Can we accelerate the adoption of AI in pharma by learning from - PowerPoint PPT Presentation

Can we accelerate the adoption of AI in pharma by learning from experiences in other industries? Dr Sam Genway 9 th October 2018

Tessella, Altran's World Class Center for Analytics We use data science to accelerate evidence-based decision making Predictive Analytics Optimization Chem and Bio Informatics Machine Learning Mathematical Modelling Artificial Intelligence Visualisation Control Theory Image Analysis Statistics Text and Sentiment Analysis Signal Processing DNA EXPERIENCE KNOWLEDGE OPERATIONS Data is in our DNA. 300 35+ years of experience Unique combination of domain UK, US, NL, of the brightest scientific knowledge, data engineering expertise, delivering 1000s of data ES, FR, PT, IT minds, 60+% hold PhDs analytics projects maths & statistics excellence

How does Tessella add value to AI & machine learning projects? Tessella focusses on delivering Tessella understands machine robust, trustworthy AI systems learning from decades of real-world backed by decades of experience in complex domains. Context engineering good practice. Quality Tessella can deliver not only the machine Tessella consultancy can help to identify learning core but also the full end-to-end, most valuable uses for AI and ensure real enabling the transition from proof-of- business value is maximised. concept to sustainable business value. Informed Advice Big Picture

AI Is Not As New As You Think It Is 2009 Google Autonomous Car demonstrates human- equivalent performance 1949 Warren Weaver proposes the idea of 2010s statistical machine translation, which now forms the basis of most translation systems 2006 AI is 50 years 1970s Natural Language 2004 DARPA Grand 2011 IBM Watson beats old as an academic Processing performance Challenge catalyses humans at Jeopardy discipline increases dramatically huge progress in 2012 Siri, Alexa and Google autonomous driving 1980s 2014 Deep Learning Assistant revolutionise our allows image recognition interaction with technology 1955 Arthur Samuel systems to surpass human 1950 Alan Turing proposes creates a chequers system 2002 Roomba 1980 Neural networks performance 2000 Kismet framework for creating and that learns to play and becomes the first used routinely on vision demonstrates evaluating intelligent machines 2015 TensorFlow released human equivalent level mainstream domestic problems emotional by Google Brain AI device 1951 First Neural interactive interfaces Network SNARC is built 2015 AlphaGo becomes 1975 Back- 2000s at MIT world champion at Go propagation kick- 1982 Hopfield 1956 Artificial Intelligence invents recurrent starts neural network 1998 Use of AI neural networks is coined as a term and application 1950s revolutionises web- academic interest explodes 1986 Decision trees search performance 1974 MYCIN AI 1951 Chess and 1973 Vision invented 1997 AI becomes world 2017 AlphaZero becomes system used in Chequers playing Controlled Robots champion at chess world champion Go and medical diagnosis algorithms run for the perform complex Chess by playing only 1986 First autonomous first time tasks 1995 Random Forests against itself. vehicles drive on the invented 1971 First Deep 1986 Hinton et al refine streets 1958 LISP programming Learning Systems use of backprop to achieve 1995 Support Vector language invented to support created 1989 Yann LeCunn uses massive improvements Machines with kernel AI research CNNs to perform 1970s trick character recognition 1959 MIT AI Lab 1994 AI becomes world is founded champion at checkers 1969 Mobile Autonomous 1990s Robots combine together 1960s 1993 1000 layer deep 1989 Watkins multiple AI systems 1965 First Expert network used to solve demonstrates Q-Learning System is produced grammar learning tasks opening the door to practical use of 1961 First Robot works reinforcement learning 1991 TD-Gammon uses on a production line reinforcement learning to 1965 Natural Language Processing able to 1963 Support create championship- understand and solve written problems, and have Vector Machines level backgammon player simple interactive conversations (ELIZA) invented

So what actually is new? New inventions , such as generative adversarial techniques, attention Data mechanisms… Ability to train hugely Big data revolution motivated abundant cheap storage and well-curated data complex networks on vast quantities of data Compute Hardware advances; ultimately emergence New applications of of GPU and TPU as effective hardware for neural networks neural networks Current AI Boom Massive R&D investment by big players (e.g. Google) and academia Widely available libraries and toolkits Algorithmic advances Initially unsupervised pre-training; more Increased public recently – batch normalization, residual understanding and blocks, optimizers… acceptance Massive investment from external sources (VCs etc.) High profile successes (e.g. Watson Jeopardy)

A view of AI maturity Innovate Focus Initial Frenzy Time from invention to application Invention Completely new Initial Frenzy Excitement and Focus As understanding grows Innovate Experts start to make AI methods and enthusiasm causes AI and toolkits mature, it’s easy to custom variants of standard paradigms are invented, techniques to be applied identify good use cases and machine learning models mostly in academia and indiscriminately and apply appropriate machine specific to the particular AI-centric companies inappropriately learning techniques to them problems of the domain. Increasing organisational maturity in AI exploitation

Organisational agility Innovate Focus Time from invention to application Invention Completely new Initial Frenzy Excitement and Focus As understanding grows Innovate Experts start to make AI methods and enthusiasm causes AI and toolkits mature, it’s easy to custom variants of standard paradigms are invented, techniques to be applied identify good use cases and machine learning models mostly in academia and indiscriminately and apply appropriate machine specific to the particular AI-centric companies inappropriately learning techniques to them problems of the domain. Potentially motivates new…

A look across industries Time from invention to application Invention Completely new Initial Frenzy Excitement and Focus As understanding grows Innovate Experts start to make AI methods and enthusiasm causes AI and toolkits mature, it’s easy to custom variants of standard paradigms are invented, techniques to be applied identify good use cases and machine learning models mostly in academia and indiscriminately and apply appropriate machine specific to the particular AI-centric companies inappropriately learning techniques to them problems of the domain. Potentially motivates new…

Automotive Challenge of adversarial images and edge-case misclassifications. Development of testing procedures to mitigate risk of Automated machine learning, dangerous errors. capturing focused training data to improve performance. Immediate application focused on image and video analysis; at the forefront of image segmentation; already in products. Time from invention to application An industry where problem-specific challenges are necessitating research which will benefit other AI applications

Energy Operational insight, such as structural integrity of assets, fault detection and operational efficiency maximisation. Careful selection of established deep learning techniques Time from invention to application Need for end-to-end solution is exemplified, approaches to AI assurance to differentiate investment from adoption

Pharma Domain-specific customisations: graph convolutional networks, exploiting attention mechanisms, MC tree search Generative models, built Broad plethora of virtual screening upon recent developments in approaches built on deep neural autoencoders, adversarial networks with various molecular networks, recurrent nets representations Time from invention to application Innovation in customising AI inventions to improve existing machine learning models and explore new capabilities

Innovate or exploit Well-specified AI task Specified from a focused use case giving and opportunity for automation, augmentation, transformation Ability to do so requires domain experts Innovation Exploit AI inventions and discoveries Through specialisation and Ability to do so requires access to deep and broad AI understanding domain experts and use case

The bigger picture Well-specified AI task Specified from a focused use case giving and opportunity for automation, augmentation, transformation Ability to do so requires domain experts Innovation Exploit AI inventions and discoveries Through specialisation and Ability to do so requires access to deep and broad AI understanding domain experts and use case Software User experience, linking data, maintenance Data Compute