Universal Machine Learning

Abstract

Deep learning approaches a vision problem the way a person does – trained on examples rather than programmed with rules – and Tom Hess sets that against manual inspection point by point: fast enough to keep pace with in-line production, consistent across every shift and every line, and improving as more examples arrive after deployment. The documented output is what makes it defensible: yield improvement through real-time process control, analytics for offline improvement work, external reporting to the customer, and traceability for failure analysis later. A pump assembly case study shows the labelling and marking workflow, where the operator supplies the labels and the system returns location, classification and match scores without a vision expert in the loop.