Artificial Intelligence (2016) - AMP New Ventures
7 likes2,302 views
Artificial intelligence (AI) has evolved from a theoretical concept to a practical technology that enhances various industries, including healthcare, agriculture, and customer service. Key developments in AI include advancements in machine learning, natural language processing, and computer vision, which enable tasks such as automated decision-making and improved human-computer interaction. However, concerns about AI's potential risks, including job displacement and existential threats, underscore the need for responsible management as AI technology continues to grow.
Artificial Intelligence (2016) - AMP New Ventures
- 1. AMP New Ventures Perspective on Artificial Intelligence September 2016
- 2. 2 Artificial Intelligence is already everywhere. It powers our smartphones, drives our cars and sorts our newsfeeds. Companies globally and across industries are participating in the race for true AI, to reduce operational costs, make faster, more accurate decisions and personalise customer experiences. Perspective
- 3. Sections 1. Definition 2. Branches 3. Applications 4. Why now 5. Risks 6. Startups (Examples) 3
- 4. 4 Artificial Intelligence The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition and decision-making Branches Applications Why nowDefinition RIsks Startup Examples
- 5. 5 Recent leaps of progress in AI has triggered an explosion of startups Source: Venture Scanner
- 6. 1956 John McCarthy coins ‘Artificial Intelligence’ at Dartmouth Conference Theory 6 1950 Alan Turing publishes paper on concept of machine intelligence 1995 US Department of Defence uses Predator UAV in Balkan war 1997 IBM’s Deep Blue wins chess against World Champion Gary Kasparov 2011 IBM Watson computer defeats Jeopardy game show champions 2011 Debut of Virtual Assistants Apple Siri and Microsoft Cortana Jan 2014 Deep Mind Team’s algorithm wins Atari games May 2015 Google self-driving cars complete 1M miles autonomously June 2015 Deep Mind teaches program how to read AI has materialised from a theory in 1950 to widespread technological applications that we use today in our daily lives March 2016 AlphaGo beats Go Grandmaster Lee Sedol in a 5 game series. Strong AI: Machine Learning Selected Milestones of AI PresentWeak AI: Expert Systems Deep Learning
- 7. RIsksWhy nowBranches ApplicationsDefinition 7 Branches Artificial Intelligence Startups examples
- 8. 8 Machine Learning (Learn) Computer Vision (See) Speech Recognition (Hear) Natural Language Processing (NLP) (Communicate) Expert Systems (Think) Motion Planning (Move) AI can be split by unique human capabilities
- 9. AI listens, thinks and communicates... 9 Speech Recognition is the process of mapping audio speech data to textual sentences or key phrases. As humans can speak 150 words per minute on average, but can only type 40, speech recognition has great potential in computer efficiency. As more voice usage data becomes available, speech recognition accuracy will get better and better. In 2010, accuracy for technology companies hovered around 70, and today sits between 95 and 99. Natural language processing (NLP) focuses on human– computer interaction, enabling computers to derive meaning from human language input; and also generate natural language responses. Today, machines proficiently understand natural language syntax but face great challenge in interpreting sentiment (i.e. sarcasm, excitement). Expert Systems emulate human expert decision-making abilities. It allows the computer to solve for complex problems by reasoning about knowledge, navigating if–then rules. (Communicate) (Think) (Listen) From the creators of Siri, Viv enables developers to create anything on top of its, conversational interface, making ‘her’ smarter.
