gaoyunquan Goto Github PK

awesome-interview

Collection of awesome interview references.

buptgraduatethesis

deeppavlov

An open source library for deep learning end-to-end dialog systems and chatbots.

dynamicvino

DynamicVINO is an open-source C++ library for establishing easy-to-use, extensible and scalable deep learning inference system for Intel CPU, GPU, NCS, FPGA based on Intel' s OpenVINO toolkit

edgetpu-facenet

Implement SSD and FaceNet on Edge TPU Accelerator

eigen

Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.

hand_detection

using Neural Networks (SSD) on Tensorflow. This repo documents steps and scripts used to train a hand detector using Tensorflow (Object Detection API). As with any DNN based task, the most expensive (and riskiest) part of the process has to do with finding or creating the right (annotated) dataset. I was interested mainly in detecting hands on a table (egocentric view point). I experimented first with the [Oxford Hands Dataset](http://www.robots.ox.ac.uk/~vgg/data/hands/) (the results were not good). I then tried the [Egohands Dataset](http://vision.soic.indiana.edu/projects/egohands/) which was a much better fit to my requirements. The goal of this repo/post is to demonstrate how neural networks can be applied to the (hard) problem of tracking hands (egocentric and other views). Better still, provide code that can be adapted to other uses cases. If you use this tutorial or models in your research or project, please cite [this](#citing-this-tutorial). Here is the detector in action. <img src="images/hand1.gif" width="33.3%"><img src="images/hand2.gif" width="33.3%"><img src="images/hand3.gif" width="33.3%"> Realtime detection on video stream from a webcam . <img src="images/chess1.gif" width="33.3%"><img src="images/chess2.gif" width="33.3%"><img src="images/chess3.gif" width="33.3%"> Detection on a Youtube video. Both examples above were run on a macbook pro **CPU** (i7, 2.5GHz, 16GB). Some fps numbers are: | FPS | Image Size | Device| Comments| | ------------- | ------------- | ------------- | ------------- | | 21 | 320 * 240 | Macbook pro (i7, 2.5GHz, 16GB) | Run without visualizing results| | 16 | 320 * 240 | Macbook pro (i7, 2.5GHz, 16GB) | Run while visualizing results (image above) | | 11 | 640 * 480 | Macbook pro (i7, 2.5GHz, 16GB) | Run while visualizing results (image above) | > Note: The code in this repo is written and tested with Tensorflow `1.4.0-rc0`. Using a different version may result in [some errors](https://github.com/tensorflow/models/issues/1581). You may need to [generate your own frozen model](https://pythonprogramming.net/testing-custom-object-detector-tensorflow-object-detection-api-tutorial/?completed=/training-custom-objects-tensorflow-object-detection-api-tutorial/) graph using the [model checkpoints](model-checkpoint) in the repo to fit your TF version. **Content of this document** - Motivation - Why Track/Detect hands with Neural Networks - Data preparation and network training in Tensorflow (Dataset, Import, Training) - Training the hand detection Model - Using the Detector to Detect/Track hands - Thoughts on Optimizations. > P.S if you are using or have used the models provided here, feel free to reach out on twitter ([@vykthur](https://twitter.com/vykthur)) and share your work! ## Motivation - Why Track/Detect hands with Neural Networks? There are several existing approaches to tracking hands in the computer vision domain. Incidentally, many of these approaches are rule based (e.g extracting background based on texture and boundary features, distinguishing between hands and background using color histograms and HOG classifiers,) making them not very robust. For example, these algorithms might get confused if the background is unusual or in situations where sharp changes in lighting conditions cause sharp changes in skin color or the tracked object becomes occluded.