I’m looking for a specific photo of a friend I took on my iPhone a few years ago. There are thousands of images to search, but Apple’s Photos app focuses on the right people. Within seconds you will find the image you are looking for.
There is a lot of work going on behind the scenes to make this happen, including facial recognition, image analysis, and automatic tagging. We save effort by guessing what we want and need and acting on those guesses in real time.
Companies like Apple, Google, FedEx, Uber, and Netflix have spent years building systems and architectures to make the user experience easier, more personal, and more intuitive. In some cases, artificial intelligence can be used to make critical decisions almost instantaneously or make predictions in real time, enabling businesses to improve results in the moment.
According to a 2022 Deloitte survey, 94% of business leaders say AI is critical to their success.
So why is building successful AI applications such a big challenge for most organizations? It boils down to three big hurdles. Putting the wrong data on the wrong infrastructure at the wrong time.
Hurdles to AI success
According to McKinsey, 56% of companies have adopted AI, but only 12% have succeeded in achieving superior growth and business transformation with AI, as Accenture notes in its report.
There are many obstacles to successfully embedding AI into real-time applications, most of which relate to one central element: data.
Many traditional ML/AI systems and the results they produce rely on data warehouses and batch processing. The result: “Ingesting” this historical data into a machine learning system requires a complex set of technologies, data movements, and transformations.
The data that enters an ML model is called features (measurable properties that can be used for analysis) and are typically based on data stored in application databases or written to log files. Transformations such as scaling values or calculations based on previous records (for example, moving averages when records are generated) are often required.
This generally slows the flow of data from input to decision to output, which can lead to missed opportunities and customer churn. It can also leave perceived cybersecurity threat patterns undetected and unmitigated. The challenge can be summed up as having an inadequate dataset and being supported by an inadequate infrastructure that moves too slowly.
wrong data
Due to the sheer volume of data (and associated costs), data must be aggregated for ease of transfer and availability. Simply put, aggregated or overly transformed data prevents organizations from easily identifying appropriate actions in real time, reducing the likelihood of achieving desired outcomes. This reduces an organization’s ability to find answers to new questions, predict outcomes, and adapt to rapidly evolving conditions.
Data scientists are forced to use coarse-grained datasets that drive fuzzy predictions, especially in discrete contexts like customer sessions, that don’t lead to expected business impact. There can also be important events that are not recognized and do not feed functionality when the application is reconfigured or the data source evolves. This missing data leads to uninformed decisions regarding model selection. This can lead to poor prediction accuracy, or worse, models using the wrong data can lead to wrong decisions.
Finally, aggregation focuses on creating existing functionality. New feature engineering (processing data required for model selection and training) requires going back to raw data for various aggregations. This extra processing significantly slows down the data scientist’s work and lengthens the experimentation process.
… with inadequate infrastructure
A second challenge is related to the current ML infrastructure that powers AI initiatives and the inability to process datasets at scale. The quality of the model and its results improve with the amount of event data ingested. Organizations often have to handle an enormous volume of events that their legacy infrastructure just can’t keep up with.
The set of trained models that feed them to perform inference becomes complex. Especially since you need to move data between each one. Attempting to handle the scale required for high-quality predictions pushes traditional architectures to their breaking point. It’s also painstakingly slow, unreliable, and costly. All of this threatens the value and impact of increasingly mission-critical apps.
… at the wrong time
Another obstacle is that data is processed too slowly to have a significant impact. Current architectures require data processing by multiple systems to deliver the model. This introduces latency that impacts AI initiatives in a number of ways.
- A model’s output cannot change the course of an evolving situation. For example, you may have made an offer to a customer when the conversion rate dropped and the customer bought something else.
- The time it takes to deliver a model to get results does not match the expectations of digital experiences and automated processes. It may take several days before the data is ready to process. In a competitive market, data this old is irrelevant at best and dangerous at worst (think ride-sharing apps that apply surge tolls during crises and disasters).
- Data scientists don’t have access to the latest data. This can affect model results and may require data scientists to spend valuable time looking for additional data points and sources.
Much of the current ML infrastructure is too costly, too complex, and too slow to deliver your applications. Also, regulatory changes may ultimately require organizations to provide more detailed explanations of how models were trained to reach certain decisions. This level of visibility is not possible with current architectures due to the processing, aggregation, and various tools involved.
Many infrastructure problems lie on the way data reaches AI-driven applications. The answer to this problem, simply put, is to do the opposite.
Bring AI to your data
Leaders like the companies mentioned at the beginning of this article thrive on collecting large amounts of real-time data from their customers, devices, sensors, or partners as they pass through their applications. This data is used for model training and serving. These companies act on this data in the moment, serving millions of customers in real time.
Another key factor in Reader’s success is the fact that it collects all data at the most granular level as time-stamped events. This means that you don’t just have a lot of data. It also allows us to understand what happened when over time.
Leading companies such as Netflix, FedEx, and Uber are “bringing AI where the data is,” so that they can provide inference where applications live. This means embedding ML models into your applications, aggregating events in real-time through a streaming service, and publishing this data to your ML models. And they have a database (in the case of his three leaders above, the high-throughput open-source NoSQL database Apache Cassandra) that can store huge amounts of event data.
With the right integrated data platform, your ML initiative will have the right infrastructure and the right data. Data engineers and data scientists can “get out of the silos” and align the process of feature engineering, model experimentation, training, and inference with force predictions. These processes still require many tools, but they all work from the same data foundation.
Leveraging massive amounts of event data to deliver models and applications, the most successful AI-powered applications lead the industry in differentiating themselves by continually improving the experience they deliver to their end users. increase. Our ability to serve millions of customers and become smarter allows our customers to define their markets.
Learn how DataStax enables real-time AI.
