Now? The Rise of Fast Computers So, the rise of fast computers with specialized processors such as graphics processing units or GPUs has enabled many companies, not just giant tech companies, but many many other companies to be able to train large neural nets on a large enough amount of data in order to get very good performance and drive business value. What is the most important idea in AI? Machine Learning What is Supervised Learning?  A to B mappings Input to Output mappings What enables machine learning to work so well? What is Data A Table of Data (Dataset) Size of House (Square Feet) Price ($1000) 523 115 645 150 708 210 1034 280 2290 355 2545 440 A B A Table of Data (Dataset) Size of House (Square Feet) # of Bedrooms Price ($1000) 523 1 115 645 1 150 708 2 210 1034 3 280 2290 4 355 2545 4 440 A B Data is often unique to your business Data is often unique to your business, and this is an example of a dataset that a real estate agency might have that they tried to help price houses. It's up to you to decide what is A and what is B, and how to choose these definitions of A and B to make it valuable for your business. Another example If you have a certain budget and you want to decide what is the size of house you can afford, then you might decide that the input A is how much does someone spend and B is just the size of the house in square feet, and that would be a totally different choice of A and B that tells you, given a certain budget, what's the size of the house you should be maybe looking at. A Table of Data (Dataset) Size of House (Square Feet) # of Bedrooms Price ($1000) 523 1 115 645 1 150 708 2 210 1034 3 280 2290 4 355 2545 4 440 B A Acquiring data Manual labeling Acquiring data From observing behaviors of humans User ID Time Price ($) Purchased 4783 Jan 21 08:15.20 7.95 yes 3893 Mar 3 11:30.15 10.00 yes 8384 Jun 11 14:15.05 9.50 no 0931 Aug 2 20:30.55 12.90 yes Acquiring data From observing behaviors of machines Machine Temperature Pressure (psi) Machine Fault 17987 60 7.65 N 34672 100 25.50 N 08542 140 75.50 Y 98536 165 125 Y Input A Input B Acquiring data Download from websites / partnerships ○ Thanks to the open internet you can find so many datasets available for free online Computer vision or image datasets Self driving car datasets Speech recognition datasets Medical imaging datasets ○ Keep in mind licensing and copyright Use and misuse Give me three years to build up my IT team, we're of data collecting so much data. Then after three years, I'll have this perfect dataset. We'll do AI then. What’s wrong with this approach? Use and misuse It turns out that's a really bad strategy. of data Once you've started collecting some data, go ahead and start showing it or feeding it to an AI team. Then the AI team can give feedback to your IT team on what types of data to collect and what types of IT infrastructure to keep on Example Maybe an AI team can look at your Machine Temperature Pressure (psi) Machine Fault factory data and say, "Hey. You know 17987 60 7.65 N what? If you can collect data from this 34672 100 25.50 N big manufacturing machine, not just 08542 140 75.50 Y once every ten minutes, but instead once every one minute, then we could 98536 165 125 Y do a much better job building a Input A Input B preventative maintenance systems for you.” Use and misuse "Hey, I have so much data. Surely, an AI team can make it of data valuable." What’s wrong with this statement? Use and misuse Unfortunately, this doesn't always work out. of data More data is usually better than less data, but I wouldn't take it for granted that just because you have many terabytes or gigabytes of data, that an AI team can actually make that valuable. Don't throw data at an AI team and assume it will be Data is Messy Not a Not a cat cat Cat Cat  If you have bad data, then Data is Messy the AI will learn inaccurate things. Data problems: Incorrect labels Missing values Multiple types of data Unstructured Data: Images, audio, text  Size of House # of Price Example (Square Feet) Bedrooms ($1000) 523 1 115 You can have incorrect labels or just 645 1 0.001 incorrect data. For example, this 708 unknown 210 house is probably not going to sell for 1034 3 unknown $0.1 just for one dollar. unknown 4 355 Or, data can also have missing values 2545 unknown 440 such as we have here a whole bunch of unknown values. This is structured data. Machine Learning vs Data Science Size of House (Square Feet) # of Bedrooms # of Bathrooms Newly Renovated Price ($1000) 523 1 2 N 115 645 1 3 N 150 708 2 1 N 210 1034 3 3 Y 280 2290 4 4 N 355 2545 4 5 Y 440 A B Running AI System A software that which automatically returns output B for input A. If you have an AI system running, serving dozens or hundreds of thousands or millions of users, that's usually a machine-learning system. Data Science If you want to have a team analyze your dataset in order to gain insights. The output of a data science project is a set of insights that can help you make business decisions So, a team might come up with conclusions like: "Hey, did you know if you have two houses of a similar size, they've a similar square footage, if the house has three bedrooms, then they cost a lot more than the house of two bedrooms, even if the square for this is the same." Data Science "Did you know that newly renovated