How to avoid falling down the Data Cascade?

Model is King, but Data is what’s flowing in his veins. An exciting study points out that project teams in the AI industry are primarily driven to construct and refine engaging AI models (create "magic" with the algorithms), yet the task of managing the data necessary for the training and operation of these models is often viewed as mundane. This aspect of AI development is not seen as 'glamorous' and fails to receive the spotlight it deserves, to the point where it can jeopardize the success of the project itself. Many AI initiatives struggle to even begin, for this precise reason. That’s a worrying thought for anyone endeavoring to weave machine learning into the fabric of their business operations. We underscore some of the pivotal findings from the study and intertwine them with insights from our own journey through numerous AI projects. If you're keen on exploring related topics or potential applications of AI further, you're invited to check here. This can enhance your understanding and potentially offers new perspectives on the integration and utilization of AI in various domains.

This report is based on the results of a study on practices and structural factors among 53 AI practitioners in India, the US, and East and West African countries, applying AI to high-stakes domains, such as landslide detection, suicide prevention, and cancerdetection.

AI projects in the study, by industry:

• Health and wellness (19) (e.g., maternal health, cancer diagnosis, mental health)
• Food availability and agriculture health (10) (e.g., regenerative farming, crop illness)
• Environment and climate (7) (e.g., solar energy, air pollution)
• Credit and finance (7) (e.g., loans, insurance claims)
• Public safety (4) (e.g., traffic violations, landslide detection, self driving cars)
• Wildlife conservation (2) (e.g., poaching and ecosystem health)
• Aquaculture (2) (e.g., marine life)
• Education (1) (e.g., loans, insurance claims)
• Robotics (1) (e.g., physical arm sorting)
• Fairness in ML (1) (e.g., representativeness)

What is a Data Cascade?

The study is focused on ‘Data Cascades’ - situations where an AI project goes wrong, specifically because of a data problem, not a model problem.

A ‘cascade’ is an event that is causing a ‘technical debt’ - which means the project hits an unforeseen obstacle requiring extra effort for getting the desired results, or derailing the project entirely or momentarily, or at least significantly lessening the efficiency of the solutions in use.

Staggering, but true: 92% of all the practitioners asked in the study experienced at least one event which could be identified as a Data Cascade.

Of course, these are not your everyday automation solutions, but fairly complex endeavours.

The writers identified a ‘high-stake’ AI project as something that is characterised by:

• high accountability
• the requirement for inter-disciplinary work
• constrained resources

You could say that in the business world, appropriate resources and expertise should be available, but the reality is that sometimes you have to make things work under less than ideal circumstances. We strongly believe that every business leader can learn from these examples.

The practitioners tried their best to get things done, obviously. And, because they are human beings lacking the complete perspective of their projects, they sometimes tried to make the model work even though the data was partially compromised. The authors of the study say that the practiotiners ‘did not appear to be equipped to recognise upstream and downstream people issues’.

Clear communication is key, protocols are essential

Quite simply, when the developers design a process, the feedback of the field workers are often not integrated early enough. The people who are responsible for the running and maintenance of the AI project itself are not trained thoroughly enough to realize the importance and value of  appropriate data handling. Data becomes fuzzy, scarce, and it’s dependent on the partners.

Data work is often ‘taken for granted’, it’s not properly recognized. The practitioners quoted in the study described it as ‘time-consuming, invisible to track, and often done under pressures to move fast due to margins— investment, constraints, and deadlines often came in the way of focusing on improving data quality’.

The study points out that it’s sometimes ‘difficult to get buy-in from clients and funders to invest in good quality data collection and annotation work, especially in price-sensitive and nascent markets like East and West African countries and India. ‘

It’s a bit like running a mill, but not caring about the quality of the grain. Which will adversely affect the quality of the flour, for sure.

Commitment cannot waver through any phases, including the data phase

The study goes further: ‘Clients expected ‘magic’ from AI—a high performance threshold without much consideration for the underlying quality, safety, or process—which led to model performance ‘hacking’ for client demonstrations among some practitioners. ‘ We’re sure that there are many AI project managers out there who can relate - you want the project to succeed, but you also want to do it within the limitations of the expectations coming from further up the chain of command. Cutting corners can be tempting!

