According to a recent Deloitte study, 82% of companies that have already invested in AI have gained a financial return on their investment. For companies among all industries, the median return on investment from cognitive technologies is 17%.
AI is transforming daily life and business operations in a way seen during previous industrial revolutions. Current products are being enhanced (according to 44% of respondents), internal (42%) and external (31%) operations are being optimized and better business decisions are being made (35%).
With that in mind, it is better to see the “Trend” as a larger and more significant development than a particular technology or advancement. That’s why chatbots or autonomous cars are not so much seen as particular trends, but rather as separate threads in the fabric that is AI.
That distinction aside, here are five of the most significant and inspiring artificial intelligence trends to watch in 2019.
1. Chatbots and virtual assistants ride the lightning
The ability to process natural language is widely considered a hallmark of intelligence. In 1950, Alan Turing proposed his famous test to determine if a particular computer is intelligent by asking the ordinary user to determine if his conversational partner is a human or a machine.
The famous test was initially passed in 1966 by ELIZA software, though it had nothing to do with natural language processing (NLP) – it was just a smart script that seemed to understand text. Today’s NLP and speech recognition solutions are polished enough not only to simulate understanding but also to produce usable information and deliver business value.
While still far from perfect, NLP has gained a reputation among businesses embracing chatbots. PwC states that customers prefer to talk with companies face-to-face but chatbots are their second preferred channel, slightly outperforming email. With their 24/7 availability, chatbots are perfect for emergency response (46% of responses in the PwC case study), forwarding conversations to the proper employee (40%) and placing simple orders (33%). Juniper Research predicts that chatbots will save companies up to $8bln annually by 2022.
NLP is also used in another hot tech trend–virtual assistants. According to Deloitte, 64% of smartphone owners say they use their virtual assistant (Apple Siri, Google’s Assistant) compared to 53% in 2017.
Finally, Gartner has found that up to 25% of companies will have integrated a virtual customer assistant or a chatbot into their customer service by 2020. That’s up from less than 2% in 2017.
2. Reducing the time needed for training
Academic work on AI often focuses on reducing the time and computing power required to train a model effectively, with the goal of making the technology more affordable and usable in daily work. The technology of artificial neural networks has been around for a while (theoretical models were designed in 1943), but it works only when there are enough cores to compute machine learning models. One way to ensure such cores are present is to design more powerful hardware, though this comes with limitations. Another approach is to design new models and improve existing ones to be less computing hungry.
AlphaGo, the neural network that vanquished human GO champion Lee Sidol, required 176 GPUs to be trained. AlphaZero, the next iteration of the neural network GO phenom, gained skills that had it outperforming AlphaGo in just three days using 4 TPUs.
Expert augmented learning is one of most interesting ways to reduce the effort required to build reinforcement-based models or at least ones that are reinforcement learning-enhanced. Contrary to policy-blending, expert augmented learning allows data scientists to channel their knowledge not only from another neural network but also from a human expert or another machine. Researchers at deepsense.ai have recently published a paper on using transfer learning to break Montezuma’s Revenge, a game that reinforcement learning agents had long struggled to break.
Another way to reduce the time needed to train a model is to optimize the hardware infrastructure required. Google Cloud Platform has offered a cloud-based tailored environment for building machine learning models without the need for investing in on-prem infrastructure. Graphics card manufacturer Nvidia is also pushing the boundaries, as GPUs tend to be far more effective in machine learning than CPUs.
Yet another route is to scale and redesign the architecture of neural networks to use existing resources in the most effective way possible. With its recently developed GPipe infrastructure, Google has been able to significantly boost the performance of Generative Adversarial Networks on an existing infrastructure. By using GPipe, researchers were able to improve the performance of ImageNet Top-1 Accuracy (84.3% vs 83.5%) and Top-5 Accuracy (97.0% vs 96.5%), making the solution the new state-of-the-art.
Read the source article at deepsense.ai.