State of AI 2021 report is an annual report delivered by Nathan Benaich and Ian Hogarth who are professionals and investors working in the AI industry. The report covers four categories: research, talent, industry and politics, each of them describing major trends both from industrial and scientific perspective.
In this article, I summarize 3 things that interested me the most in the report in terms of how AI progressed and will progress in terms of:
- retail use cases,
- safety and privacy,
- and production first paradigm.
AI for retail
Retail AI use cases are highly dependent on fast and cost-effective processing of a key data type used there – images. And there are at least 3 trends that will support it.
The first one is self-supervised learning, which is a way to train neural networks on non-labelled data. The scenario goes as follows – we use non-labelled data to train the network to understand a concept of image, and we use for that only information encoded inside the data (hence “self”). Then we can add a very small portion of labelled data to adjust a network to our task. Additionally, research showed it is possible to have competitive results even when the non-labelled portion of data is also relatively small. This creates a potential for less restrictive data requirements for retail models.
Another problem occurring in retail is the necessity to present products from different perspectives, and therefore – additional time that you have to spend on photographing it. It is closely related to the deep learning problem called view synthesis. In 2021, a method called Neural Radiance Fields (NeRF) achieved the new state of the art and allowed the generation of multiple views of an object presented in one static image. It also makes it easier to create those perspectives without changing the background, meaning we only manipulate an object of interest.
Last but not least, a problem persisting in retail, especially in selling products second-hand, is the quality of photos. Very often photos are dark and have distracting background, but sellers may not have time nor resources to provide better ones. This is where ClipDrop comes in handy. ClipDrop allows you to extract a desired object from a photo and remove any distraction-making background, dropping it onto a more neutral one. Apart from solving a specific retail problem, this step may also be beneficial for the environment by making it easier for people to quickly sell their spare things instead of throwing them out.
Safety and Privacy
Every data related privacy breach and every major failure of an AI system causing human harm are raising the concern that stronger regulations in AI should be required. The European Union seems to be a particular supporter of regulations which led to introduction of the AI Act.
One of the propositions included is categorizing AI products into 3 categories: prohibited, high- and low-risk, which would not only apply to products of EU companies, but also foreign companies that provide their software to EU citizens. All algorithms would be therefore scrutinized by EU organs, and the higher the risk, the more requirements and expectations towards the system would apply. For example high-risk algorithms would need to be transparent, secure and well-documented.
The main problem however is that some of the requirements are still a topic of research in the scientific community. Apart from that, legal and economic circumstances lead the authors of the report to believe that the EU may not have yet the means to enforce such a law. But it is something to keep in mind.
Apart from propositions of policy-makers, there are also solutions presented by the industry to tackle the problem of data privacy. We need to remember that data privacy is not only important from the perspective of stealing the data, but also generating new data to assume someone’s identity (you can gain some interesting perspective on deepfakes here). One of the answers is federated learning. It is a type of decentralised learning where we do not send user data to the centralised algorithm to train it, but we train its instances on users’ devices. This mechanism is already used by Brave, a computer browser which does not share user data to the cloud to adjust ads targeting, but adjusts them on-device.
Production first AI
Year 2021 showed an even bigger increase in machine learning automation tools, such as automatic model creation or data versioning. And it supports a new paradigm that emerges in AI – data-centric AI. This paradigm shifts our attention from model to data by introducing data iteration process instead of model search as a core component of the project.
Many AI projects failed due to the fact that the main focus was an algorithm and not data (aka model-centric). At some point of the implementation, our data were fixed, the model improved, and after deployment to production creators realised that model does not reflect data and current situation anymore. Or never did because they were a reflection of someone’s biases. Those lessons were learned by both small and big players, such as Amazon through its hiring bot or Microsoft through its tweeting bot Tay.
Data-centric approach is different. One of its principles is active benchmarking that assumes the model is fixed and we share it with potential users that can interact with it and produce data that are more reflecting real world situations. It also allows us to measure if a problem we’re tackling is prone to concept shifts – the situations where data, or rather phenomenon they describe, change over time and it has to be adjusted to those time differences.
So if your team is not yet familiar with MLOps and data tracking tools – it’s definitely high time to catch up.
Conclusions
The State of AI report shows multiple directions where the AI industry and research are heading, and it’s always good to be up to date with predictions made by the authors – they have a high chance of happening. But if you don’t want to work in AI as a developer but still love the area, the report gives one other tip in terms of what professions we may need in the future. It appears that we see a rise in research communication – Twitter threads are driving citations up to 3 times, so the necessity for AI and tech writers that can communicate easily rises.
So do me a favour – grab your pen (or keyboard) and write this article that crossed your mind a long time ago.
Sources (all found in State of AI Report)
[1] State of AI 2021 Report
[2] M. Niemeyer, A. Geiger, GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields
[3] ClipDrop