homes have a 15% premium, and this can help you make decisions such as, given a similar square footage, do you want to build a two bedroom or three bedroom size in order to maximize value? " "Is it worth an investment to renovate a home in the hope that the renovation increases the price you can sell a house for?" The output of a data science project is a set of insights that can help you make business decisions, such as what type of house to build or whether to invest in renovation. Machine Learning vs Data Science Machine Learning “Field of study that gives computers the ability to learn without being explicitly programmed.” - Arthur Samuel (1959) A machine learning project will often result in a piece of software Formal Definition of Data Science Data science is the science of extracting knowledge and insights from data. So, the output of a data science project is often a slide deck, the presentation summarizes conclusions for executives to take business actions or summarizes conclusions for a product team to decide how to improve a website. Example of ML vs DS in the online ad industry Large platforms have AI that quickly tells them what’s the ad you’re most likely to click on. This is a machine learning system. It inputs information about the user and about the ad and outputs whether the user will click on the ad or not. These systems run 24/7 and drive ad revenue for these platforms. Example of ML vs DS in the online ad industry If analyzing data tells you, for example, that the travel industry is not buying a lot of ads, but if you send more salespeople to sell ads to travel companies, you could convince them to use more advertising, then that would be an example of a data science project. The data science conclusion results in the executives deciding to ask a sales team to spend more time reaching out to the travel industry. Deep Learning AI and related disciplines Machine Learning Data Science Deep Learning / Neural Network Supervised Learning Un supervised learning Reinforcement Learning What makes a company AI company? Strategic data acquisition Unified datawarehouse Pervasive automation New roles such as MLE AI Transformation Deciding about a new project Technical diligence Is it feasible project? Can AI do that? Pretty much any thing you can do with a second of thought can be automated using supervised learning Supervised learning tasks Examples of what ML can and can’t do? Identifying the intent of the customer - Possible Writing an emphatic response to customer’s email – Not possible or difficult Technical diligence rules You are learning a simple concept Do you have large training data More examples Self driving car ○ Input is from sensors, camera ○ Output where are the other cars Recognizing gesture of traffic police, construction work, people– not possible Critical application requires good accuracy X-ray diagnosis Diagnosing a disease from X-ray images– possible Diagnosing a disease after reading a book Strengths and weakness of ML Works when, Learning a simple concept Lots of data available Doesn’t work when, Learning a complex concept Asked to work on new type of data such as X-ray images in different conditions and angles Demand prediction based on price Price -> Demand can be modeled using a neural network using a neuron (Perceptron model) Network of neurons (ANN) ○ Price ○ Shipping Cost ○ Marketing ○ Meterial Face recognition Pictures comprise pixels Color images and channels A neural network corresponds to pixels Earlier layers will detect edges, then lobes and then objects Speech Recognition Key steps of Echo / Alexa Collect data Labelled voice Train model Iterate many times Deploy the model Get more data and update model Machine Learning changing job functions Sales Identifying sales opportunities Prioritizing Manufacturing line manager Optimize manufacturing Machine learning can spot defects Recruiting Identify how people prefer recruitment Spot good candidates Marketing Optimize website A/B testing Recommendation system ○Agriculture What to plant? Precision agriculture How to chose an AI project? Brainstorming framework Automate task rather than job Automating call center: picking phone, emails, issue refund, call routing Automating radiologist: X-ray, mentoring other doctors, consulting, Main drivers of business value What are the main pain points in your business? Is it always necessary to have big data? Having more data is good With small datasets you can make progress 10, 100 or 1000 data points can be a good start Ethical diligence Is this going to make society better? Build Vs Buy ML projects can be inhoused or outsourced DS projects are generally inhoused Buy industry standard, only build specialized products How to work with AI team Specify your acceptance criteria 95% accuracy Training, validation and Test dataset Don’t expect 100% accuracy Limitations of ML Insufficient data Mislabeled data Ambiguous labels (human perception) Machine Learning frameworks CPU Vs GPU Edge Deployme nt Case Studies Steps or AI pipeline Trigger word: Hey Device Speech Recognition: Tell me a joke Intent Recognition: joke, time, music, weather Log of training instances, variation in text Execute joke Activity Hey device, set timer for 10 minutes ○ What is the intent? ○ Extract duration ○ What command is to execute Smart speaker