“With a lot of access to AI technology, just ten lines of code, you can embed any AI algorithm to perform complex functions like object detection, object recognition and great attributes,”said Navin Dhananjaya, Chief Solutions Officer at Ugam, A Merkle Company, during his talk at Machine Learning Developers’ Summit (MLDS) titled, “The 3 Building Blocks of AccessibleAI”.
The 3 building blocks of accessible AI are:
- AI-ready no code/low code cloud platform
- AI-trained multidisciplinary ecosystem
- Integrated AI operations
He said the companies are reaping the benefits of using AI with high returns on investment across use cases. However, using AI comes with its own set of challenges:
The challenges include:
- Lack of a good mix of native data platforms, AI toolsets and partly cloud technologies.
- Niche skilled manpower (data scientist, data SMEs) requirement.
- No formal training or understanding of AI among business teams.
- Minimal performance monitoring and reuse of AI models.
- Inconsistent and time consuming process for getting data and models into production.
Navin Dhananjaya said accessible AI requires standardised end-to-end integrated tool sets on cloud; wider adoption; more interpretable and explainable AI; continuous performance monitoring of AI models; and an integrated operation to make it successful.
Navine mentioned how an online footwear fashion house got 700,000 odd SKUs processed across three major product categories in just about nine weeks with improved productivity by 70 percent.
Below are the details on how they used the 3 building blocks of accessible AI.
- AI ready no-code or low-code cloud platform to do data prep, models available to be configured, trained, deployed and monitored all in one system.
- AI trained multidisciplinary ecosystem – augment the workforce where they had people to validate, annotate the outputs, and help in actual active learning of the models.
- Integrate AI Ops, Data Ops, ML Ops and Decision Ops.
Upshots
- Improved productivity by 70% where they managed to process 700,000 SKUs across 3 major product categories in 9 weeks against usual 30 weeks.
- The team had 6 members with no data science knowledge.
- Team could build their own computer vision models, validate outputs and re-train models to get the output
- Latest algorithms available in the production
- Quality and standardisation of output was assured
- Centralised latest ML pipeline available for other delivery streams for similar use cases, multiplying the impact achieved in the first delivery
The 3 building blocks have better success if technology, people and process are in alignment.
Takeaways
- Mature AI-ready no code or low-code cloud platform- scalable for users and use cases.
- Augmented AI-trained multidisciplinary ecosystem – to scale deliverables and enable faster business adoption.
- Integrated AI Ops process – to facilitate standardisation and automation.
Accessible AI maturity roadmap
- Basic AI: Early stage pilot – integrated data.
- Tactical AI: short term, non critical AI in place – advanced pilots – data apps in place.
- Strategic AI: AI process and use cases partly automated by platform – ML data pipelines – Mature data ops and ML ops process.
- Transformational AI: Mature AI platform – AI people ecosystem in place – Mature and Integrated OPS (data ops, ML Ops, ML Ops)- Decision taken based on AI Recommendations.
- Optimise AI implementations: Measuring the success defined and tracking the impact and constantly measuring the ROI.