Three trends that shaped the work of data practitioners in 2022
There’s never been a more exciting time to work in data. As artificial intelligence (AI) and machine learning (ML) continue to evolve, the potential of what can be achieved appears to be near limitless. Innovations in AI will continue to be some of the most transformative technologies impacting how we live. Generative AI is already pushing technology into a realm thought to be exclusive to the human mind: creativity.
As we look ahead to some of our work on the horizon, we’ve reflected on some of the trends from 2022 that we’re most excited about. Three of our data experts have each chosen their most impactful development from the past 12 months that they think every data practitioner should be aware of in 2023.
1) Multi/Cross-modality — navigating via language
Karan Wanchoo, Data Science Specialist
The last year has seen great leaps towards embodied autonomous agents, essentially robots intelligent enough to accurately execute instructions that are not wholly defined via complex code. Instead, ‘agents’ are told to do something and simply do it. At a basic level, this might be, for example, the ability to navigate an unfamiliar place after receiving the basic, human-language instruction told to ‘go down the hall and turn right into the kitchen.’
While the work is still in its infancy, with challenges to overcome, a recent paper hypothesized that a language-driven agent could navigate any unseen and unfamiliar household environment. This paper presents a novel AI architecture entailing two cross-modal attention modules that learn language-informed representations.
Building on the hypothesis, their tests show it is possible to successfully blend language modality (instructions) and vision modality (household map perceived by the agent) and take us closer to the realization of true embodied intelligence.
2) Participatory AI — creating fair systems
Carla Vieira, Data Engineer
Participatory AI can counter the ingrained, multi-generational bias ingrained in historical data. It goes beyond the developers to engage the broadest possible group of stakeholders. Co-creation is the key to empowerment, ensuring that a new tool, system, model, or application can authentically back up its claims of safety, legitimacy, and responsibility.
By involving those affected, we can rethink AI-based products, validate their value and avoid creating products that exacerbate social inequalities due to biases in the development process. Also, it allows for building trust with its users. AI-based products have more than just financial impact; they also create social impact for their users.
The development of AI will increasingly include fairness, compliance with AI ethics regulation laws and user co-participation as requirements. Those developing AI products need to be prepared for that when it happens. This trend is set to be the norm in 2023 and beyond.
3) MLOps — monitoring models and avoiding technical debt
Devanshi Verma, Data Scientist
MLOps is a growing and important discipline. It’s the heart of the operation, really — continuing to beat and keep things running even after some of the ‘cooler’ initial model development has been done. MLOps includes monitoring inference data for ‘drift’ — a tendency to differ from the training data on which the model was built — which could cause ML-driven applications’ performance to deteriorate over time.
Monitoring and updating is also a crucial best practice for mitigating bias and improving fairness. Proper data and model monitoring enable teams to detect shifts leading to quicker intervention. Given the implication of model drift and technical debt, MLOps will avoid the tedious process of manually comparing a selected window of live traffic with a baseline.
In a world as fast-paced as AI and ML, the wheel of progress continues to roll. We’re eager to hear about any other trends you believe are worthwhile to watch. Please leave your suggestions in the comments.