functions Play music Volume up/ down Make call Current time Units conversion Simple question These specialized execution routines are written by software engineer Self driving car AI teams AI team may have 100s of engineers A small team can have four or five members Example roles Software Engineers Execute joke, Set timer Machine Learning Engineer Machine Learning Researcher Extend state-of-the-art Applied ML scientist in between ML researcher and ML Engineer Data Scientist Provide insights  Data Engineer ○ Organize data ○ Data is saved in cost effective way ○ We have lot of data, scalability is important AI Product Manager ○ What to build and feasible Get started with a small team 1 Software engineer 1 ML engineer / Data scientist No body but your self AI Transformation playbook Execute a pilot project to gain momentum Build an in-house AI team Provide broad AI training Develop an AI strategy Develop internal and external communication Execute pilot project Start the fly wheel Show traction with in 6-12 months Can be in-house or out-sourced Build an in-house AI team Develop tools that could be useful company wide Under CIO, CTO,CDO, CAIO Provide broad AI training Resources Online courses Books Curate rather than create content Develop an AI strategy Leverage AI to create an advantage specific to your company Design strategy that align with virtuous cycle of AI Blue River – precision agriculture AI needs to be specialized or verticalized to your industry sectory Don’t compete with giants Creating a strategy Strategic data acquisition Unified data warehouse – Pull data into single repository, software can connect the dots Create network effect and platform advantages Uber, Careem, Facebook  Low cost strategy High value strategy Develop internal and external communications Investor relations Government relations Consumer / user education Talent / recruitment Internal communication Common pitfalls Take your first step Get friends to learn about AI Start brainstorming projects Hire a few ML / DS to help Hire or appoint an AI leader (VP AI, CAIO) Discuss with CEO about possibilities of AI transformation AI Application areas Computer Vision ○ Image Classification / Object recognition Face recognition ○ Object detection ○ Image segmentation ○ Tracking Natural language processing Text classification (Spam / Non spam) Sentiment recognition Information retrieval Web search Named entity recognition Machine translation Part of speech tagging  Parsing Speech Speech to text Trigger / wake word detection Speaker ID Speech synthesis (text-to-speech / TTS) Robotics General machine learning Unsupervised learning  Supervised learning needs lot of data 10,000 defected coffee mug, human can easily do that with few examples Transfer learning Reinforcement learning Also useful in Games Not as much as economic value as supervised learning Generative Adversarial Network (GAN) Synthesize new images from scratch Entertainment industry, film, animation Knowledge graph AI & Society AI is super power Goldilock rule Neither too optimistic nor pessimistic Don’t over spend on unnecessary danger AI winter AI can’t do every thing, but will transform industries Limitations of AI Performance limitations With small amount of data Explainability is hard (sometimes doable): How should we trust Humans are also not good at explaining Barrier to acceptance Biased through biased data Adversarial attacks AI can learn unhealthy stereotype Learn from internet ○ Man: Woman as Father: Mother ○ Man: Woman as King: Queen ○ Man : Computer programmer as women: Home maker ○ Man and woman can equally become programmer Why bias matters Hiring tool that discriminates against woman Facial recognition working better for specific ehtnicity Bank loan approavals Toxic effect of reinforcing unhealthy stereotypes Combating bias Technical solution Zero out bias Use less biased or more inclusive data Transparency or auditing process Diverse workforce Creates less biased applications Adversarial attacks Spam Filters Hate speech filter Physical attacks Adversarial defenses Cost to defend Slow speed May not be any incentive to attack, so should we invest in defense? Zero-sum against adversaries Adverse uses of AI Deep Fakes Synthesizing videos Video of Obama ○Undermining of democracy and privacy Oppressive surveillance ○Generating fake comments ○Spam Vs anti-spam, Fraud Vs anti-fraud AI & Developing economy Developing economies gradually moved up the ladder Lower end ladder are susceptible to automation such as agriculture Trampoline to move higher rungs Leapfrog Example of mobile phone Mobile payments ○Online education How developing economies can build AI? US and China leading But AI communities are still immature Focus on AI to strengthen country’s vertical industries Instead of focusing on AI in general, use AI where you are already good at Public private partnership Invest in education AI and impact on jobs AI is automation on steroids Jobs displaced by 2030 400-800 mn Jobs created by 2030 555- 890 mn Is your job amenable to automation? Some solutions to counter AI impact on jobs Conditional basic income: provide a safety net Life long learning society Political solutions ○ Legalization Work at intersection of your current joband AI Summary What is AI? Building AI projects Building AI in your company AI & society