(see [here for a review](https://www.cse.unr.edu/~bebis/handposerev.pdf) paper on hand pose estimation from the HCI perspective) With sufficiently large datasets, neural networks provide opportunity to train models that perform well and address challenges of existing object tracking/detection algorithms - varied/poor lighting, noisy environments, diverse viewpoints and even occlusion. The main drawbacks to usage for real-time tracking/detection is that they can be complex, are relatively slow compared to tracking-only algorithms and it can be quite expensive to assemble a good dataset. But things are changing with advances in fast neural networks. Furthermore, this entire area of work has been made more approachable by deep learning frameworks (such as the tensorflow object detection api) that simplify the process of training a model for custom object detection. More importantly, the advent of fast neural network models like ssd, faster r-cnn, rfcn (see [here](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md#coco-trained-models-coco-models) ) etc make neural networks an attractive candidate for real-time detection (and tracking) applications. Hopefully, this repo demonstrates this. > If you are not interested in the process of training the detector, you can skip straight to applying the [pretrained model I provide in detecting hands](#detecting-hands). Training a model is a multi-stage process (assembling dataset, cleaning, splitting into training/test partitions and generating an inference graph). While I lightly touch on the details of these parts, there are a few other tutorials cover training a custom object detector using the tensorflow object detection api in more detail[ see [here](https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/) and [here](https://towardsdatascience.com/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9) ]. I recommend you walk through those if interested in training a custom object detector from scratch. ## Data preparation and network training in Tensorflow (Dataset, Import, Training) **The Egohands Dataset** The hand detector model is built using data from the [Egohands Dataset](http://vision.soic.indiana.edu/projects/egohands/) dataset. This dataset works well for several reasons. It contains high quality, pixel level annotations (>15000 ground truth labels) where hands are located across 4800 images. All images are captured from an egocentric view (Google glass) across 48 different environments (indoor, outdoor) and activities (playing cards, chess, jenga, solving puzzles etc). <img src="images/egohandstrain.jpg" width="100%"> If you will be using the Egohands dataset, you can cite them as follows: > Bambach, Sven, et al. "Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions." Proceedings of the IEEE International Conference on Computer Vision. 2015. The Egohands dataset (zip file with labelled data) contains 48 folders of locations where video data was collected (100 images per folder). ``` -- LOCATION_X -- frame_1.jpg -- frame_2.jpg ... -- frame_100.jpg -- polygons.mat // contains annotations for all 100 images in current folder -- LOCATION_Y -- frame_1.jpg -- frame_2.jpg ... -- frame_100.jpg -- polygons.mat // contains annotations for all 100 images in current folder ``` **Converting data to Tensorflow Format** Some initial work needs to be done to the Egohands dataset to transform it into the format (`tfrecord`) which Tensorflow needs to train a model. This repo contains `egohands_dataset_clean.py` a script that will help you generate these csv files. - Downloads the egohands datasets - Renames all files to include their directory names to ensure each filename is unique - Splits the dataset into train (80%), test (10%) and eval (10%) folders. - Reads in `polygons.mat` for each folder, generates bounding boxes and visualizes them to ensure correctness (see image above). - Once the script is done running, you should have an images folder containing three folders - train, test and eval. Each of these folders should also contain a csv label document each - `train_labels.csv`, `test_labels.csv` that can be used to generate `tfrecords` Note: While the egohands dataset provides four separate labels for hands (own left, own right, other left, and other right), for my purpose, I am only interested in the general `hand` class and label all training data as `hand`. You can modify the data prep script to generate `tfrecords` that support 4 labels. Next: convert your dataset + csv files to tfrecords. A helpful guide on this can be found [here](https://pythonprogramming.net/creating-tfrecord-files-tensorflow-object-detection-api-tutorial/).For each folder, you should be able to generate `train.record`, `test.record` required in the training process. ## Training the hand detection Model Now that the dataset has been assembled (and your tfrecords), the next task is to train a model based on this. With neural networks, it