We see the AI space poised for an acceleration in adoption, driven by more sophisticated AI models being put in production, specialized hardware that increases AI’s capacity to provide quicker results based on larger datasets, simplified tools that democratize access to the entire AI stack, small tools that enables AI on nearly any device, and cloud access to AI tools that allow access to AI resources from anywhere.
Integrating data from many sources, complex business and logic challenges, and competitive incentives to make data more useful all combine to elevate AI and automation technologies from optional to required. And AI processes have unique capabilities that can address an increasingly diverse array of automation tasks, tasks that defy what traditional procedural logic and programming can handle—for example: image recognition, summarization, labeling, complex monitoring, and response.
In fact, in our 2019 surveys, more than half of the respondents said AI (deep learning, specifically) will be part of their future projects and products—and a majority of companies are starting to adopt machine learning.
The line between data and AI is blurring
Access to the amount of data necessary for AI, proven use cases for both consumer and enterprise AI, and more-accessible tools for building applications have grown dramatically, spurring new AI projects and pilots.
To stay competitive, data scientists need to at least dabble in machine and deep learning. At the same time, current AI systems rely on data-hungry models, so AI experts will require high-quality data and a secure and efficient data pipeline. As these disciplines merge, data professionals will need a basic understanding of AI, and AI experts will need a foundation in solid data practices—and, likely, a more formal commitment to data governance.
That’s why we decided to merge the 2020 O’Reilly AI and Strata Data Conferences in San Jose, London, and New York.
New (and simpler) tools, infrastructures, and hardware are being developed
We’re in a highly empirical era for machine learning. Tools for machine learning development need to account for the growing importance of data, experimentation, model search, model deployment, and monitoring. At the same time, managing the various stages of AI development is getting easier with the growing ecosystem of open source frameworks and libraries, cloud platforms, proprietary software tools, and SaaS.
New models and methods are emerging
While deep learning continues to drive a lot of interesting research, most end-to-end solutions are hybrid systems. In 2020, we‘ll hear more about the essential role of other components and methods—including Bayesian and other model-based methods, tree search, evolution, knowledge graphs, simulation platforms, and others. We also expect to see new use cases for reinforcement learning emerge. And we just might begin to see exciting developments in machine learning methods that aren’t based on neural networks.
New developments enable new applications
Developments in computer vision and speech/voice (“eyes and ears”) technology help drive the creation of new products and services that can make personalized, custom-sized clothing, drive autonomous harvesting robots, or provide the logic for proficient chatbots. Work on robotics (“arms and legs”) and autonomous vehicles is compelling and closer to market.
There’s also a new wave of startups targeting “traditional data” with new AI and automation technologies. This includes text (new natural language processing (NLP) and natural language understanding (NLU) solutions, chatbots, etc.), time series and temporal data, transactional data, and logs.
And traditional enterprise software vendors and startups are rushing to build AI applications that target specific industries or domains. This is in line with findings in a recent McKinsey survey: enterprises are using AI in areas where they’ve already invested in basic analytics.
Handling fairness—working from the premise that all data has built-in biases
Taking a cue from the software quality assurance world, those working on AI models need to assume their data has built-in or systemic bias and other issues related to fairness—like the assumption that bugs exist in software, and that formal processes are needed to detect, correct, and address those issues.
Detecting bias and ensuring fairness doesn’t come easy and is most effective when subject to review and validation from a diverse set of perspectives. That means building in intentional diversity to the processes used to detect unfairness and bias—cognitive diversity, socioeconomic diversity, cultural diversity, physical diversity—to help improve the process and mitigate the risk of missing something critical.
Machine deception continues to be a serious challenge
Deepfakes have tells that automated detection systems can look for: unnatural blinking patterns, inconsistent lighting, facial distortion, inconsistencies between mouth movements and speech, and the lack of small but distinct individual facial movements (how Donald Trump purses his lips before answering a question, for example).
But deepfakes are getting better. As 2020 is a US election year, automated detection methods will have to be developed as fast as new forms of machine deception are launched. But automated detection may not be enough. Detection models themselves can be used to stay ahead of the detectors. Within a couple months of the release of an algorithm that spots unnatural blinking patterns for example, the next generation of deepfake generators had incorporated blinking into their systems.
Programs that can automatically watermark and identify images when taken or altered or using blockchain technology to verify content from trusted sources could be a partial fix, but as deepfakes improve, trust in digital content diminishes. Regulation may be enacted, but the path to effective regulation that doesn’t interfere with innovation is far from clear.
To fully take advantage of AI technologies, you’ll need to retrain your entire organization
As AI tools become easier to use, AI use cases proliferate and AI projects are deployed, and cross-functional teams are being pulled into AI projects. Data literacy will be required from employees outside traditional data teams—in fact, Gartner expects that 80% of organizations will start to roll out internal data literacy initiatives to upskill their workforce by 2020.
But training is an ongoing endeavor, and to succeed in implementing AI and ML, companies will need to take a more holistic approach toward retraining their entire workforces. This may be the most difficult, but most rewarding, process for many organizations to undertake. The opportunity for teams to plug into a broader community on a regular basis to see a wide cross-section of successful AI implementations and solutions is also critical.
Retraining also means rethinking diversity. Reinforcing and expanding on how important diversity is to detecting fairness and bias issues, diversity becomes even more critical for organizations looking to successfully implement truly useful AI models and related technologies. As we expect most AI projects to augment human tasks, incorporating the human element in a broad, inclusive manner becomes a key factor for widespread acceptance and success.