It seems like we are facing a systemic problem which makes AI projects so results-oriented that an unrealistic expectation is being developed by the clientside. The study points out that AI trainings and courses think of data as a 100% clean and available input material, but in real life, you never see such things as ‘clean data’.

Within IT, we’ve got data scientists for that, but when the input source is not fully digital, like in many of the high-stakes projects, their efforts come at a stage where it might be too late. This leads to a whole lot of AI projects going unfinished.

Causes of a cascade

So what are the main causes for data cascades? What can derail an AI project before it even gets going?

1. Physical world effects

Weather, wind, sand and dirt can mess with sensors and camera images, easily. These effects should be thoroughly analyzed before the data flow even starts, because even surprisingly small contamination can cause disturbances and lead to cascades.

We need to make sure that the physical limits and risks are thoroughly accounted for, and we have regulations in place for handling them.

2. Inadequate application-domain expertise

When the people working on the data input systems have to make decisions that are beyond their skills and knowledge, they will be forced to make guesses and they will make mistakes. Discarding, correcting, merging or a full sequence restart requires careful consideration to do effectively, which is unrealistic to expect from someone without proper preparation. This kind of problem plagued a lot of AI projects in the study, from healthcare to insurance. If the ‘ground truth’ is set incorrectly, the model will not be able to provide meaningful results.

The workers responsible for managing data entry should receive proper training or should be helped by qualified professionals, even with realtime remote guidance if necessary.

3. Conflicting reward systems

Sounds simple, but the researchers found that the tasks of data collection were simply added to the regular duties of the professionals, without them being compensated for it. It was extra work for them, for no extra pay. Obviously, data quality suffered.

4. Poor data documentation

The study found that there were several cases where data collection was subpar, and the practitioners had to make assumptions to fill the void,  and sometimes discard whole datasets - in one case, 4 months of valuable medical research data.

The problem is that standards and conventions can be different between organizations, groups of professional, and even minor differences can trigger a data cascade, so this should definitely be a focus point of every project where data gathering and analysis happens within different teams.

Conclusions

As it is written in the study: ‘Our results indicate the sobering prevalence of messy, protracted, and opaque data cascades even in domains where practitioners were attuned to the importance of data quality.  We need to move from current approaches that are reactive and view data as ‘grunt work’. We need to move towards a proactive focus on data excellence.’

There should be clarified processes and standards to make sure data quality is sound through every phase. This requires a proper infrastructure and appropriate incentives for everyone involved.

‘Despite the primacy of data, novel model development is the most glamorised and celebrated work in AI—reified by the prestige of publishing new models in AI conferences, entry into AI/ML jobs and residency programs, and the pressure for startups to double up as research divisions.’ Model work gets almost all the spotlight, so business leaders should not make this mistake. Carefully consider every segment of the pipeline, and do not let yourself swayed away from allocating the right amount of resources to the data work.

This probem starts at the education level. AI literacy is almost synonymous with model development, so there aren’t enough graduates skilled in the art of working with data, according to a research from 2016.

There’s an old meme quote in the NLP community, which unironically highlights the problems around the perception of data work: “Every time I fire a linguist, the performance of the speech recognizer goes up” . This is a prime example of ‘cutting corners’, generating quick value while sacrificing long term development.

When the participants of the study have been asked about good practices, they basically mentioned nothing new or revolutionary, they just brought up tools and methods that are already used in software development, but the data side sometimes simply doesn’t get that treatment:

• shared style guides
• thorough documentation
• peer review
• clearly assigned roles

All of these solutions ‘compound uncertainty’, help projects avoiding data contamination.

What should you do to save your own precious AI project from tumbling down a data cascade? We hope we’ve been able to give you some ideas, and direct your attention towards the data side of the AI projects - it surely is the unsung hero which can make or break the value generation process!

The cover image is from the study itself. If you are interested and want to know more, but don't have time to read the whole 15 pages, here's a 5 minute video about the findings of the research!