Appen Ltd. · ASX:APX
Appen — Investor Day (May 2020)
31 slides from Appen's investor day presentation, May 19, 2020
Slide 1: Investor Technology Day
seeeeee eee eeeeee seeeeee eee eceees seeeeee se eees see ecceee Pe eeereey ee eeeees eee reeces seeveeces PPePErErey ereeee ny se eeeeees eeee eeeee sO eeeseeees seeesees seeee ee ee eee eee ee ere re re see eee sees eee eeeeeee eee eceeeee eT eee eS ee TEST ere eer errr eres PETSVTSESES ELSES ELEC
Slide 2: For personal use only
For personal use only • • • • • • •
Slide 3: Presenters
For personal use only Mark Brayan Wilson Pang Meeta Dash Ryan Kolln CEO CTO VP Product VP Corporate Development
Slide 4: AI applications: speech, NLU, vision, search, recommendations
For personal use only Speech-to-text AI: converts the raw audio file into text Natural Language Understanding (NLU): turns words into an actionable instructions Computer Vision: Label image attributes to make them searchable beyond product name Search Relevance: Return the optimal response for the r
Slide 5: AI model building and refresh cycle
For personal use only Model optimisation and refresh Collect Label data Train AI model Prediction Evaluate data for for training model output training Model building Real world data
Slide 6: Relevance, language, image
For personal use only • Relevance • Language • Image
Slide 7: Search relevance example
For personal use only “AZ” , “Phoenix” ,
Slide 8: Search relevance example
For personal use only “AZ” , “Phoenix” ,
Slide 9: Image annotation JSON example
Corresponding JSON file iudgable" dat tarted_a nothing to.box": { peat confidence’ ‘annotation*: [] '2020:05-14T06:29:04+00: 00 '2020-05-14T06:48: 26.00: 00 judgabte Gyelist 681, \"h 5: //drive .google.com/uc?id DST eGqzhQbDG0s wKeSED Location of image file Time and date Location Worker identifier C
Slide 10: AI performance correlates with volume of training data
Al performance is correlated with the volume of data used for training Performance Data Deep Neural Networks Medium Neural Networks Shallow Neural Networks Traditional Machine Learning appen
Slide 11: Data volume and deep learning performance
For personal use only Source: Sun, C. et al., Revisiting Unreasonable Effectiveness of Data in Deep Learning Era, https://arxiv.org/abs/1707.02968, Aug. 2017.
Slide 12: Volume of data is critical, but it also needs to be current
Volume of data is critical, but it also needs to be current “Model drift” is a common in Al models Refreshing training data ensures optimal model performance 8 8 a ——. 5 5 mS ~ E E 2 2 boy boy a a . 3B s ‘ nel nel fo} fo} \ = = ‘ Time Time * ~34% of models need to be refreshed monthly! * Large porti
Slide 13: Even the best AI models are far from perfect
Even the best Al models are far from perfect Textu Fox Squirrel Grasshopper Overgeneralization Source: https://arxiv.org/pdf/1907.07174.pdf
Slide 14: Vendors
For personal use only Vendor 1 Vendor 2 Vendor 3 Vendor 4 Vendor 5
Slide 15: Our mission
_Our mission is tomake =. -Alworkinthe realworld 2 by creating large a eS . . _ volumes of high-quality s training data faster. @o @6-% rn FOr p << appen
Slide 16: Technology combined with a crowd of over 1m
We combine our industry leading technology with a crowd of over 1m and deep internal expertise 55). 238 fs oS! yw appen
Slide 17: The Appen Platform
For personal use only The Appen Platform Crowd management Client Workspace Annotation Tools oe oe. eee
Slide 18: Crowd Management
Crowd Management provides a frictionless experience for crowd workers and project managers @- © &@ &® we Bs Crowd Registration Training Project Ticketing Payments Social Workers Qualification Campaign Report Ticketing Quality = Recruiting Project Project Mgmt Assurance Setup Managers appen
Slide 19: Client Workspace
Client Workspace helps data scientists manage training data preparation efficiently nape 08 2 & oo loan : “ny Ox) fa il Data Workflows Interface People Quality at Monitor & Raw Data Ingestion Design Scale Report Annotated Data {ely Enterprise Grade Secure Platform ES (SOC2, GDPR, |ISO27001, HIPPA Co
Slide 20: Annotation tools support a wide variety of use cases
Annotation tools support a wide variety of data labelling uSe CASES % & Ag altaya H Pa H i) By Pe LA) be ; Content Speech & Natural Language Computer Data Relevance Language Understanding Vision Collection Search Voice Name Entity Image LiDAR Mobile Relevance Transcription Extraction Transcription A
Slide 21: Platform building blocks
For personal use only 21 a Ww Crowd Client Annotation management Workspace Tools Our platform provides the building blocks to provide large volumes of high-quality training data faster eee. oe oe. eee
Slide 22: Utilise AI to automate and increase crowd productivity
For personal use only Utilise AI to automate and increase crowd productivity
Slide 23: Increasing product breadth
Increasing our product breadth will enable us to capture greater market share 1. Product breadth Examples of how we are increasing product breadth “* Support a wider variety of data annotation types, including vertical specific capabilities ** Improve our low-touch service models to increase our add
Slide 24: Crowd productivity
Utilise Al and automation to improve the throughput of our crowd 2. Crowd productivity Examples of how we are increasing crowd throughput ** Incorporate Al into our data labelling process to radically increase the output quality and volume of our crowd appen
Slide 25: Internal efficiency
Apply Al to our internal operations to deliver large-scale high- quality data 3. Internal efficiency Examples of how we are increasing internal efficiency ** Leveraging Al to improve the crowd recruitment process “* Improve data quality by automatically matching crowd workers with tasks better suite
Slide 26: Technology for customers, crowd workers and employees
Technology plays a critical role for our customers, crowd workers, and employees @ (ny Crowd workforce ©) Employees appen
Slide 27: Benefits to the business: product breadth
For personal use only Benefits to the business ❖ Customisable platform to support annotation tasks across a wide variety of data types Broader product set enables us to capture a wider share of the market ❖ Labelling automation improves speed and quality data annotation Greater integration with our
Slide 28: Benefits to the business: crowd productivity
For personal use only Benefits to the business ❖ Predictive matching of crowd workers to tasks improves worker experience and output Accelerated onboarding improves our ability to fulfill project demands ❖ Automated resume screening accelerates onboarding More tenured crowd workers ramp faster and r
Slide 29: Benefits to the business: internal efficiency
For personal use only Benefits to the business ❖ Automated quality management through AI validation Internal workforce productivity lowers project overhead costs ❖ Faster and more flexible project setup (for managed service customers using our platform) Improved quality delivers better customer outc
Slide 30: Technology is essential to delivering high-quality training data faster
Technology is essential to the delivery of large volumes of high-quality training data faster Build the platform Accelerate scalability Revenue growth and quality Crowd 1. Product breadth * Increased market share by management supporting more use cases and customers * Diversified customer base and .
Slide 31: For personal use only
For personal use only