Stanford College’s Human-Centered Articial Intelligence has lately printed its complete annual assessment of worldwide AI expertise, funding and adoption developments.
What can AI do now?
AI continues to advance in leaps and bounds, and in lots of areas is considerably extra expert than human benchmarks. Nonetheless, AI appears to be plateauing in additional complicated duties – which the AI Index suggests means a breakthrough now relies on an as-yet undiscovered new AI expertise.
The AI outperformers are:
- Facial recognition: In 2017, a number of the top-performing facial recognition algorithms had error charges of over 50%. However by 2021, the top-performing apps registered an error price of simply 0.1%, which means that for each 1,000 faces, the mannequin accurately recognized 999. Even in a COVID world, the highest apps solely carried out 5 to 16% worse on masked faces in comparison with unmasked ones.
- Voice to textual content: In 2021, for each 100 phrases that top-performing transcription apps heard, they accurately transcribed 99.
- Language translation: translation apps have now reached a workably proficient degree. Since 2015, there was a 23.7% enchancment in English-French translation potential and a 68.1% enchancment in English-German translation potential (after lastly mastering Teutonic grammar).
- Medical AI: Medical picture segmentation refers back to the potential of AI programs to section objects of curiosity, comparable to organs, lesions, or tumours in medical pictures. AI programs at the moment are able to accurately segmenting colonoscopy polyps at a price of 94.2%, representing an 11.9 share level enchancment since 2015.
- Studying comprehension: In 2021, the perfect AI scores 95.7 on studying comprehension, exceeding the human benchmark of 91.2.
- Deepfakes: Many AI programs can now generate faux pictures which can be indistinguishable from actual ones. In 2012, the top-performing programs might accurately establish 69.9% of deepfakes however in 2021 that quantity elevated to 97.7%.
- Human pose recognition: That is the duty of estimating totally different positions of human physique joints (arms, head, torso, and many others.) from a single picture. Human pose estimation can be utilized to facilitate exercise recognition for functions comparable to sports activities analytics, crowd surveillance, CGI growth, digital atmosphere design, and transportation: for instance, an AI-controlled aircraft figuring out the physique language indicators of an airport runway controller. In 3 dimensional pose recognition, again in 2015, the top-performing app was making a median per joint error of 16 centimetres, half the dimensions of an ordinary faculty ruler, however in 2021 this fell to 1.9 centimetres, lower than the dimensions of a median paper clip.
- Sentiment recognition: This includes AI recognising the sentiment of a bit of textual content (very destructive, destructive, impartial, constructive, very constructive). Sentiment evaluation will be easy if sentences are worded clearly and unambiguously, comparable to “I dislike winter climate.” Nonetheless, sentiment evaluation can grow to be tougher when AI programs encounter sentences with flipped buildings or negations, comparable to “to say that disliking winter climate will not be actually my factor is totally inaccurate.” By 2021, the top-performing apps estimate sentiment accurately 9 out of 10 instances, in comparison with solely 7 out of 10 instances again in 2015.
- Visible process recognition: This includes AI recognising actions like strolling, waving, or extra complicated duties like making ready a salad (which requires an AI system to acknowledge and chain collectively discrete actions like slicing tomatoes, washing the greens, making use of dressing, and many others). In 2021, AI accurately recognised 89.6% of easier duties however this fell to 82.2% accuracy for extra complicated duties (pouring wine is outwardly a fancy process!). Nonetheless, most strikingly, within the single 12 months of 2021, the accuracy hole between AI recognising easy and sophisticated duties narrowed from 27 factors to solely 7.4 factors, which implies that enhancements in AI efficiency on harder-to-recognise duties is going on extra quickly than efficiency on simpler duties.
Nonetheless, on extra complicated duties, AI nonetheless has a method to go. For instance, Visible Common sense Reasoning (VCR) includes AI programs answering difficult questions on situations introduced from pictures, and likewise offering the reasoning behind their solutions. Within the picture under, the AI is requested why is one visitor pointing his finger at one other visitor when the waiter turns up:
Funding and R&D developments
In 2021, the best AI funding globally got here by way of personal funding (totalling round US$93.5 billion), adopted by mergers and acquisitions (round US$72 billion), public choices (round US$9.5 billion), and minority stake (round US$1.3 billion).
Personal funding greater than doubled from 2020, which was the best year-over-year improve since 2014. Personal funding in AI in Australia was US$1.25 billion.
The dimensions of offers additionally elevated markedly. In 2020, there have been 4 funding rounds price $500 million or extra, however in 2021, there have been 15. On the similar time, the variety of newly funded AI corporations continued to fall 12 months on 12 months, from 1,051 corporations in 2019 and 762 corporations in 2020 to 746 corporations in 2021.
Over the past 5 years, most personal funding globally in AI has been poured into medical and healthcare AI ($28.9 billion); adopted by information administration, processing, and cloud ($26.9 billion); fintech ($24.9 billion); and retail ($21.95 billion). Nonetheless, in 2021, the best personal funding in AI shifted to information administration, processing, and cloud (round $12.2 billion), 2.6 instances the funding in 2020.
The variety of AI-related patents filed in 2021 is greater than 30 instances increased than in 2015, a compound annual progress price of 76.9%. After a late begin, East Asia and Pacific (primarily China) accounted for 62.1% of all patent purposes. America recordsdata patents at one-third the speed of China. But regardless of the geo-politics, the USA and China had the best variety of cross-country collaborations in AI analysis in 2021, growing 5 instances since 2010.
On constructing an AI-skilled workforce, Australia appears to lag behind comparable economies. In 2021, New Zealand had one of many world’s highest growths in AI hiring—2.42 instances higher in 2021 in contrast with 2016. In 2021, Australia was just one.14 instances higher than its 2016 degree.
The AI talent penetration price within the workforce means the prevalence of AI abilities throughout occupations (primarily based on LinkedIn member’s use AI abilities of their job!). India led the world within the price of AI talent penetration—3.09 instances the worldwide common from 2015 to 2021. However Australia was solely at 0.77 of the worldwide price.
Nonetheless, in some excellent news for Australia, among the many 15 nations listed, the AI talent penetration charges of females are increased than these of males in India, Canada, South Korea, Australia, Finland, and Switzerland.
Sluggish AI adoption by enterprise
Whereas AI expertise and funding races forward, a survey of enterprise leaders globally (by McKinsey) discovered that economy-wide adoption of AI is way slower. Globally, the adoption price solely elevated by 6% in 2021. Developed Asia Pacific, which incorporates Australia, noticed adoption improve by solely 3%.
Globally, the best adoption was in product and/or service growth for prime tech/telecommunications (45%), adopted by service operations for monetary providers (40%), service operations for prime tech/telecommunications (34%), and threat perform for monetary providers (32%).
The most important boundaries to enterprise adoption in 2021 have been cybersecurity (55% of respondents), adopted by regulatory compliance (48%), explainability (41%), and private/particular person privateness (41%). However considerations over cybersecurity dangers fell a bit of from the 2020 survey.