At Rise of AI 2017 in Berlin, six AI startups pitch their products back to back in a demo session, each followed by questions from a panel and the audience. Gestoos presents gesture-recognition software for digital signage and cars, built from small labeled datasets and refined with live deployment data, and reports a 1,500-license retail rollout. German Autolabs shows Chris, a voice-and-gesture assistant for cars that reads messages and adapts to driving context, funded through a recent Kickstarter campaign. Inspirient demonstrates software that automatically scans business spreadsheets and sensor data to generate presentation slides of anomalies, illustrated with the Volkswagen emissions dataset and a DB Schenker shipping-container example. Searchink pitches AI-based document process automation that classifies and extracts data from paper and digital document streams for corporates, citing access to roughly 900 million documents from a founder's earlier company. Twenty Billion Neurons describes a crowd-acting platform that pays contributors to film labeled action videos, used to build the largest known hand-gesture dataset and a webcam-based gesture demo. Aisoy Robotics closes with a live demo of Aisoy1 Kid, an emotional educational robot with an internal 'mentor' system, reporting over 3,000 users mainly in Spain, France and the US.
Key takeaways
- Gestoos trains gesture-recognition models from as few as five users and ten samples each, then refines them with live deployment data instead of large pre-collected datasets.
- German Autolabs argues touchscreens are unsafe in cars because looking at one takes three to five seconds of eyes off the road, and pitches voice-plus-gesture assistant Chris as the fix.
- Inspirient automatically generates presentation slides of data anomalies from raw spreadsheets, illustrated by flagging Volkswagen as an emissions-data outlier.
- Searchink says access to about 900 million previously annotated documents from a founder's earlier company gave it an early training-data advantage for document automation.
- Twenty Billion Neurons built a crowd-acting platform paying workers 10-12 cents per video to create the labeled action and gesture data that object-recognition datasets lack.
Detailed notes
Gestoos: gesture recognition from small datasets
The presenter, introduced as Gestoos' CEO, explains that standard deep learning needs tens of thousands of labeled examples, which is too slow and expensive for fast product iteration. Gestoos instead trains a base model from as few as five users and ten samples each, deploys it to camera sensors and screens, and improves it with live usage data once installed, reaching what he describes as 90% robustness. He cites a large digital-signage installation of 1,500 store licenses, projected to reach 16,000 by end of 2018, bookings of 520,000 euros the prior year rising toward 1.3 million, and Cisco as a solutions partner; pricing runs around 30 euros per screen per month for retail, with per-camera unit deals for larger automotive contracts.
German Autolabs: Chris, a voice assistant for cars
Patrick, co-founder and chief product officer, presents Chris, an in-car assistant addressing distracted driving; he cites that 70% of drivers use smartphones while driving and notes rising per-mile driving deaths in the US, calling it 'the Snapchat effect'. A demo video shows Chris reading messages, giving directions and rerouting around traffic. Patrick argues touchscreens are unsuitable in cars because looking at a screen takes three to five seconds of eyes off the road, and that existing voice assistants like Siri are not yet reliable enough, even in a 2017 car equipped with a factory infotainment voice system. He outlines three needs: combining voice with gesture and display, adapting interaction to driving context such as traffic density, and reducing cognitive load by having the system resolve multi-step questions itself, for example proposing lunch times based on calendar availability. The device had recently closed its Kickstarter campaign and was targeted to ship to backers by year end, reaching retail in Q1 the following year.
Inspirient: automated discovery of business data insights
The presenter frames Inspirient's mission as automating the kind of accidental discovery behind scientific breakthroughs, applied to business data. The system ingests spreadsheets or sensor data and automatically generates presentation slides highlighting notable patterns. As an example, he uploads the US EPA's vehicle emissions test dataset and shows a slide the AI generated identifying Volkswagen as a CO2-versus-NOx outlier, plus a recommended-insights view listing other outlier carmakers. He also shows the system applied unprompted to Fukushima nuclear plant sensor data and to DB Schenker's global shipping-container temperature logs, tracing routes between the Americas and Eurasia. Asked how clean input data needs to be, he says Inspirient runs analysis on imperfect data anyway and flags missing data in a footnote, leaving the user to judge whether the result is good enough.
Searchink: AI for document process automation
Co-founder Sophie explains that Searchink goes into large and small corporates and speeds up administrative document processes by up to 80%, citing that over 60% of order management, over 80% of claims management and over 70% of expense management data is still extracted manually. The platform classifies and clusters document streams, then extracts the most relevant information via an API. A first customer pool held 30 million documents, and the team's access to roughly 900 million previously annotated documents from a founder's earlier robotics company gave it an early training-data advantage. Searchink reported prior-year project-based revenue of roughly 600,000-700,000 euros and said onboarding a new customer vertical takes about five weeks; it does not process unstructured text such as emails, focusing instead on already-formalized document types.
Twenty Billion Neurons: crowdsourced video data for computer vision
A co-founder explains that standard object recognition, such as labeling cats and dogs, is largely solved, but machines still cannot interpret complex human actions because labeled data for that does not exist. Twenty Billion Neurons built what he calls the first crowd-acting platform, where the company issues an action label, such as an object falling or a hand gesture, and pays crowd workers 10-12 cents per video to film themselves performing it. The approach generated roughly 110,000 videos and, applied to hand gestures, produced what the presenter calls the largest crowdsourced hand-gesture dataset, demonstrated live via a standard laptop webcam recognizing swipe, zoom and other gestures without a depth sensor. The Toronto- and Berlin-based team numbered 18 people at the time.
Aisoy Robotics: an emotional robot for children's education
The Aisoy Robotics team demonstrates Aisoy1 Kid, describing it as the first emotional robot built for education, designed to build an affective bond with a child, make its own decisions, and adapt to the user. It includes a system called Mentor, which proposes content aimed at developing a child's multiple intelligences. In a live, voice-driven demo the robot introduces itself, plays music, reacts with simulated emotions such as surprise or shyness, and supports learning to code via Scratch. The company reports more than 3,000 users, mainly in Spain, France and the US, with UK expansion planned, and closes on the line that education is not possible without emotion.
Quotes
Touch in the car is not a good user interface, no matter how big and how fancy your touch panel is.
Patrick, German Autolabs
We go into big corporates but also small companies. We look at their administrative processes and we speed them up to 80 percent.
Sophie, Searchink
Questions answered in this talk
How does Gestoos build gesture recognition without large training datasets?
According to Gestoos' presenter at Rise of AI 2017, the company builds a base model from only five users and ten samples each, deploys it to camera sensors and screens, and improves accuracy using live data gathered after deployment rather than upfront mass data collection.
Why does German Autolabs think touchscreens are wrong for cars?
Patrick of German Autolabs explains at Rise of AI 2017 that using a touchscreen requires looking at it, finding and pressing a button and getting feedback, which takes three to five seconds of eyes off the road, so the company built the voice-and-gesture assistant Chris instead.
How does Inspirient generate insights automatically from business data?
At Rise of AI 2017, Inspirient's presenter describes uploading a dataset, such as US vehicle emissions test data, and having the AI automatically produce presentation slides identifying outliers and patterns, as demonstrated by a slide flagging Volkswagen as a CO2-versus-NOx outlier.