VentureBeat TRANSFORM:

Learning about AI trends & implementation strategies

The following report provides a summary of the VentureBeat TRANSFORM AI conference with key takeaways from fireside chats with big names in technology including AirBnB, LinkedIn, Google, Trustpilot, Johnson & Johnson, Gap, and Facebook. Finally, interesting startups are showcased at the end.

Executive Summary

VentureBeat TRANSFORM is a global conference that showcased the latest development in computer vision, business AI, implementing AI, intelligent RPA, and more in San Francisco. The event attracted over 900 attendees, 120 speakers, 48 sessions, and 50 startups comprised of innovators, investors, tech leaders, entrepreneurs, and corporate professionals,

The conference featured guest speakers from big names in technology and venture capital that shared personal anecdotes, perspective on AI trends, as well as advice on ethics and implementation of AI. Many speakers emphasized the importance of making AI accessible, preventing bias in AI, the need for an AI culture, and developing the right framework for implementing AI and using it at scale within organizations.

Companies big and small are recognizing and embracing the value of AI across all business functions. It is being applied in more organizations and at greater scale, but has also created new challenges, including but not limited to governance, culture, education, and deployment. It is clear that AI has finally crossed the chasm of adoption, at least for the early majority.

Evolution of Innovation Speed

Matt Marshall- the founder of VentureBeat gave the audience a welcome speech and a recap of the history of AI (the term was coined in 1956 and referred to as software that can learn by training where deep learning is a subset of that) and evolution of the speed of innovation across time.

Marshall spoke about how the adoption rate of companies using AI has increased three-fold in the past 12 months, much faster than before.

Companies using AI over time

2016 to 2018

• 1 in 25 companies

2019

• 1 in 3 companies

According to Marshall innovations cross the adoption chasm when 25% of the population is reached. This achievement took AI less time than previous technological innovations:

· The web: 7 years

· Smartphones: 10 years

· AI: 3 years

Why has the speed of innovation increased so dramatically? Because of advances in chip architecture, research, the explosion of data, cloud computing, and platform companies with opensource software as enablers.

Marshall held a survey under businesses to identify insights of how business think around AI. A lot of responses mentioned that there are still barriers to implementing AI to ensure success. Fifty percent stated there was not enough talent and resources to achieve this. Furthermore, companies recognized that AI is usually implemented per business line (42%) and takes time to implement before showing results; a majority expected an ROI between 6 and 12 months.

Finally, a large number of businesses stated that being pragmatic with AI, educating the teams on AI, and ensuring deployment are their focus areas to ensure success with AI.

AirBnB Fireside Chat

The CTO of AirBnB, Vanja Josifovski, talked with moderator Matt Marshall. One of the first topics that was addressed was whether companies should build and use their own AI tech stack or use an off-the-shelf platform from large tech companies. Josifovski said that building the whole AI tech system requires different specialists from various company departments, and as AI Is more commoditized it’s easier to monetize than before. Asked if AI will be dominated by a few big players or startups and whether the in-house technology would prevail over opensource, Josifovski had an interesting perspective:

“Big players took opensource technology and brought it in-house with their own customizations”. “The benefit of running the technology yourself enables you to own the data, but I would question whether dependency on owning the data yourself is important for future growth”.

According to Josifovski, a company like AirBnB can learn a lot from small amounts of data as 90% of what they do is combining and automating business processes. They are focused on building multiple learning systems where it’s currently difficult to produce tangible results.

As bias is a big topic with AI the moderator asked how AirBnB handles this. Josifovski answered saying that AI and ML are a reflection of society, the way it learns is to look at the world and capture the current state — which are results of societal development. Therefore, if bias is in society it would typically also be present in AI and ML. AirBnB tries to tackle this by participating in technical communities and including them in developing solutions. He said it’s possible to avoid key problems in bias by excluding populations of dimensions that are insignificant for the AI or ML model.

LinkedIn Fireside Chat

The VP of AI of LinkedIn, Deepak Agarwal, talked with moderator Souvik Ghosh, Principal Staff Engineer and Scientist (LinkedIn) on how LinkedIn uses AI across their organization.

