Scaling to AI Breakthroughs Featured Signal of Change: SoC1249 August 2021
AI developers are struggling to advance AI and address various issues—including AI biases, data problems, and poor methodologies—that inhibit commercialization. As industry stakeholders look for practical AI technologies, researchers and technology firms are exploring new strategies to enable AI systems with greater flexibility, accuracy, and reasoning capabilities. Leading AI developers envision that the production of larger models in tandem with advances in high‑performance-computing systems will enable the next significant jump in AI performance. AI companies and researchers believe that by increasing the scale of existing AI models, training data sets, and dedicated hardware, they can enhance the performance and efficiency of state-of-the‑art technologies and enable new opportunities in a wide range of areas, including natural-language processing and safety-critical computer vision.
Researchers and technology firms are exploring new strategies to enable AI systems with greater flexibility, accuracy, and reasoning capabilities.
During the annual conference of the Beijing Academy of Artificial Intelligence (BAAI; Beijing, China) in early June 2021, BAAI researchers presented the Wu Dao 2.0 multimodal deep‑learning model, which can perform natural-language processing, text generation, image generation, image recognition, and a wide variety of other tasks. The researchers trained the model on 1.75 trillion parameters, making the model an order of magnitude larger than OpenAI's (San Francisco, California) GPT‑3 (Generative Pre‑trained Transformer 3) language model. During the conference, BAAI chairman Dr. Zhang Hongjiang suggested that large AI models and more powerful computational systems will be key enablers of artificial general intelligence and other advances in AI. Other AI developers, including Google (Alphabet; Mountain View, California), have emphasized a similar sentiment and have proposed that large AI models and data sets and further advances in high‑performance computing might enable significant progress in deep learning and address existing limitations. However, an industry shift to larger models and data sets will likely necessitate innovations in time‑efficient training techniques and dedicated AI chips. Several research groups are investigating new algorithms that may optimize the performance of industry-standard AI accelerators, including graphics-processing units and application-specific integrated circuits. For example, researchers at the Massachusetts Institute of Technology (Cambridge, Massachusetts) Computer Science and Artificial Intelligence Laboratory developed an optimized algorithm that improves on traditional matrix-multiplication operations in use to enable fast and efficient AI. By removing multiplication and addition steps, the researchers were able to run the algorithm 100 times faster than existing matrix-multiplication techniques and 10 times faster than current approximation methods. Although the algorithm has undergone testing on only a low‑end central processing unit at a small scale, the researchers claim that it is deployable with no run‑time overhead and includes all the dependent components for summing low‑bit-width integers. Importantly, the researchers anticipate that in comparison with current AI accelerators, new hardware implementations based on the algorithm might enable significant improvements in efficiency. The innovation might inspire new software and hardware solutions that help address current performance bottlenecks for natural-language processing, object recognition, and other AI tasks.
In addition to the advances to state-of-the‑art methodologies that are occurring, research is emerging that may inspire new AI models and computational architectures that enable disruptive advances in the field of AI. For example, researchers at Google and Harvard University (Cambridge, Massachusetts) developed a searchable 3D map (connectome) of a small part of the cerebral cortex of the human brain. The map, which accurately describes neural connections, represents only one‑millionth of the entire volume of the human brain; however, it is one of the most complete maps of the human brain that has ever seen development, featuring 50,000 cells, 130 million synapses, and various other cellular components. Although larger connectomes of the human brain will likely take years to develop, this small-scale connectome might inspire innovations in neuromorphic AI hardware and new algorithms that simulate the structure of the human brain. Researchers anticipate that employing brain‑like neuromorphic architectures might be one way to enable AI models that have better predictive capabilities and flexibility than traditional deep‑learning technologies have. However, although neuromorphic technologies might enable disruptive advances, adoption of such technologies would require significant changes to chips, software, high‑performance-computing systems, and other technologies currently in use. Significant advances in AI‑model reasoning, flexibility, accuracy, and speed promise to change the way that users interface with AI systems in their everyday lives. Such systems promise to become more conversational, predictive, and intuitive to use. Researchers continue to conceptualize how future AI systems might affect labor and everyday life. In a recent publication, Google researchers envisioned how future iterations of AI‑enabled search engines might provide reliable information and better understand questions through advances in natural-language understanding. Specifically, the researchers propose the synthesis of conventional informational theory with large pretrained language models as a solution to the limited answering capabilities of existing models. Other research groups have suggested similar strategies, and if the new approach is successful, it might enable accessible cloud-based platforms that can provide end users with accurate expert-level analysis. Such a system promises to have important implications for knowledge workers and strategists who rely on repositories of accurate open‑source information to perform their jobs.
The AI-research community is at a crossroads, exploring various methodologies to deploy practical and reliable AI that can deliver on complex enterprise needs and serve as a usable and functional tool for general users. On the one hand, a brute-force approach that includes the scaling of accelerators, data sets, computational systems, and every other enabling technology for AI might bring about significant advances in AI. On the other hand, state-of-the‑art AI models may have reached the point at which a new technology paradigm is necessary to enable the next significant leap in progress. In this case, unconventional models and new modes of computing might be the key to unlock AI's disruptive potential and address core performance barriers. Both paths present significant challenges for technology developers. Scaling AI models and data sets promises to increase training times, driving the need for powerful dedicated AI chips and other hardware that can efficiently train and run large models. Alternatively, the pursuit of novel AI models and computational systems will require technology developers to encourage private industry to adopt such systems, creating significant hurdles to commercialization. Because demand for practical AI is increasing across nearly every industry, promising solutions are bound to arise.