Prof. Tom Yeh Explains Two-Tower MLP by Hand, From Recommenders to CLIP

ProfTomYeh · x · 2026-09-14

In the 11th installment of his "AI by Hand" series, Prof. Tom Yeh breaks down the Two-Tower MLP with an interactive hand-drawn diagram: two inputs pass through independent MLPs whose outputs are concatenated and linearly projected. The pattern powers recommendation systems (user/item towers) and contrastive learning, and its most far-reaching use is multimodal—one tower reads images, the other text, and training pulls matching pairs into one shared space. That's exactly how CLIP works, reducing image-text matching to simple vector arithmetic. Interactive diagram available on byhand.ai.

Original post →

More from Research

Research channel →