- 10. Sees, moves and learns... 10 Computer vision is the ability to electronically perceive and understand image/video sources, extract meaningful information and take action. Up until now, image recognition has been driven by rules-based categorisation. Today, machines are fed data so they build their own vision. Motion Planning is the process of forming a strategy of action sequences to achieve a desired movement, typically for execution by intelligent agents, autonomous robots and unmanned vehicles. Today, we are at advanced levels of simple motion planning problems, such as ‘move from A to B, while avoiding collision with any obstacles.’ Machine learning is training computers with datasets to recognise patterns, develop algorithms and self-improve. Machine Learning has been central to today’s unprecedented momentum in AI, as it enables the progress of other AI branches. (See) (Move) (Think)
- 11. 11 Machine Learning techniques are used to create self-learning capabilities 1. Raw Data is formatted and cleaned so scientific conclusions can be drawn without error/skew. Accuracy and insights increase with relevance and amount of data. 2. Algorithms are applied for statistical analysis. This includes things like regression models and decision trees. The results are examined and algorithms are re- iterated until a best model emerges that produces the most useful results. Under the hood 3. A Chosen Model is now used to produce probability scores (usually between 0 and 1) that can be used to make decisions, solve problems and trigger actions. Source: Azure Supervised Learning :Data is labelled and there is a specific outcome Unsupervised Learning : Insights are drawn from data without a specific purpose
- 12. 12 The goal of AI is to create Strong and Broad platforms using Machine Learning techniques Strong AI Weak AI Executes tasks within a rules-based programmed domain Narrow AI Built to perform limited, specific tasks Broad AI Systems that can be applied to many contexts Self-improves through Machine Learning based on raw input Goal Knows one thing and improves Knows many things and improves Knows many thingsKnows one or limited things
- 13. RIsksWhy nowApplicationsBranchesDefinition 13 Applications Artificial Intelligence Startup examples
- 14. Agriculture • Drone vision to monitor crop conditions like water stress, nutrient condition, plant population, soil moisture content etc. • Predicting pest and disease outbreaks using data • Drones capable of delivering customized fertilizers and pesticides based on the requirement of each plant • Autonomous GPS guided harvesting systems • Facial recognition for livestock (e.g. cows) 14 Healthcare • Expert systems to instantly weigh factors in a patients circumstance and shortlist possible diagnoses with confidence ratings • Surgery Robotics to assist in the operating theatre • Virtual nurses and remote patient monitoring • Data to streamline the selection process of drug development to show investigators which developments show the most promise • Insight and pattern induction from huge data deposits from connected devices Military • Unmanned drones providing sustained surveillance and swift precise attacks on high-value targets • Small robots are used for missions to counter improvised explosive devices • Systems for faster collection and information analysis to improve reaction and decision-making time to implement effective military actions • Smart pilot helmets (e.g. F35 fighter jet helmet) Manufacturing • Computer vision with robotics to automate assembly line tasks • Computer vision and machine learning to track and isolate physical fault causes • Mail routing using computer vision based on human written (and often badly) postal codes • Data-driven rapid prototyping for 3D printing AI is being adopted across all industries
- 15. 15 Customer Service Chat Bots NLP powered chat bots used to answer general FAQ and action simple tasks, reducing volumes and waiting times for customers Predictive Credit Analysis Machine Learning algorithms are applied to credit scores and other personal data to assess risk for loan applications and loan pools as a whole Insurance Underwriting Underwriting AI systems are used to automate the underwriting process and utilise wider and more granular data such as health and social media Personal Budgeting AI is used to recognise and report personal spending patterns, detailing location, merchant and spend category. Alerts can be pushed for irregular fees and patterns Algorithmic Trading Investment managers use trading Algorithms to automatically place trades, generating profits at speeds that are humanly impossible Fraud Detection By analysing historical transaction data, models can be built to detect fraudulent patterns. These models can then be applied to real-time financial transactions and be given fraud scores. Operations and Risk Product Sales & Marketing Customer Service Marketing AI used to personalise offers, A/B test advertising content, and decide when is the optimal time to release that content AI is being deployed across the Financial Services value chain Robo-Advice Automated financial advice and investment portfolio rebalancing based on risk profile and life stage
- 16. 16 41 startups bringing AI to Fintech Source: CB Insights
- 17. Branches Why now RIsksApplicationsDefinition 17 Why now? Artificial Intelligence Startup examples
