Reasoning About Machine Intuition

This talk discusses the resurgence of Machine Learning and neural networks from multiple digital perspectives, including:

• product & design,
• iterative delivery,
• organisational design and knowledge management, and
• governance and risk.

I chose to use “intuition” to distinguish ML’s capability for pattern recognition in narrow tasks from other descriptions of intelligence. Find slides here.

Outline

Machine intuition, ftw

• Performance leaps in translation, autonomous vehicles, and negotiations
• Besting humans in video games, go and poker

What just happened?

• Neural networks, how they work, and their history
• Recent confluence of factors:
  • Data
  • Compute
  • Network designs

So, what’s next?

• Explosion of micro-intuitors
• Humans and machines complementing each other
  • Humans: fast learners and flexible thinkers
  • Machines: wide learners and scalable thinkers
• Distributed machine intuition
  • Edge inference for speed, efficiency and privacy
  • Federated cloud training for scale and privacy
• Pro-active recognition of risks
• More decision-making tasks outsourced, like more computational tasks have already been outsourced

Designing products with intuition

• Customer perception and trust (Ellen Enkel)
  • Operational safety & data security as a hygiene factor
  • An application that makes sense, it usable, and testable
  • Autonomy & control, pro-active communication, gradual introduction
• Take a human-centred approach (Google PAIR)
  • Machine learning doesn’t solve everything
  • Prototype with real people instead of an algorithm
  • Design with the system’s failure in mind
  • Get feedback, forever
• Choose the right problems – various canvas techniques for elaboration
• Consider structuring business scenarios as gameplay for machines

Developing technology with intuition

• Avoid simply stirring the pile of linear algebra
• The new new new product development game
  • Inner verification loop, aka building the thing right
    • From analysis driving coding to satisfy tests
    • To curated data driving model learning to satisfy metrics
  • Wrapped in the good old fashioned outer validation loop, aka building the right thing with customer feedback
• Data curation considerations
• Model training considerations
• Other success factors:
  • Diverse teams
  • Role of data science
  • Architecture
  • Continuous delivery
  • Choice of tooling
• A good task to have?

Building organisations with intuition

• Introduce a basic knowledge classification system
  • Know what – explicit, rule-based
  • Know how – tacit, feel-based
  • Know why – explicit, cause-based
• Introduce a simplified Nonaka cycle of organisational learning with these types knowledge – repeatedly applying the what leads to deep knowledge of the how which provides the insight for the why, and the next round of what
• Now add machine “know how” alongside human “know how”
• If we rely too much on machine “know how”, we lose the human “know how”
  • Then lose the human “know why”
  • Then lose the improved human “know what”
• So we must deliberately design for learning when deploying machine intuition in organisations
• And focus on augmenting humans (or complementing, as above) – machines bettering humans, not just besting humans
• Consideration of changing job design, from cognitive hammers to cognitive nail guns

Managing risk & ethics with intuition

• Some machine failure modes
  • Training set bias
  • Spurious correlations
  • Compare to black swans and pareidolia in humans
• Opaque box nature of decisions
  • “Know how” but not “know why”
• Potential societal implications (SciAm)
  • Democracy 2.0, or
  • Feudalism 2.0
• Some responses
  • Weapons of Math Destruction
    • Align with customer objectives
    • Avoid or reduce opaque models
    • Don’t naively port solutions between applications
    • Avoid applications that create their own reality through feedback loops
  • GDPR right to explanation
  • Corporate mission and values aligned position on avoiding harm

Conclusion

• Machines outperforming humans in narrow tasks
• Due to a confluence of recent developments; innovation continues
• Potential improvements have significant implications for product development and organisational design
• Potentially huge benefits for society, but risks to be managed
• Better understanding of machine intuition is the key to all this