is possible to use a process called [transfer learning](https://www.tensorflow.org/tutorials/image_retraining) to shorten the amount of time needed to train the entire model. This means we can take an existing model (that has been trained well on a related domain (here image classification) and retrain its final layer(s) to detect hands for us. Sweet!. Given that neural networks sometimes have thousands or millions of parameters that can take weeks or months to train, transfer learning helps shorten training time to possibly hours. Tensorflow does offer a few models (in the tensorflow [model zoo](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md#coco-trained-models-coco-models)) and I chose to use the `ssd_mobilenet_v1_coco` model as my start point given it is currently (one of) the fastest models (read the SSD research [paper here](https://arxiv.org/pdf/1512.02325.pdf)). The training process can be done locally on your CPU machine which may take a while or better on a (cloud) GPU machine (which is what I did). For reference, training on my macbook pro (tensorflow compiled from source to take advantage of the mac's cpu architecture) the maximum speed I got was 5 seconds per step as opposed to the ~0.5 seconds per step I got with a GPU. For reference it would take about 12 days to run 200k steps on my mac (i7, 2.5GHz, 16GB) compared to ~5hrs on a GPU. > **Training on your own images**: Please use the [guide provided by Harrison from pythonprogramming](https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/) on how to generate tfrecords given your label csv files and your images. The guide also covers how to start the training process if training locally. [see [here] (https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/)]. If training in the cloud using a service like GCP, see the [guide here](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_on_cloud.md). As the training process progresses, the expectation is that total loss (errors) gets reduced to its possible minimum (about a value of 1 or thereabout). By observing the tensorboard graphs for total loss(see image below), it should be possible to get an idea of when the training process is complete (total loss does not decrease with further iterations/steps). I ran my training job for 200k steps (took about 5 hours) and stopped at a total Loss (errors) value of 2.575.(In retrospect, I could have stopped the training at about 50k steps and gotten a similar total loss value). With tensorflow, you can also run an evaluation concurrently that assesses your model to see how well it performs on the test data. A commonly used metric for performance is mean average precision (mAP) which is single number used to summarize the area under the precision-recall curve. mAP is a measure of how well the model generates a bounding box that has at least a 50% overlap with the ground truth bounding box in our test dataset. For the hand detector trained here, the mAP value was **[email protected]**. mAP values range from 0-1, the higher the better. <img src="images/accuracy.jpg" width="100%"> Once training is completed, the trained inference graph (`frozen_inference_graph.pb`) is then exported (see the earlier referenced guides for how to do this) and saved in the `hand_inference_graph` folder. Now its time to do some interesting detection. ## Using the Detector to Detect/Track hands If you have not done this yet, please following the guide on installing [Tensorflow and the Tensorflow object detection api](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md). This will walk you through setting up the tensorflow framework, cloning the tensorflow github repo and a guide on - Load the `frozen_inference_graph.pb` trained on the hands dataset as well as the corresponding label map. In this repo, this is done in the `utils/detector_utils.py` script by the `load_inference_graph` method. ```python detection_graph = tf.Graph() with detection_graph.as_default(): od_graph_def = tf.GraphDef() with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid: serialized_graph = fid.read() od_graph_def.ParseFromString(serialized_graph) tf.import_graph_def(od_graph_def, name='') sess = tf.Session(graph=detection_graph) print("> ====== Hand Inference graph loaded.") ``` - Detect hands. In this repo, this is done in the `utils/detector_utils.py` script by the `detect_objects` method. ```python (boxes, scores, classes, num) = sess.run( [detection_boxes, detection_scores, detection_classes, num_detections], feed_dict={image_tensor: image_np_expanded}) ``` - Visualize detected bounding detection_boxes. In this repo, this is done in the `utils/detector_utils.py` script by the `draw_box_on_image` method. This repo contains two scripts that tie all these steps together. - detect_multi_threaded.py : A threaded implementation for reading camera video input detection and detecting. Takes a set of command line flags to set parameters such as `--display` (visualize detections), image parameters `--width` and `--height`, videe `--source` (0 for camera) etc. - detect_single_threaded.py : Same as above, but single threaded. This script works for video files by setting the video source parameter videe `--source` (path to a video file). ```cmd # load and run detection on video at path "videos/chess.mov" python detect_single_threaded.py --source videos/chess.mov ``` > Update: If you do have errors loading the frozen inference graph in this repo, feel free to generate a new graph that fits your TF version from the model-checkpoint in this repo. Use the [export_inference_graph.py](https://github.com/tensorflow/models/blob/master/research/object_detection/export_inference_graph.py) script provided in the tensorflow object detection api repo. More guidance on this [here](https://pythonprogramming.net/testing-custom-object-detector-tensorflow-object-detection-api-tutorial/?completed=/training-custom-objects-tensorflow-object-detection-api-tutorial/). ## Thoughts on Optimization. A few things that led to noticeable performance increases. - Threading: Turns out that reading images from a webcam is a heavy I/O event and if run on the main application thread can slow down the program. I implemented some good ideas from [Adrian Rosebuck](https://www.pyimagesearch.com/2017/02/06/faster-video-file-fps-with-cv2-videocapture-and-opencv/) on parrallelizing image capture across multiple worker threads. This mostly led to an FPS increase of about 5 points. - For those new to Opencv, images from the `cv2.read()` method return images in [BGR format](https://www.learnopencv.com/why-does-opencv-use-bgr-color-format/). Ensure you convert to RGB before detection (accuracy will be much reduced if you dont). ```python cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB) ``` - Keeping your input image small will increase fps without any significant accuracy drop.(I used about 320 x 240 compared to the 1280 x 720 which my webcam provides). - Model Quantization. Moving from the current 32 bit to 8 bit can achieve up to 4x reduction in memory required to load and store models. One way to further speed up this model is to explore the use of [8-bit fixed point quantization](https://heartbeat.fritz.ai/8-bit-quantization-and-tensorflow-lite-speeding-up-mobile-inference-with-low-precision-a882dfcafbbd). Performance can also be increased by a clever combination of tracking algorithms with the already decent detection and this is something I am still experimenting with. Have ideas for optimizing better, please share! <img src="images/general.jpg" width="100%"> Note: The detector does reflect some limitations associated with the training set. This includes non-egocentric viewpoints, very noisy backgrounds (e.g in a sea of hands) and sometimes skin tone. There is opportunity to improve these with additional data. ## Integrating Multiple DNNs. One way to make things more interesting is to integrate our new knowledge of where "hands" are with other detectors trained to recognize other objects. Unfortunately, while our hand detector can in fact detect hands, it cannot detect other objects (a factor or how it is trained). To create a detector that classifies multiple different objects would mean a long involved process of assembling datasets for each class and a lengthy training process. > Given the above, a potential strategy is to explore structures that allow us **efficiently** interleave output form multiple pretrained models for various object classes and have them detect multiple objects on a single image. An example of this is with my primary use case where I am interested in understanding the position of objects on a table with respect to hands on same table. I am currently doing some work on a threaded application that loads multiple detectors and outputs bounding boxes on a single image. More on this soon.

