hiejulia / Deep-Learning

Deep learning

DEEP-LEARNING

• Generate song lyrics using RNN
• Browser AI : running AI within browser
• Image classifier
• MNIST digit recognition
• XOR gate operation
• Music recommendation system with DL
• NLP with BERT
• Plate detection
• Optical character recognition
• Self driving car
• Action recognition

DL algorithms

• CNN
• RNN
• LSTM
• Transfer learning
• Siamese network
• Activation function
  • tanh
  • signmoid
  • ReLU
  • Leaky ReLu
  • Swish
  • softmax
  • exponential linear unit function

Data

• audio, image
• text processing

Model architecture

• image classificaion : ResNet, MobileNet
• text classification : BERT, XLNet
• image segmentation : Deep Lab, UNet
• image translation : Pix2Pix
• object detection : YOLO, Mask R CNN
• Speech generation : WaveNet

DL on the cloud

• Cluster management on Cloud

• AWS SageMaker

• AWS batch

• AWS Lambda, Farate

• Serverless for DL

• DL framework

  • automate training
  • allow export & import models
  • enable graph definition

• Train model using Tensorflow

  • MNIST dataset

• pretrained model

  • import/ export
  • deploy: preprocess data

• TF repo:

  • image classification
  • deep speech
  • lm2txt
  • https://github.com/tensorflow/models
  • text summarization
  • vid2depth

• import pretrained CNN model

  • inception-v3
    • 3.46% error rate

• Train DL model from Computer Vision Task

  • image classification task
  • Pretrained CNN model for Computer Vision Task
  • AWS SageMaker to train DL CNN model

• Image Data pipeline on cloud

  • AWS Fargate
  • AWS Step Function with AWS Fargate using Serverless Framework
  • DL training pipeline for the CNN

• DL training pipeline on the cloud

  • AWS lambda for DL inference
  • DL inference pipeline for CNN

• Microsoft custom vision.AI

• Google AutoML

• Clarifai

• IBM visual recognition

• Apple create ML

• NVIDIA Digits

• Amazon Rekognition

Distributed DL

• Scalable DL

DL production pipeline

• Deployment

Framework

• tensorflow, lite (mobile, embedded devices)
• pytorch
• MXNet
• CoreML

Tech stack

• Python

• TF / TF GPU

• Pillow

• Optional : NVIDIA GPU

  • install driver, CUDA, cuDNN, tensorflow-gpu
  • lambda stack : https://lambdalabs.com/lambda-stack-deep-learning-software

• sklearn

• https://keras.io/

hardware

• CPU

• GPU : large network

  • titanx

• FPGA

• ASIC

• TPU

• Edge TPU : wearable

• NPU : smartphone, digital signal processing

• Embedded AI projects (e.g., flood surveying drone, autonomous wheelchair) → Accelerators like Google Coral, Intel Movidius with Raspberry Pi, or GPUs like NVIDIA Jetson Nano, all the way down to $15 microcontrollers (MCUs) for wake word detection in smart speakers

• Cloud : Google cloud instane (n1 8 instance with NVIDIA Tesla K8 addon)

• AWS GPU instance - p2.xlarge instance

Image classification with Keras

• Improve accuracy from 90% to 97% by using transfer learning

image search

• feature extract
• search on image dataset (caltech)
• scale large dataset
• optimized system
• mainstream product
• Deep NN

Resource

• https://colab.research.google.com/github/googlecolab/colabtools/blob/master/notebooks/colab-github-demo.ipynb#scrollTo=WzIRIt9d2huC

• Open Neural Network Exchange (ONNX)

• https://runwayml.com/

• https://researchcode.com/

• https://arxiv.org/

• Case study

  • Flickr
    • similarity search feature based on same semantic meaening
    • ANN algo : Locally optimized product quantization
  • Pinterest
    • associated metadata
  • Spotify
    • NN music and creating automatic playlist and radia station
  • Image captioning feature in SeeingAI