eetop.cn_T9 - Generative AI on Edge Devices Models, Hardware and Systems.pdf

1、 2025 IEEE International Solid-State Circuits ConferenceGenerative AI on Edge Devices:Models,Hardware and SystemsPaul WhatmoughSenior Director,EngineeringQualcomm I 2025 IEEE International Solid-State Circuits ConferenceExamples of Discriminative ModelsGenerative AI on Edge Devices:Models,Hardware a

2、nd Systems2 of 56Paul WhatmoughMLPGeneral ClassificationCNNImage ClassificationRNNSpeech RecognitionInputsOutputsHidden LayerInputConvolutionPoolingFeature MapsInputsOutputsLong Short-Term Memory(LSTM)Cells 2025 IEEE International Solid-State Circuits ConferenceWhat is Generative AI?Generative AI on

3、 Edge Devices:Models,Hardware and Systems3 of 56Paul WhatmoughNeural Networks that learn a simplified representation of their training dataAble to create a new output that follows a similar distribution,conditioned on a promptArtificial IntelligenceMachine LearningDeep LearningGenerative AISaharia e

4、t al.,“Photorealistic text-to-image diffusion models with deep language understanding”,NeurIPS 22“A cute corgi lives in a house made of sushi.”2025 IEEE International Solid-State Circuits ConferenceDiscriminative and Generative ModelsDiscriminative ModelsLearns a boundary between classesMaximizes th

5、e conditional probability:P(Y|X)Directly estimates P(Y|X)Cannot generate new dataClassification tasks onlyGenerative ModelsLearns the input data distributionMaximizes the joint probability:P(X,Y)Estimates P(X|Y)to find P(Y|X)using Bayes ruleCan generate new dataGenerative AI on Edge Devices:Models,H

6、ardware and Systems4 of 56Paul Whatmough 2025 IEEE International Solid-State Circuits ConferenceExamples of Generative ModelsLarge Language Model(LLM)Text-to-TextLarge Vision Model(LVM)Text-to-imageLrg Multimodal Model(LMM)Text/Image-to-Text/ImagePaul WhatmoughGenerative AI on Edge Devices:Models,Ha