{
  "docId": "019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "docSlug": "cc8183ec02431b7a",
  "documentTitle": "2022 Air Street Capital The State of AI Report 2022",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 39,
  "pageCount": 114,
  "prevPage": 38,
  "nextPage": 40,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 1,
  "notes": "Discusses the technical approach of masking large patches of pixels to train Vision Transformers (ViT).",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc/39",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc#slide-39",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Self-supervision techniques used to train transformers on text are now transposed almost as is to images and are achieving state of the art results on ImageNet-1K.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a217-8e00dfda792f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.3,
        "x": 0.65,
        "y": 0.315
      },
      "kind": "image",
      "text": "Diagram showing input image, encoder, decoder, and target output.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "45ac243d-7737-4933-9917-0440a47574cb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "Self-supervision techniques used to train transformers on text are now transposed almost as is to images and are achieving state of the art results on ImageNet-1K.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "121ad688-395d-4272-9696-cbe7291e30c1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.6,
        "x": 0.02,
        "y": 0.32
      },
      "kind": "list",
      "text": "Transformer-based autoencoder LMs are trained to predict randomly masked words in large text corpora. This results in powerful models that are SOTA in language modeling tasks (e.g. BERT).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "76aafda1-32bd-4719-86a2-3d811bd7667e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.6,
        "x": 0.02,
        "y": 0.46
      },
      "kind": "list",
      "text": "While masking a word in a sentence makes the sentence nonsensical and creates a challenging task for LMs, reconstructing a few randomly masked pixels in images is trivial thanks to neighbouring pixels.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "77bc2bef-56dc-48a5-83b0-e137cc0a5176",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.95,
        "x": 0.02,
        "y": 0.74
      },
      "kind": "list",
      "text": "Self-supervised learning isn't new to computer vision. But this work is further evidence that SOTA techniques in language transition seamlessly vision. Can domains unification be pushed further?",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7afbdc65-4d0b-4b3c-9253-74bd902f6e1c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.6,
        "x": 0.02,
        "y": 0.6
      },
      "kind": "list",
      "text": "The solution: mask large patches of pixels (e.g. 75% of the pixels). Meta use this and other adjustments to pre-train a ViT-Huge model on ImageNet-1K and then fine-tune it to achieve a task-best 87.8% top-1 accuracy.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "88453179-95a2-4e5f-9d92-446ea2c1dd3a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "top-1 accuracy: 87.8%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a217-907c1ba865ef",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.8,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "...but the inevitable vision and language modeling unification continues...",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "deed9f40-4a6a-4187-942a-e42df9a15325",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Analytical method",
      "slug": "analytical-method",
      "agent": null,
      "layer": "slide",
      "matchId": "a296fcee-cb55-46dd-bf3b-7ebc9a278add",
      "evidence": "The slide presents a technical analysis of self-supervised learning techniques used in vision and language modeling.",
      "confidence": 0.7
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 80,
      "from": 5,
      "beatId": "019dd95a-0682-776c-8e35-2557e3799e96",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Definitions, exec summary, Research and Industry sections inventory the state of AI.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 93,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-35abf646aa8b",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Research, Industry, Politics sections accumulate signals.",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 43,
      "from": 37,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-afe811e71ad6",
      "evidence": "SayCan, ConvNets vs ViT, vision-language unification, Gato, IRIS, NeRFs as parallel evidence.",
      "position": 6,
      "objective": "Show transformers becoming the universal cross-modality architecture",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 78,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}