Fun fact: LinkedIn has over 40,000 engineers

The moderator asked Agarwal what it really means to democratize AI. He stated that AI and ML will have a considerable impact on shaping society for the future. He referred to the early industrial revolution in the 18th century as example of democratizing where the invention of the steam engine increased productivity tremendously. This meant that early adopters that had capital were able to build businesses around those machines and had the expertise to design, build, and operate them.

The same goes for AI where initially only a handful of people can work with it, but you have to push for it to become a level plain field (what is happening more and more now). Right now, there are about 40m professional developers in the world working on democratizing AI for a larger audience.

On a positive note Agarwal mentioned a couple of things that are on the horizon:

· Transferred learning and machine teaching

· Reinforcement and supervised learning

· The constraint of data labelling will become less after which the computer power may still be a limiting factor

Finally, he mentioned it was key to engage in the public debate around AI and to show your unique perspective to create a balanced dialogue.

Google Cloud AI Fireside Chat

The Head of Cloud AI of Google, Andrew Moore, talked with moderator Jana Eggers, CEO of Nara Logics about several topics: AI as the key to human survival, what it means to create a successful AI project, how AI is organized at Google, and a rapid-fire discussion.

Is AI the key to human survival? Moore believes that “AI definitely isn’t our crisis, but our potential.” Although AI is still far from reaching human creativity, humans and machines working together with people are safer than not. For example, in climate change events such as a flood, building a robot that can grab rocks for emergency actions, or carry people out of the water is where we have technology on our side. Eggers noted, “It’s a mistake when we say that a machine is better than humans — we should say machines and humans together are better than anything else.”

In terms of how Google organizes their AI, Moore said it’s based on a sense of urgency — the number of problems in the world and the number of people that are dealing with that problem as much as they deal with technology. According to Moore, you should only start once you have a clear and measurable view of the goal and benefits, which will make it easier to form a team. Secondly, team motivation is also very important and a project or goal should make the team excited as that drives the goal, the team, and the subsequent benefits.

The top two high level things Moore says to people around AI ethics:

• Read a book on how to have ethical debates in organizations regarding AI
• Trends in AI help to follow ethical considerations

Trustpilot Fireside Chat

Ramin Vatanparast (The Chief Product Officer of Trustpilot), talked with moderator Seth Colaner (AI Editor at VentureBeat) about using AI to safeguard consumer content and feedback.

Vatanparast shared that Trustpilot works between consumers and companies to collect reviews from both sides to develop insights for businesses to improve and enable consumers to make better decisions — their mission is to be the most trusted review platform on the market. Over the last several years, the focus of Vatanparast has been on making the business AI-based/AI-centric and applying the right processes, resources and tools to get to the next level.

Today, internet users are consuming a lot more information than they did a few years ago, and can find it difficult to trust what they are reading (e.g. the proliferation of fake news). According to a 2007 study surveying trust in organizations and the government, consumer trust dropped by a staggering 43%.

Vatanparast shared that the most difficult problem the company is facing is how to create and establish trust, but are working towards a positive outcome:

· 86% of consumers in the UK and US trust content from Trustpilot

· 92% of online consumers read reviews

· 85% researches companies online and their reviews before purchase

· 89% says a review influences their decision

· 75% share their experience after making a purchase

Early on, Trustpilot developed a data warehouse to establish their AI and solidify what they wanted to achieve and why. They credited starting out small, but thinking big, and have built on 3 layers within their organization:

1. Building an AI culture & becoming data-centric
2. Defining, analyzing, and processing data of the customer journey
3. Applying AI to product solutions

Vatanparast reflected on the organizational challenges of implementing AI within the organization, and how they handled organizing all that unstructured data. In the beginning of their AI journey, focus was very important (especially on the outcomes), demystifying AI, and agreeing that it was okay to fail in the beginning and learn valuable lessons. Ultimately, he stressed the importance of each company connecting AI to their strategy early on. Furthermore, looking at the data quality (what and why) and processing in line with the desired outcome, as well as the technical side was key to the development and implementation of AI within an organization. According to Vatanparast, “It’s not just about building models to give output, but what action you want to take with the results from the AI models.”

