visual representation of something” Process of identifying and detecting an object or a feature in a digital image or video For humans vision seems easy, our brains are incredibly good at understanding images
easy task, algorithmical techniques, this is not machine learning (not too much intelligence) Object class recognition, scene understanding/interpretation: hard task, advanced machine learning techniques needed Different levels of difficulty
copies of the same neuron "...a computing system made up of a number of simple, highly interconnected processing elements, which process information by their dynamic state response to external inputs”1 1. "Neural Network Primer: Part I" by Maureen Caudill, AI Expert, Feb. 1989 TL;DR A Basic Introduction To Neural Networks: http://pages.cs.wisc.edu/~bolo/shipyard/neural/local.html TL;DR Conv Nets: A Modular Perspective http://colah.github.io/posts/2014-07-Conv-Nets-Modular/ TL;DR Using neural nets to recognize handwritten digits http://neuralnetworksanddeeplearning.com/chap1.html Deep CNN: https://www.tensorflow.org/tutorials/deep_cnn Convolutional neural network
CNN 1990’s • read zip codes, digits • AlexNet • won the ImageNet ILSVRC challenge in 2012 • top 5 error of 16% • ZF Net • won the ImageNet ILSVRC challenge in 2013 • improvement on AlexNet by tweaking the architecture hyperparameters • GoogLeNet (aka Inception) • won the ImageNet ILSVRC challenge in 2014 • Developed at Google • dramatically reduced the number of parameters in the network (4M, compared to 60M AlexNet ) TL;DR Convolutional Neural Networks for Visual Recognition: http://cs231n.github.io/convolutional-networks/#case Justin Johnson Stanford Vision Lab - http://cs.stanford.edu/people/jcjohns/ Andrej Karpathy Research Scientist at OpenAI http://cs.stanford.edu/people/karpathy/
an academic benchmark for computer vision. Inception-v3 reaches 3.46% 1 (top-5 error rate) 1. Rethinking the Inception Architecture for Computer Vision https://arxiv.org/pdf/1512.00567.pdf TL;DR Inception in Tensorflow: https://github.com/tensorflow/models/tree/master/inception ImageNet http://image-net.org/ How well do humans do on ImageNet Challenge? http://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/ Inception Model - convolutional neural network (CNN)
Based on data flow graphs: • Nodes: mathematical operations • Graph edes: data arrays (tensors) communicated between them • Flexible architecture, deploy computation to one or more CPUs or GPUs • APIs for constructing and executing a TensorFlow graph: • Python • C++ • Java • Go • bindings for: C#, Haskell, Julia, Ruby, Rust Tensorflow - machine learning and deep neural network library
model It runs inference on an input JPEG image. It outputs the top 5 predictions along with their accuracy probabilities. Source Code: https://github.com/tensorflow/models/blob/master/tutorials/image/imagenet/classify_image.py Image Recognition
2 weeks with 8 GPUs NVIDIA Tesla K40s ) • Splits the batch across the GPUs • Updates model parameters synchronously by waiting for all GPUs to finish processing a batch of data How to train from scratch: https://github.com/tensorflow/models/blob/master/inception/README.md#how-to-train-from-scratch Train your own image classifier: https://research.googleblog.com/2016/03/train-your-own-image-classifier-with.html Codelabs by Google: https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0 Train from Scratch
? • You Know, For Search! • Fast and Scalable • Nested Query - wildcard Why AWS S3 (or similar object storage)? • 99.99…% durability • Scalable • Up to 5TB for bucket • Control access • Low Price
• Smart Photo Gallery: find things in your photos • Product Catalog: automatic classification of products • Content Moderation: get notified of NSFW contents • Accessibility camera app: help people with disability to see and take pictures