Artificial General Intelligence and The Potential Role of Edge AI

On July 22, 2019, Microsoft announced an investment of US$1 billion into OpenAI, a San Francisco-based non-profit Artificial Intelligence (AI) research group that focuses on the development of Artificial General Intelligence (AGI). This announcement brings two of the largest names in the AI industry together: Microsoft owns Azure, one of the most popular cloud AI platforms, and is behind many key AI initiatives, including Open Neural Network Exchange (ONNX) and Project Brainwave, while OpenAI has achieved significant breakthrough in deep learning technology, creating one of the most advanced text generators based on its GPT-2 Framework and a gaming AI that has defeated human players in multiplayer online battle arena video games.

In addition, the amount that Microsoft invested is a noteworthy sum. To put this into context, ABI Research estimated the total global venture capital funding in AI to be US$10.7 billion and US$18.4 billion in 2017 and 2018 respectively. Given the significant progress that the industry has achieved as a whole in the domain of machine vision, Natural Language Processing (NLP) and AI chipset, Microsoft’s US$1 billion investment is going to generate great research and applications. OpenAI can leverage Microsoft’s cloud infrastructure for its AI research, while getting a great channel to commercialize their AI technologies.

The key focus of this new partnership is on AGI. Currently, most, if not all, of the commercially successful AI products run on narrow AI, or AI that focuses on one specific task and executes that task very well. This includes customer service chatbots, smart industrial cameras for defect detection, and self-driving cars. On the contrary, AGI is designed to mimic human intelligence, enabling the AI to learn and make decisions like a human. This is a vision that has been around since the inception of IBM’s Watson, but the industry has made little progress thus far.

OpenAI and Microsoft’s vision is for AGI is to help humans solve currently intractable multidisciplinary problems, including global challenges such as climate change, more personalized healthcare, and education, but designing an AGI is extremely challenging. As such, there is no unified consensus across the industry regarding the timeline for AGI’s readiness and commercialization. Many fundamental works for AGI are still in their early stages, as AGI will require nascent machine learning techniques. Here are some of the key challenges of implementing AGI:

• The Need to Multi-Task: Currently, AI is trained to produce a single output under a set of clearly defined inputs. In order for AI to be able to multi-task, it needs to be able to adapt the trained model to produce similar or expected outputs when given a different but related set of data. This requires transfer learning, which involves the transfer of training instance, feature representation, parameters, and relational knowledge from the existing trained AI model to a new one that addresses the target task. Such process has a great impact on AGI’s accuracy and reliability if implemented sub-optimally.
• The Capability to Self-Optimize: AGI also needs to have the ability to self-manage its resource demands. AGI is expected to be more efficient in domains such as memory management, power consumption, and even model selection and hyperparameter tuning in some automated machine learning scenarios. This would require AGI to also identify errors within its system. While techniques like reinforcement learning can be used to reward the right behavior and penalize the wrong decisions, such optimization models are still in very early stages.
• The Ability to Handle Incomplete Information: AGI needs to make decisions under circumstances with incomplete data and often on the go. This requires AGI to be heuristic and, in some cases, creative. Depending on how the AGI is designed, the AGI performance will suffer from the lack of data. In comparison, narrow AI is trained and thoroughly tested before it is deployed in commercial settings to ensure its reliability and accuracy. This is not a luxury that AGI can expect to have.
• The Need for More Powerful AI Software and Hardware: Microsoft is partnering with OpenAI to form one of the strongest hardware-software vendor tie-ups in their ambition to tackle AGI. However, AGI may require stronger hardware that is designed to handle probabilistic computing by nature, such as quantum computing and neuromorphic chipset.

Nonetheless, all these shortcomings clearly have not deterred Microsoft from partnering with OpenAI on AGI. In recent years, Microsoft has poured in many resources to develop its AI capabilities in the form of Cortana. As mentioned in the ABI Insight A Fresh Look at Human-Machine Interaction (IN-5504), Microsoft showcased Cortana’s ability to hold a natural conversation by assimilating and presenting contextual information in real time, thanks to its acquisition of NLP startup Semantic Machines, at Microsoft Build 2019. It is reasonable to expect that Microsoft is looking to create an AGI out of Cortana.

While all these developments happen in the cloud environment, ABI Research believes that the edge has a role to play. Admittedly, AGI at the edge will serve a very different role. Instead of solving climate change and personalized healthcare challenges, edge AGI may serve as a smart mobile robot in public safety or the central brain for a dark factory or warehouse.

It is important to note that many innovations and nascent AI techniques are currently being trialed and tested on narrow AI at the edge, often in environments with poor data connectivity, limited data, and demand for high mobility and low latency. These include localization and navigation models on mobile robots, connected home applications on smart home gateways, and manufacturing operation optimization and predictive maintenance models in on-premises servers in factories. The richness of data collected by end devices and residing in gateway and on-premises servers can be useful in developing AGI. Like humans, AGI makes decisions based on data collected from multiple sources in real time, and therefore requires the inclusion of characteristics and features of narrow AI models dedicated to edge.

In addition, edge AI hardware has grown more powerful over time. With the demand for low latency and localized AI processing capabilities, there will always be demand for edge-based AGI, be it on a device, gateway, or on-premises server. Federated or distributed learning can provide incremental upgrades to AGI. Nevertheless, these developments are still in the nascent stage, even in edge devices, so it will take a while before they will make it into AGI.

As such, AGI is still a long-term vision for the industry. It is definitely exciting to see what will come forth from the partnership between Microsoft and OpenAI. As the hardware provider in this tie-up, it would be interesting to see the direction Microsoft wants to take with its investment in AGI in the future, whether the cloud AI giant will invest in new AI computing architecture, beef up its edge AI offerings, or, like many of its competitors, start to develop its own cloud and edge AI chipsets.