Johnson & Johnson Fireside Chat

Marc Leibowitz (Global Head of Digital of J&J), and Shelia Anderson (CIO of Corporate Functions Technology from Liberty Mutual), talked with moderator Jaime Fitzgibbon (Innovation Strategist & Director of The San Francisco Innovation Hub from Booz Allen Hamilton) about the path to digital transformation for large corporates.

The discussion kicked-off with both panelists sharing insights into their respective organizations’ digital transformation progress. Lebowitz mentioned that the AI team reports directly to the CIO and operates in 200 countries in a decentralized manner. He emphasized that the culture is not very data-driven in healthcare and that driving cultural change has been a big part of the past 3 years for him. Anderson added to that by noting that

“digital transformation at Liberty Mutual was like building a plan while flying.”

She joined the company to lead the company’s digital transformation and her focus was all about testing, learning, and having the momentum to keep going.

When asked whether AI was moving more towards a forward-facing role or back-office role, Leibowitz shared that J&J made a lot of thoughtful plans, but the reality was that most of them have gone nowhere. He said,

“Once you get into the businesses and incumbents and teams it becomes messy, challenging, and uncomfortable for people to solve problems differently. It became more ad-hoc.”

They started trying to reinvent J&J and soon noticed it was too much change, too fast for people. Leibowitz made the scope more manageable to inflect the growth curve and keep the scope achievable by focusing on areas like automation and predictive analytics. Regulation is often used as excuse for fight against change/disruption as Leibowitz wanted to go beyond optimizing current operations.

When asked about how nurturing or acquiring talent fits into the digital transformation, Anderson noted that they have “champions” and use a multi-pronged approach. Even though they are in Boston with a large amount of talent, their challenge is that there are fancy startups with better branding that attracts talent away from Liberty Mutual. This is why they have programs for training to educate internally, encourage movement of teams, and attract new hires through accelerator affiliations like Techstars. In addition, Liberty Mutual is constantly assessing compensation (which gets people in the door), but the stickiness of jobs is generally key for talent to stay. Leibowitz added to that by sharing he spends a lot of time in acquiring talent and re-training existing talent — the mindset of employees is different now, as many have a 2–3 year outlet and then re-evaluate.

What worked well for each company during their digital transformation:

• Marc Leibowitz, J&J: The combination of a subject expert with a business expert to solve problems in a unique way
• Sheila Anderson, Liberty Mutual: Partnering with the right people. For example, Liberty Mutual successfully partnered with renowned university (MIT) to help solve challenges creatively

Gap, Inc Fireside Chat

Chris Chapo (The SVP of Data & Analytics from Gap), talked with Deborah Leff (Global Leader and Industry CTO for Data Science and AI from IBM) about what it means to “Do AI” and use it successfully.

Chapo shared the following definitions when asked to define AI:

• Business definition: creating computers or systems that mimic humans, can solve problems like people can but faster, cheaper, and more effective than humans can. For example, as a human in healthcare to stay current in their medical field, it requires reading material of 40 hours a week. This is an area where computers can support humans to see things we can’t see and take less time to learn new content and analyze data
• Data science definition: type of tools and breathe of algorithms, ML, NLP, and using predictive analytics to solve problems

Small companies can chip away market share from larger companies by using their speed in their advantage. According to Leff, the company that picks up and runs with AI faster than the competitor can create a competitive advantage, regardless of size.

“Data science has been around since the 1950’s but it means something different to different people”

• If you are talking to VCs, you are talking AI
• If you are talking internally about a project, you are talking ML
• If you are talking with a data scientist about a project, you are talking regression analysis

The interpretation is an important point for AI projects. Chapo prefers not to talk about AI, but focus on the problem first. He highlighted a use case of Gap where he discussed the lifecycle from a clothing collection from ideation to store delivery. Normally, this process takes about 56 weeks, but Gap used AI to shrink the length considerably.

