# Niki Parmar > ~1986– · Computer Scientist, Transformer Co-author > > **Recorded contribution:** Transformer co-author; image Transformer ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Computer scientist Niki Parmar was one of eight authors of the Transformer paper and later first-authored Image Transformer, extending attention-based autoregressive modeling to images. She co-founded Adept AI with Ashish Vaswani and others. Public biographical data is sparse, so this profile makes no unsupported claim about her birth year and anchors her contribution in inspectable papers. Parmar's Transformer co-authorship belongs beside her later Image Transformer work, which showed that the same autoregressive attention principle could model image pixels or local image neighborhoods. This extended attention beyond text while exposing the severe sequence-length cost of treating images token by token. ## 2. The problem inherited Sequence models needed parallel long-range interaction, and image generation needed to model dependencies among many pixels without relying only on local convolutions. ## 3. The central contribution Parmar co-created the Transformer and led Image Transformer research applying self-attention to conditional and unconditional image generation. ## 4. Reconstruct the mechanism 1. Represent sequence elements or image pixels/patches as ordered vectors. 2. Use masked self-attention to condition each predicted element on permitted prior context. 3. Combine several attention heads with feed-forward transformations and residual paths. 4. Train an autoregressive likelihood and sample outputs one position at a time. ## 5. What changed downstream - The Transformer established a general attention-based architecture across modalities. - Image Transformer helped demonstrate that self-attention could model visual dependencies beyond natural-language sequences. - Image Transformer helped establish attention as a cross-modal modeling primitive and anticipated later vision Transformers and autoregressive image generators that changed tokenization and scale trade-offs. ## 6. Attribution, limits, and uncertainty - Both contributions are multi-author results and should not be rewritten as individual sole invention. - The registry's approximate birth year is not verified by these sources; autoregressive image modeling is computationally expensive and later methods followed different trade-offs. - Pixel-level likelihood can devote capacity to imperceptible detail, sequential sampling is slow, and the approximate registry birth year lacks support in the cited technical record. ## 7. Reconstruction lab Flatten a four-by-four binary image into a sequence and train a masked next-pixel model. Compare raster and block ordering to expose how representation changes the learned dependency. Compare raster and local-block ordering on a tiny image and measure how each ordering changes visible context, dependency length, and generation latency. ## 8. Evidence trail - [Attention Is All You Need](https://arxiv.org/abs/1706.03762) — NeurIPS - [Image Transformer](https://arxiv.org/abs/1802.05751) — ICML / arXiv - [Transformer proceedings record](https://papers.nips.cc/paper/7181-attention-is-all-you-need) — NeurIPS --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*