{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "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": 185,
  "pageCount": 213,
  "prevPage": 184,
  "nextPage": 186,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 0,
  "notes": "The slide describes a specific technical method (CrossMax) to improve model robustness.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/185",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-185",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The resulting model gains a considerable adversarial robustness without any adversarial training or extra data.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d20-a484e197847a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.3,
        "w": 0.45,
        "x": 0.52,
        "y": 0.5
      },
      "kind": "image",
      "text": "Diagram showing the CrossMax process: input image -> multi-resolution expansion -> standard classifier backbone -> CrossMax top-k ensembling.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e4efd8a1-b145-4e4e-802d-d0e069201f37",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.65,
        "w": 0.45,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "list",
      "text": "To improve the robustness of image classifiers to adversarial attack, a Google DeepMind team drew inspiration from biological visual systems, specifically the concept of microsaccades (small, involuntary eye movements).\nThey feed the model multiple smaller, slightly blurrier versions of the same image. This improves robustness without needing special training.\nCrossMax Ensembling combines predictions from different layers of the model.\nEven if an adversarial attack confuses the final output, the predictions from earlier layers are often still accurate. By combining these, the model becomes stronger against attacks.\nThe proposed method achieves state-of-the-art (SOTA) adversarial accuracy on the CIFAR-10 and CIFAR-100 datasets without adversarial training.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "66445bc8-1fe0-41ae-83a5-c56cd8dc2245",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.53,
        "x": 0.44,
        "y": 0.81
      },
      "kind": "source-note",
      "text": "Figure 2 | Combining channel-wise stacked augmented and down-sampled versions of the input image with robust intermediate layer class predictions via CrossMax self-ensemble. The resulting model gains a considerable adversarial robustness without any adversarial training or extra data.",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d0e65137-13f1-40dc-9a8e-a70895f9240a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.05,
        "y": 0.13
      },
      "kind": "title",
      "text": "It’s not just foundation models that face adversarial attacks",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "26850988-76be-4fbd-8e32-b1d860a48cc6",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 204,
      "from": 154,
      "beatId": "019dd95a-0682-776c-8e35-605159a069e6",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore consequences and risks",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 204,
      "from": 175,
      "beatId": "019dd95a-0682-776c-8e35-74f05e198faf",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Return",
      "beatSlug": "voyage-return-the-return",
      "evidence": "Safety section returns to risks/governance debate",
      "position": 4,
      "confidence": 55,
      "parentBeatName": "Resolution",
      "parentBeatSlug": "resolution"
    }
  ],
  "loops": [
    {
      "to": 186,
      "from": 182,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-76aa50b143e8",
      "evidence": "Attack surface, Pliny red team, LLM-on-LLM patching, adversarial robustness, stealth attacks",
      "position": 22,
      "objective": "Show jailbreaking arms race across attack surfaces",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}