1 of 29

AI від Apple

Як Apple розкриває потенціал AI: Секрети Apple Intelligence та ключові дослідження

2 of 29

Hello

there!

Maksym Kmet

Senior Data Scientist

3 of 29

We befriend people with technology

4 of 29

$9M+

З 24 лютого 2022 року задонатили на підтримку України

30M+

Понад 30 мільйонів користувачів наших продуктів

180+

Країн, в яких живуть наші користувачі

5

Кожен п’ятий Mac на планеті має хоча б один продукт MacPaw

2008

Рік заснування компанії

5 of 29

ML Research

ML Frameworks

Advanced

Hardware

Як Apple стає AI-first company

6 of 29

Браковані NVIDIA GPU у новому Macbook та судові позови

Apple AI:

Передумови

2008

2016

2017

2020

2024

Apple Intelligence локально на iPhone 15 Pro

Apple Silicon (M1)

Neural engine для прискорення ML inference

GPU від AMD у Macbook

7 of 29

ML research

  • REALM
  • FerretUI
  • Open-ELM
  • DCLM-7B

А що крім заліза?

Frameworks for ML

  • CoreML
  • MLX
  • AxLearn

8 of 29

ML Frameworks

9 of 29

build and train a model with the Create ML app

use Core ML Tools to convert models built with other libraries into the Core ML

fine-tune on-device with users data

ML framework used across Apple products for fast inference with easy integration of pre-trained models on the edge

CoreML (2017)

10 of 29

build and train a model with the Create ML app

use Core ML Tools to convert models built with other libraries into the Core ML

fine-tune on-device with users data

ML framework used across Apple products for fast inference with easy integration of pre-trained models on the edge

CoreML (2017)

11 of 29

MLX is an array framework optimized for the unified memory architecture of Apple silicon. The NumPy-like API makes it familiar to use and flexible.

MLX (2024)

12 of 29

MLX is an array framework optimized for the unified memory architecture of Apple silicon. The NumPy-like API makes it familiar to use and flexible.

MLX (2024)

13 of 29

AXLearn is a library built on top of JAX and XLA to support the development of large-scale deep learning models

  • Training of models with up to hundreds of billions of parameters across thousands of accelerators
  • Apple Intelligence foundation models trained with AxLearn

AxLearn (2024)

14 of 29

ML

Research

15 of 29

On-screen Entities (ex. phone number on a restaurant webpage)

Background Entities (ex. alarm ringing or music in the background)

Conversational Entities (entities from the conversation with user (ex “Call Mom” or “Show path to the first restaurant from the list”)

Reference resolution as language modeling.

Understand what user is referring to in the conversation even if it is not explicitly stated.

Small model Flan-T5 matches GPT-4 on benchmarks (their own)

ReALM (2024)

16 of 29

Elementary Tasks:

Referring:

  • OCR
  • Icon recognition
  • Widget classification

Grounding:

  • Widget listing
  • Find text
  • Find icon
  • Find widget

Advanced Tasks:

  • Detailed description
  • Conversation perception
  • Conversation interaction
  • Function inference

Multimodal large language model for enhanced understanding of mobile UI screens. (encoder CLIP-ViT-L + Vicuna + projection layers + �visual sampler)

Ferret-UI (2024)

17 of 29

Open-ELM (May 2024)

  • Family of small open source LLM 270M-3B parameters (good for mobile)
  • Underwhelming performance compared to competitors

DCLM-7B (July 2024)

  • Relatively small open source LLM
  • Beats Mistral 7B in some of the benchmarks and gets close of Llama 3 and Gemma
  • Positively perceived in the community

Large Language Models

Open-ELM and DCLM-7B

18 of 29

How Apple �Intelligence works

19 of 29

Were our guesses correct?

Apple Intelligence

Features

20 of 29

  • On-screen context awareness
  • Taking actions
  • Reference resolution from complex phrases
  • Background context: emails, photos etc

…sounds like ReALM or Ferret-UI

Context Awareness

21 of 29

  • Writing tools: emails summarization, proofreading, rewriting
  • Notifications prioritization
  • Smart replies
  • All working locally

…sounds like Open-ELM

Language generation

22 of 29

  • Custom avatars generation
  • Images from description
  • Photo “clean up” function

Image generation

23 of 29

What we actually �know

24 of 29

Architecture

of Apple Intelligence

25 of 29

Architecture

of Apple Intelligence

26 of 29

On-device models: foundation models (3B params) with a set of small LoRA adapters

App Intents ToolBox: predefined functions such as Open Email, Start playing music, create a new Reminder. Each LoRa adapter is for a specific type of intents

Server models: The stack of Apple models in the Private Cloud

Cloud Compute: larger foundation LLM + LoRa

Orchestration: routes user request �to on-device model/ server /ChatGPT

Semantic index: is a vector database �of personal context (emails, photos, contacts …)

Architecture

of Apple Intelligence

27 of 29

described a photo of my beautiful baby as “group of people in a room with white walls.”

photo “clean up” function made me look bald

misinterprets the meanings of text, inverts people’s names

lifted to priority a scam email about suspended account

from a column by The Washington Post

Apple Intelligence can be kind of dumb

28 of 29

  • Foundation models
  • Adapters
  • Orchestration
  • Personal context database
  • Local vs server vs ChatGPT

  • Core ML
  • MLX

  • ReaLM
  • FerretUI
  • Open-ELM
  • DCLM-7B

Apple Intelligence

Brief history �of Apple computational powers

ML Frameworks

ML Research

29 of 29

Q&A