This article is more than 1 year old

Alibaba gazes into crystal ball to spy coming advances in AI and silicon photonics

Machine learning to propel us into glorious era of scientific discovery

Alibaba has published a report detailing a number of technology trends the China-based megacorp believes will make an impact across the economy and society at large over the next several years. This includes the use of AI in scientific research, adoption of silicon photonics, the integration of terrestrial, and satellite data networks among others.

The Top Ten Technology Trends report was produced by Alibaba's DAMO Academy, set up by the firm in 2017 as a blue-sky scientific and technological research outfit. DAMO hit the headlines recently with hints of a novel chip architecture that merges processing and memory.

Among the trends listed in the DAMO report, AI features more than once. In science, DAMO believes that AI-based approaches will make new scientific paradigms possible, thanks to the ability of machine learning to process massive amounts of multi-dimensional and multi-modal data, and solve complex scientific problems. The report states that AI will not only accelerate the speed of scientific research, but also help discover new laws of science, and is set to be used as a production tool in some basic sciences.

As evidence, the report cites that fact that Google's DeepMind has already used AI to prove and propose new mathematical theorems and assisted mathematicians in areas involving complex mathematics.

Renewable energy

One unusual area where DAMO sees AI having an impact is in the integration of energy from renewable sources into existing power networks. Energy generated from renewable sources will vary depending on weather conditions, the report states, which are unpredictable and may change rapidly, thereby posing challenges for integration of renewable energies such as maintaining a stable output.

DAMO states that AI will be essential to solving these challenges, in particular being able to provide more accurate predictions of renewable energy capacity based on weather forecasts. Intelligent scheduling using deep learning techniques should be able to optimise scheduling policies across energy sources such as wind, solar, and hydroelectric.

The use of big data and deep learning technologies will be able to monitor grid equipment and predict failures, according to the report, so perhaps in the near future you will blame the AI when the power cuts out just as you are trying to binge-watch Line of Duty.

Smaller scale

DAMO also believes that we will see a shift in the evolution of AI models, away from large-scale pre-trained models such as BERT and GPT-3 that require huge amounts of processing power to operate and therefore consume a lot of energy, to smaller-scale models that will handle learning and inferencing in downstream applications.

According to this view, the cognitive inferencing in foundational models will be delivered to small-scale models, which are then applied to downstream applications. This will result in separately evolved branches from the main model that have developed their own perception, decision-making and execution results from operating in their separate scenarios, which are then fed back into the foundational models.

In this way, the foundational models continually evolve through feedback and learning to build an organic intelligent cooperative system, the report claims.

There are challenges to this vision, of course, and the DAMO report states that any such system needs to address the collaboration between large and small-scale models, and the interpretability and causal inference issues of foundational models, as the small-scale models will be reliant on these.

Silicon photonics

Silicon photonics has been just around the corner for many years now, promising not just the ability for computer chips to communicate using optical connections, but perhaps even using photons instead of electrons inside chips. DAMO now expects we will see the widespread use of silicon photonic chips for high-speed data transmission across data centres within the next three years, and silicon photonic chips gradually replacing electronic chips in some computing fields over the next five to ten years.

The continuing rise of cloud computing and AI will be the driving factors for technological breakthroughs that will deliver the rapid advancement and commercialisation of silicon photonic chips, the report states.

Silicon photonic chips could be widely used in optical communications within and between data centres and optical computing. However, the current challenges of silicon photonic chips are in the supply chain and manufacturing processes, according to DAMO. The design, mass production, and packaging of silicon photonic chips have not yet been standardised and scaled, leading to low production capacity, low yield, and high costs.

Private computing

Privacy is another area where DAMO believes we will see advances in the next few years. It states that techniques already exist that allow computation and analysis while preserving privacy, but widespread application of the technology has been limited due to performance bottlenecks and standardisation issues.

The report predicts that advanced algorithms for homomorphic encryption, which enables calculations on data without decrypting it, will hit a critical point so that less computing power will be required to support encryption. It also foresees the emergence of data trust entities that will provide technologies and operational models as trusted third parties to accelerate data sharing among organisations.

Satellite connectivity

Another prediction from DAMO is that satellite-based communications and terrestrial networks will become more integrated over the next five years, providing ubiquitous connectivity. The report labels this as satellite-terrestrial integrated computing (STC), and states that it will connect high-Earth orbit (HEO) and low-Earth orbit (LEO) satellites and terrestrial mobile communications networks to deliver "seamless and multidimensional coverage."

There are major challenges to implementing all this, of course, including that traditional satellite communications are expensive and use static processing mechanisms that cannot deliver the requirements for STC, while hardware for satellite applications is not commonplace and hardware for terrestrial applications cannot be used in space.

Cloud-network-device convergence

Finally, the DAMO report predicts the rise of what it calls cloud-network-device convergence. This appears to be based on the premise that cloud platforms offer a huge amount of compute power, while modern data networks can provide access to that compute power from almost anywhere, so that endpoint devices only need provide a user interface.

Yes, it's the thin client concept emerging again, this time using the cloud as the host. Clouds allow applications to break free of the limited processing power of devices and deliver more demanding tasks, according to the report, while new network technologies such as 5G and satellite internet need to be continuously improved to ensure wide coverage and sufficient bandwidth.

Just by sheer coincidence, Alibaba Cloud already has such devices, with the handheld "Wuying" launched in 2020 and a more substantial desktop device shown off last year.

Naturally, the DAMO report expects to see a "surge of application scenarios on top of the converged cloud-network-device system" over the next two years that will drive the emergence of new types of devices and promise more high quality and immersive experiences for users. ®

More about

TIP US OFF

Send us news


Other stories you might like