AI is no longer some abstract dream for the future. It is here, now and bringing change across industries. According to the Forrester AI Readiness Study, 40 percent of the 717 businesses surveyed said they were planning to use intelligent recommendation solutions and 43 percent were planning to use AI-enhanced advanced analytics.
With breakthroughs coming thick and fast in machine learning, especially deep learning models, the AI advantage is becoming far more compelling and is spanning across a wider set of applications. However, there are still several roadblocks to wide-scale implementation of AI which are important to recognize as the appetite within the industry to integrate the technology continues to grow.
Issues with industrializing AI
Today, AI enablement is difficult and very costly. Why? First, there is no turnkey, one-size-fits-all solution for industries looking to implement AI into their systems. Companies need to custom build, train, and fine-tune their AI solutions. Building custom AI systems requires several PhDs and experts applying a mixture of experience, engineering, philosophy of mind, and maybe a little bit of voodoo to come up with elaborate configurations of deep networks. Add this to the need for custom-built processing capacity and it seems that building custom AI is something only large, well-funded companies can afford.
Second, because custom machine learning models using deep learning require such an intense training process, the cost of retraining the models becomes prohibitive. Hence, while the custom models are robust, they have limited reusability and have the tendency to become stale once the system updates. These models are “static,” in that they are pre-programmed to function in one certain way and will eventually become obsolete as the environment around them changes.
Third, companies often create models using very large datasets that need to be collected, curated, and labeled. This is not only costly in time and money resources, but also very limiting, since AI based on past data has limited applicability in an ever-changing world.
So what are we missing? How can we industrialize this promising technology to help it meet its cross-industry disruptive potential? This is a difficult problem, and scientists are hard at work tackling it.
The solution: Creating adaptive models
The biggest missing component in wide-scale AI implementation is adaptation. An AI system built on robust machine learning models should be able to adapt to changing domains, since the world is always in a state of change. There needs to be a heavier reliance on evolutionary algorithms — models that shift and change to the context where they are needed. This will help the AI system adapt to the times and be able to capitalize on each different context without having to spend time and money on retraining the models.
Evolutionary algorithms are especially important in cases where a company uses the AI to augment or drive an online user experience. In online retail, for instance, companies are employing AI-backed solutions to personalize product recommendations in real time as the user clicks through the catalog and selects different items. Without needing any historical data on the user (such as what he or she bought in the past), the AI solution can shift and change the product recommendations based on the user’s present interactions with the site. Judging present actions on the site is much more effective than using historical data to predict what the user will want next.
Why not use AI to build and design AI?
If machine-built models are more efficient and robust than human-built models, why not train AI to build AI? Many scientists are now considering using AI to come up with parameters and configurations, and basically design their own AI systems.
Researchers have used population-based and evolutionary techniques, inspired by biological evolution, to come up with designs that exceed AI designed by humans. For instance, in the world of website optimization, AI-backed solutions can “evolve” website designs and, in a process similar to natural selection, combine the best performing elements and features to produce the most optimal website for conversions.
Evolution can be paired with neural networks to substitute the slow and intensive process of back-propagation training with an online adaptation of the network weights. This means researchers train the neural network in real time, as opposed to offline over historical data. This approach is surprisingly powerful for many online problems, such as autosegmentation and hyperpersonalization of online digital media. Reinforcement learning systems are also making headway in terms of adaptation. Reinforcement learning is a form of online learning that allows the system to weight its decisions based on the most immediate context.
The future is now
Good models alone are not sufficient for the wide-scale adoption of AI. Effective AI-enablement requires breakthroughs in machine adaptation and creativity, both of which we can achieve using evolutionary algorithms, or models that can adapt to each appropriate context and train other algorithmic models to process similar information as needed. This is the future — a world where AI can build AI. Replace any human-created model with machine-generated models and your system becomes much more robust and efficient. Hence, evolutionary algorithms are an example of the “living and breathing AI” — intelligent algorithms that can adapt with the times and think outside the box.
Welcome to the future.