Home » Technical Topics » Machine Learning

ML and Causality – Why?


Machine learning is €œCompetence without Comprehension€ as famously noted by Dan Dennett, the pre-eminent philosopher of our times. There are two aspects to Machine Learning (ML) €œcomprehension€.

Artificial General Intelligence (AGI) hopes to infuse ML with comprehension. The other less lofty aspect is that WE would like to €œcomprehend€ how ML reaches its decisions and predictions! To accomplish the latter, we need €œExplainable ML€ €“ explanation is the evidence of comprehension . . .

Causation is the most important connection in the Universe €“ why? The cause-effect relationship among different entities alone provides the invariant basis for the explanation of the causal chain that leads to a decision or a prediction. ML may be able to pull together information using a deep neural network from, say, various vital signs of a patient and provide a decision (€œmove to ICU now€) or prediction (€œoxygenation level will be below threshold in 2 hours€). But when ML cannot provide the explanations, physicians cannot perform €œwhat-if€ and counterfactual thought-experiments necessary to tease out the root causes.

The conundrum in ML is that if predictions based on correlations are accurate €œenough€ and applies to related cases (generalizable) €œmost of the time€ for practical use, why bother about WHY ML worked? The problem is €œenough€ and €œmost of the time€; because they are probabilistic statements, was ML€™s effectiveness a €œchance event€ or was there an enduring basis for us to believe that it will work in other cases €“ in other words, what is the whole €œsolution space€?!

When ML is not explainable, it is hard nigh impossible to know the perimeter of the Solution Space; when we cannot surmise the extend of this space, we feel antsy! We need to comprehend the solution space coverage and reasons for it €“ Explanation is the evidence that we have comprehended the competence of our ML!

Explanation needs knowledge of cause-effect relationships . . .

CALLING on Data Scientists to incorporate Causality into your algorithms in your application domain! Here are my recent attempts to incorporate Causality into IoT –

(A) €œEvidence-based Prescriptive Analytics, CAUSAL Digital Twin and a Learning Estimation Algorithm€, April 2021 at https://arxiv.org/abs/2104.05828

  • Causality & Causal Graph; Conditions for a DAG to be a Causal Graph; Causal Discovery & Causal Estimation – Causal Graph in IoT use cases can be elicited from domain-experts
  • Non-traditional Neural Network algorithm; Causal Graph integrated into NN; Full derivation in the appendix

(B) €œStochastic Formulation of Causal Digital Twin €“ Kalman Filter Algorithm€, May 2021 at https://arxiv.org/abs/2105.05236

  • Stochastic formulation; SVAR model recast as State-space model
  • Kalman Filter & Smoother algorithm to estimate causal factors

Dr. PG Madhavan

Seattle, WA