In 2018, new hardware developments allowed more sophisticated and widespread use of AI, and software managed to make AI even more human-like. Take a look back at the developments in AI over the past year, and a look ahead to what may come next.
Defining and Debating What "Artificial Intelligence" Means
The concept of artificial intelligence has been around for centuries, such as the story of the automaton Talos from ancient Greek mythology. But in 2018, as AI seeped further into everyday life, the rapid developments in AI hardware and applications sparked debate as to its very existence. Depending on how we define both "artificial" and "intelligence", we can determine whether we see AI as a marketing myth, a legitimate stride in technology, or a harbinger of a terrifying future.
Talos, from Greek mythology, was an automoton with artificial intelligence. Painting by Thomas Bulfinch.
In one basic sense, AI can be seen as highly sophisticated software with the capacity to read data and learn. AI can process massive amounts of data at incredible speed, feats impossible for human intelligence. But AI can't yet take human factors (such as emotions) and certain contextual situation into account as it learns.
Debate over the ethics and regulation of AI heated up in 2018, as well. In May, there were resignations at Google over the controversial use of AI for military applications. And while announcing their expanded use of AI in their annual report released in August, Microsoft warned of potential liability and ethical issues.
Even with such controversy around it, AI development continued to push the boundaries of its capabilities.
Growing Applications in Human Roles
Although AI still has limitations, 2018 saw it learning to mimic nuanced human interaction and even reaching into time-honored jobs in modern society.
Last year saw Google demonstrate Google Duplex, am AI system capable of natural speech and autonomous communication. While the natural speech flow wowed onlookers at the demonstration of Duplex in May, the system received ethical criticism on the grounds of deception and data protection.
Visual representation of WaveNet's more holistic speech system approach. Screenshots courtesy of Google
Applications also expanded AI into human livelihoods, working as dentists, police patrol officers, and legal assistants. And in the next step beyond playing chess against Grandmasters, IBM's Project Debater took on human opponents in a debate over government subsidy of space exploration and telemedicine. Like Google Duplex, both AI's encroachment on the job market and its new capabilities of engaging in high-level debate garnered praise and criticism.
As the ethics of AI continue to become more complicated, the barriers to using it fell rapidly in 2018. Artificial-Intelligence-as-a-Service (AIaaS) applications became more prominent, from algorithms that recommend products while shopping online, to programs that allow anyone to deploy AI for analyzing large data sets. AI even moved into open-source applications like Mycroft's machine learning home voice assistant, bringing it into home devices.
These advancements are possible due to improving hardware with more ability to handle the heavy demands of AI.
Hardware Improvements Making AI Smarter
The requirement for AI to process large amounts of data at high speed requires more and more sophisticated hardware, and many new developments were announced in 2018.
While earlier years saw the use of cloud computing to perform most AI functions, there has been a growing interest in using edge-computing to provide more responsive AI that does not require a network connection to process data.
In the semiconductor industry, Qualcomm expanded on the use of their Artificial Intelligence Engine on their range of Snapdragon SoCs, integrating AI modules designed to run neural networks.
Screenshot of Snapdragon 845 Mobile Platform animation. Image courtesy of Qualcomm
Crossbar, a non-volatile memory technology company, partnered with Mircosemi last year to include Resistive RAM (ReRAM) on future chip designs. ReRAM's lower-power and faster performance has the potential to bring AI neural networks onto many more embedded systems.
Crossbar's visual aid of their stackable ReRAM structure. Image courtesy of Crossbar
In addition, BrainChip announced in September its production-volume Akida Neuromorphic System-on-Chip (NSoC), an artificial intelligence accelerator utilizing spiking neural networks (SNNs). By utilizing an SNN, the chip can learn more efficiently and with minimal training, mimicking neuron patterns in the human brain.
AI hardware also began growing more compact in 2018. In June, Artosyn Microelectronics licensed the CENA-XM4's low-power intelligent vision process for use in their AR9X01 AI System-on-a-Chip, enhancing AI capabilities in drones.
As hardware improves by lowering power requirements and growing smaller, more applications are becoming available as options for AI.
AI Developments in 2019
So what can we expect from AI in 2019? The future's wide open but here are some educated guesses.
AI Data Processing: From Data Centers to the Cloud to Edge Computing
As AI develops, it continually adapts to better serve the demands of its various applications. The amount of storage and memory needed for AI often requires data center-level infrastructure, many companies began utilizing the cloud for AI in 2017 in order to manage growing speed and data demands. That trend carried solidly through 2018.
2019 will most likely continue to see AI applications embracing the middle ground between data-centers and the cloud, edge computing, which moves memory and storage closer to the end devices utilizing AI.
"AI training performance and throughput is limited by available storage performance," Tom Coughlin, President of Coughlin Associates, stated in a piece on Forbes last week. "This performance can be improved with modern solid-state storage and NVMe interfaces but it also requires less movement of data. So, memory/storage must move closer together to improve AI applications." Coughlin went on to assert that edge networks will be crucial for AI as it is continually knitted into ambitious applications where safety is concerned, such as V2X (vehicle-to-X) communication where latency is paramount.
Smaller and Faster Hardware
Although there have been many advancements in AI already, the modern incarnation of the technology is still very new. Like the earliest computers that were the size of several rooms, AI hardware currently has a pretty big footprint (see the Google server racks below). In 2019, we will see that hardware shrink as it develops. Google's Tensor Processing Units (TPUs), used to power the earlier-mentioned Google Duplex, function in large, liquid-cooled pods in order to perform machine learning with the massive amount of data.
Google's liquid-cooled, 100 petaflops version 3.0 TPU pods. Image courtesy of Google via Twitter
As edge computing grows and hardware is redesigned into smaller components, look for AI to appear in even more applications in 2019.
Bigger Applications, Further Scrutiny
Beyond discussions of what AI is and ethical ways it can or should be used, there are larger conversations coming about the ethics of how AI, itself, functions. A concept that you're likely to hear more about in the coming months is algorithm bias, the idea that a system will adopt the biases of its programmer—whether the programmer is aware of it or not. This concept has been the subject of research for years, conducted by education institutions but also by some of the largest players in the AI development industry, such as IBM.
2018 already brought steps towards mitigating the issue of bias in AI. Facebook, for example, announced in May a tool called Fairness Flow that was designed to identify and eliminate bias in its proprietary AI. Last year, the University of Chicago also unveiled Aequitas, an open-source toolkit for auditing fairness in machine learning, AI, and data science. If you'd like to take part in its open-source development, you can find Aequitas on GitHub. Another open-source bias auditing tool is from Pymetrics and can also be found on GitHub.
The need for these tools is growing. With an election cycle and data security factoring heavily into the future of the United States, AI will be heavily covered in the media again in 2019 regarding big data and its uses. Further debates about regulation and ethical choices will arise as AI becomes more common and has more implications in our daily lives. In 2019, expect a great deal more scrutiny into not only AI systems and their uses, but into the people designing them.
What are your predictions for AI in 2019? Share your thoughts in the comments below.
Featured image used courtesy of GreenWaves Technologies.