- 18. 18 More Fuel Better Engineering Cheaper Material Improved engines Avalanche of Data Repurposing GPUs Cheaper Computation Stronger AI AI is booming now due to the convolution of more data, the repurposing of GPUs and cheaper computation
- 19. 6bnNetwork connections per person on earth 2.5 People have smartphones. World population = 7bn 30bnPieces of content shared on Facebook every month 19 More fuel (data) Just as human brains require dozens of examples before it can naturally distinguish cats and dogs, Artificial minds require large datasets to upskill in categorisation accuracy. Social networks, mobile phones and wearable devices, powered by improved connectivity and cloud economics, have created an explosion of data to feed AI engines. 90% of all the data in the world being generated in the past 2 years. Why is AI booming now? Avalanche of Data Data is growing at a 40% compound rate, reaching ~45 Zettabytes (ZB*) by 2020. To put things in context, 1 ZB = 1.1 Trillion Gigabytes = 2 billion years of music . More Fuel Better Engineering Cheaper Material Improved engines
- 20. 20 Why is AI booming now? Repurposing of GPUs Better Engineering Up until now, AI applications have needed to process large amounts of data in a sequential pattern, limiting processing speeds.However, n 2009, Andrew Ng’s team at Stanford discovered that GPU (Graphic Processing Units) chips, typically used for gaming, could be organised to run data processes in parallel manner. This is important as ‘neural networks’, the primary architecture of AI software today, require many different processes to take place simultaneously in parallel. To recognise images for example, every pixel must be seen in context to each other, a deeply parallel task. More Fuel Better Engineering Cheaper Material Improved engines CPU: 1-4 Serial Processing Cores GPU:100’s of Parrallel Processing Cores Serial Parallel
- 21. 21 Cheaper material Computational power has steadily become cheaper over the past 50 years as per Moore’s law, which states that overall processing power (number of transistors on an affordable CPU) for computers will double every two years. This is achieved through shrinking transistors, which in turn makes digital devices significantly cheaper and more energy- efficient to power AI applications. Why is AI booming now? Cheaper Computation Power Computer cost/performance (1992 – 2012) Microchip transistor sizes (2000 – 2020) More Fuel Better Engineering Cheaper Material Improved engines
- 22. 22 Source: CB Insights This breakthrough in AI has attracted large amounts of investment, in turn further accelerating growth AI Landscape: Global Yearly Financing History Investments in AI startups have increased nearly 6x to ~400 in 2015, up from ~70 in 2011
- 23. 23 Tech giants (Google, Facebook, Amazon, IBM) are aggressively acquiring AI startups to capture market share Race for AI: Most Active Acquirers in Artificial Intelligence Google is the most active in the space (21 companies) followed by Facebook (10 companies). Source: CB Insights By 2020, the market for machine learning applications will reach ~$40bn and 60% of those applications will run on the platform software of 4 companies (Amazon, IBM, Google and Microsoft)
- 24. Why nowApplicationsDefinition 24 Risks Artificial Intelligence Branches RIsks Startup Examples
- 25. 25 Risks of transferring responsibility and knowledge to AI Existential risk to humanity Futurist and Google’s Director of Engineering Ray Kurzweil, predicts that machines will surpass humanity in intelligence by 2029 and become ‘Superintelligent’, a powerful state that could be difficult to control and pose existential threats to humanity. Other technology leaders such as Bill Gates and Elon Musk have expressed similar concerns. Superintelligence is ranked as the 3rd highest existential threat to humanity, after Bioengineered pandemics and Nuclear war. More serious cyber attacks AI algorithms are equally as susceptible to cyberattacks as regular software. However, because AI algorithms are often depended on to make high-stakes decisions, such as driving cars and controlling robots, the impact of successful cyberattacks on AI systems could be much more devastating than attacks in the past. Elon Musk, Founder of Tesla and SpaceX, tweets concerns about AI Hackers remotely kill Jeep on a highway (July 2015)
- 26. 26 Replacing human jobs Boston Consulting Group predicts that by 2025, up to a quarter of jobs will be replaced by either smart software or robots. The first jobs most likely to be affected are industrial jobs (manufacturing, cleaning), routine information processing tasks (bookkeepers, travel agents) and basic customer service roles (call centres, cashiers). Amplification of bugs The shift from traditional programming to machine learning means that code is often self-produced in neural nets, as opposed to being hand-programmed. While this is much faster, this means the code is harder to audit, and early-stage errors or bugs can be easily amplified if undiscovered. Extra validation measures should be taken with machine learning to achieve high degrees of quality assurance. Microsoft’s Twitterbot ‘Tay’ goes rogue with tweets Jobs requiring empathy and intuition (e.g. psychologists, clergies) are least likely to be threatened by technology. Risks of transferring responsibility and knowledge to AI