Leff said that most successful projects she has seen start at the end — the desired outcome. Chapo agreed and added that they start with the solution and then add layers of complexity. He does this because small AI projects show value quicker and it helps teams become more comfortable with using AI for bigger projects later.

The panel had similar experiences around how when the first time they move an AI project to production,1 often fail. According to IBM Only 13% of AI projects reaches the production stage. Chapo shared, “Creating conditions for success is important as well as having the right stakeholders at the table from the beginning.”

The moderator brought up the hype cycle AI has seen recently, spurred by marketing. She asked why AI is not more mainstream knowing the exposure it has with companies. Chapo mentioned there are many reasons, but he highlighted that some people think that is you throw enough money at a problem with technology that success is guaranteed and that it just doesn’t work like that. He said the right leadership support and focus are required.

Leff brought up an article by the Harvard Business Review which said that AI will disintermediate jobs, but it will not replace managers. Rather, managers who use AI will replace managers that don’t use it.

In conclusion, Chapo shared his advice for people looking to start an AI project:

1. Pick a small project to start that lasts two months or less
2. Have the right team with collaborative and cross functional people
3. Leverage 3rd parties to accelerate, learn from others, and use best-practices

Facebook AI Fireside Chat

Jerome Pesenti (VP of AI at Facebook) talked with Founder & CEO from Venturebeat, moderator Matt Marshall, about their use of AI and the progress they’ve made.

Facebook uses ML and AI across all of their business units, as well as for specific hardware and software solutions. These include bots and assistants, generated content, AR effects, and VR hardware. The main goal for the company is to protect, connect, and empower people.

Pesenti shared that Facebook uses AI and ML to translate all the communication from the social media network to different languages, increase accessibility, and ensure newsfeed integrity. With so much content to review, the company also needs a lot of human moderators. Last year alone, Facebook hired 30,000 people to moderate content, but the nature and volume of the content quickly gave rise to a need for AI assistance to lighten the load.

In addition, Facebook leverages the social graph with ML to plot recommendations from friends to make it more useful. In terms of the roadmap, they use AI across all their offerings and to define next level experiences.

Pesenti shared a few thoughts on challenges he perceived with AI at Facebook:

1. Culture with the need to be data driven and use that to make decisions across the full product lifecycle
2. The volume of data — Facebook processed 3.5B images on Image Net with 85% accuracy
3. Computing at scale — the past year alone the company was able to compute 30X more in data and train 10X more workflows
4. Data cleaning where labelling and supervised data is weak

The moderator introduced the topic of bias in AI and ethics when he asked Jerome if Facebook had a process to tackle and prevent bias in data. Jerome referred to a process called Surface, Resolve, and Record.

1. Surface, ask the hard questions about fairness
2. Resolve, create a process to resolve the questions
3. Record, create a record of the process and decisions taken to resolve the questions

Facebook uses this process in their business to search for the ground truth in data, label it, make predictions, and intervene as needed. This helps them label three different versions of bias:

1. Labeling bias
2. Algorithmic bias
3. Intervention bias

Ultimately, Pesenti said that all systems in western societies have bias and that the cost of using AI has kept him up at night. When the moderator asked him what the right ratio would be between cost and benefit of AI, he believes being more cost conscious and focusing on optimization is the solution.

Interesting Startups Showcase

Doc.AI is a health company that enables users to get a full picture of their health with predictive insights, share their health data (lab tests, medical records, environmental data) for data trials securely over blockchain, and earn money for contributing to science.

Brainworks is a medical AI company that develops AI to read vital signs, such as blood pressure, heart rate and respiration by using peoples’ smartphone or laptop camera.

Mesmer is an AI company in Robotics Process Automation for Development (RPAD) that uses deep learning and AI-bots to speed up customer experience testing for developers so they can spend more time developing versus testing code.

D-ID develops de-identification solutions for facial images to protect privacy of users while preserving functionality. D-ID removes unnecessary information from biometric data in facial images and ensured decryption or reverse-engineering is impossible.