- 27. Branches Applications Why nowDefinition 27 Startup examples Artificial Intelligence RIsks Startup Examples
- 28. 28 Banjo Gods eye view Description Banjo delves through public social media posts and uses algorithms to identify deviations from the normal activity at a given location. Apart from breaking news, Banjo’s use cases include things such as track disease outbreaks and predict insurance claim in natural disaster events. Banjo is now used by thousands of news outlets, insurance firms, security contractors and more. How it works The company divided the globe into 35 billion football-field- size squares and spent years determining baseline activity levels for each portion of the virtual grid. Now, any deviation from this baseline triggers an alert to the Banjo team. Why it matters During the Boston bombing on April 15, 2013, the Banjo team were able to instantaneously look at the scene in real time and identify people of interest just minutes after the bombing occurred. Inception: 2011, California (US) Social Media activity heat map used at Banjo HQ Banjo’s computer vision classification Funding to Date: US $121m
- 29. 29 Affectiva Emotion as a service Description Affectiva offers a cloud based solution that reads facial expressions, which it calls “Emotion as a Service”. Its emotion analytics platform ‘Affdex’ is used by one third of Fortune Global 100 companies and over 1,400 brands (Unilever, Kellogg's, MARS etc.) to understand consumer emotional engagement, optimise business processes and improve customer experiences. How it works Affectiva has collected the world's largest repository of emotion data – 3.2 million faces analysed from 75+ countries amounting to more than 12 billion emotion data points. Why it matters Affectiva allows developers to create hyper-personalized experiences across multiple industries. For example, in gaming, developers can create adaptive games that change based on a player’s mood. In healthcare, clinical researchers can develop applications that respond to a patient’s emotional state. Video communication platforms can even modify presentations in real- time, based on an audience’s engagement. Funding to Date: US $33.72m Inception: 2009, Massachusetts (US) Testing advertisement reception using Affectiva software Affectiva’s facial analysis to label emotional states
- 30. 30 Jibo Every family needs a Robot Description Jibo is the world’s first social robot for the home, at 11 inches tall and weighing 3 pounds. It’s uniquely empathetic in the way it takes voice commands, recognises individuals, takes photos/videos, answers queries and more. How it works Jibo uses machine learning, speech and facial recognition, and natural language processing to learn from its interactions with people. Jibo will familiarise with individuals, recognising voice print and appearance, and alter its behaviour accordingly. Why it matters Interest from larger players in the smart home and entertainment fields has grown since Jibo's 2014 reveal. In May 2016, Jibo’s team released an SDK (software developer kit) that allows developers to create their own skills for Jibo. Jibo is a step ahead of Amazon’s Alexa or Apple’s Siri in that it is built to coexist socially with humans, a step closer towards fictional characters such as Starwars’ R2D2. Funding $33.72m (FTD) Location Massachusetts (US) Founded 2009 Founder Rana el Kaliouby Inception: 2012, Massachusetts (US) Funding to Date: US $52.3m Jibo
- 31. 31 Prisma AI with a paintbrush Description Prisma uses machine learning algorithms to instantly transform smartphone into stylized artworks based on unique artistic and graphical styles. How it works Styles are extracted from artworks are mashed with photo data using neural networks on a blank canvas to produce a final new image. This is not to be confused with ‘filters’ as used in Instagram. Why it matters This counters the argument that ‘machines can never develop creativity’, as Prisma’s art has become virally popular. The app is now being used in ~30 countries, with 300,000 installs across 10 of those countries per day. Inception: 2016, Moscow (Russia) Funding to Date: US $1m - $2m
- 32. 32 ZestFinance Big data credit scoring Description Founded by ex-CIO of Google, Douglas Merrill, ZestFinance applies algorithms to thousands of data points to make a credit decision within seconds. Its loan product ‘Basix’ can approve personal loans ($3000 - $5000) in minutes. How it works In evaluating borrowers, ZestFinance pulls data from various credit agencies and other sources, looking at factors such as college attendance, online restaurant ratings, phone bills and even the way you type online. This allows the company to re- create the holistic view of the borrower. Why it matters Alternative credit scoring allows Fintechs to lend to borrowers typically not served by banks due to a lack of credit history. For example, ZestFinance’s ‘Basix’ lends to near-prime borrowers who just miss the cut to borrow from banks. Secondly, the speed of data crunching means loans can be funded to customers within minutes, much faster than traditional bank processes. Finally, according to ZestFinance, ‘all data is credit data’. Inception: 2009, California (US) Funding to Date: US $112m Douglas Merrill, Founder (ZestFinance)