hashingdeeplearning

Codebase for "SLIDE : In Defense of Smart Algorithms over Hardware Acceleration for Large-Scale Deep Learning Systems"

hostpital-on-palm

江苏省人民医院掌上医院客户端

image-classifier

Image Classifier Going forward, AI algorithms will be incorporated into more and more everyday applications. For example, you might want to include an image classifier in a smartphone app. To do this, you'd use a deep learning model trained on hundreds of thousands of images as part of the overall application architecture. A large part of software development in the future will be using these types of models as common parts of applications. In this project, you'll train an image classifier to recognize different species of flowers. You can imagine using something like this in a phone app that tells you the name of the flower your camera is looking at. In practice, you'd train this classifier, then export it for use in your application. We'll be using this dataset of 102 flower categories. When you've completed this project, you'll have an application that can be trained on any set of labelled images. Here your network will be learning about flowers and end up as a command line application. But, what you do with your new skills depends on your imagination and effort in building a dataset. This is the final Project of the Udacity AI with Python Nanodegree Prerequisites The Code is written in Python 3.6.5 . If you don't have Python installed you can find it here. If you are using a lower version of Python you can upgrade using the pip package, ensuring you have the latest version of pip. To install pip run in the command Line python -m ensurepip -- default-pip to upgrade it python -m pip install -- upgrade pip setuptools wheel to upgrade Python pip install python -- upgrade Additional Packages that are required are: Numpy, Pandas, MatplotLib, Pytorch, PIL and json. You can donwload them using pip pip install numpy pandas matplotlib pil or conda conda install numpy pandas matplotlib pil In order to intall Pytorch head over to the Pytorch site select your specs and follow the instructions given. Viewing the Jyputer Notebook In order to better view and work on the jupyter Notebook I encourage you to use nbviewer . You can simply copy and paste the link to this website and you will be able to edit it without any problem. Alternatively you can clone the repository using git clone https://github.com/fotisk07/Image-Classifier/ then in the command Line type, after you have downloaded jupyter notebook type jupyter notebook locate the notebook and run it. Command Line Application Train a new network on a data set with train.py Basic Usage : python train.py data_directory Prints out current epoch, training loss, validation loss, and validation accuracy as the netowrk trains Options: Set direcotry to save checkpoints: python train.py data_dor --save_dir save_directory Choose arcitecture (alexnet, densenet121 or vgg16 available): pytnon train.py data_dir --arch "vgg16" Set hyperparameters: python train.py data_dir --learning_rate 0.001 --hidden_layer1 120 --epochs 20 Use GPU for training: python train.py data_dir --gpu gpu Predict flower name from an image with predict.py along with the probability of that name. That is you'll pass in a single image /path/to/image and return the flower name and class probability Basic usage: python predict.py /path/to/image checkpoint Options: Return top K most likely classes: python predict.py input checkpoint ---top_k 3 Use a mapping of categories to real names: python predict.py input checkpoint --category_names cat_To_name.json Use GPU for inference: python predict.py input checkpoint --gpu Json file In order for the network to print out the name of the flower a .json file is required. If you aren't familiar with json you can find information here. By using a .json file the data can be sorted into folders with numbers and those numbers will correspond to specific names specified in the .json file. Data and the json file The data used specifically for this assignemnt are a flower database are not provided in the repository as it's larger than what github allows. Nevertheless, feel free to create your own databases and train the model on them to use with your own projects. The structure of your data should be the following: The data need to comprised of 3 folders, test, train and validate. Generally the proportions should be 70% training 10% validate and 20% test. Inside the train, test and validate folders there should be folders bearing a specific number which corresponds to a specific category, clarified in the json file. For example if we have the image a.jpj and it is a rose it could be in a path like this /test/5/a.jpg and json file would be like this {...5:"rose",...}. Make sure to include a lot of photos of your catagories (more than 10) with different angles and different lighting conditions in order for the network to generalize better. GPU As the network makes use of a sophisticated deep convolutional neural network the training process is impossible to be done by a common laptop. In order to train your models to your local machine you have three options Cuda -- If you have an NVIDIA GPU then you can install CUDA from here. With Cuda you will be able to train your model however the process will still be time consuming Cloud Services -- There are many paid cloud services that let you train your models like AWS or Google Cloud Coogle Colab -- Google Colab gives you free access to a tesla K80 GPU for 12 hours at a time. Once 12 hours have ellapsed you can just reload and continue! The only limitation is that you have to upload the data to Google Drive and if the dataset is massive you may run out of space. However, once a model is trained then a normal CPU can be used for the predict.py file and you will have an answer within some seconds. Hyperparameters As you can see you have a wide selection of hyperparameters available and you can get even more by making small modifications to the code. Thus it may seem overly complicated to choose the right ones especially if the training needs at least 15 minutes to be completed. So here are some hints: By increasing the number of epochs the accuracy of the network on the training set gets better and better however be careful because if you pick a large number of epochs the network won't generalize well, that is to say it will have high accuracy on the training image and low accuracy on the test images. Eg: training for 12 epochs training accuracy: 85% Test accuracy: 82%. Training for 30 epochs training accuracy 95% test accuracy 50%. A big learning rate guarantees that the network will converge fast to a small error but it will constantly overshot A small learning rate guarantees that the network will reach greater accuracies but the learning process will take longer Densenet121 works best for images but the training process takes significantly longer than alexnet or vgg16 *My settings were lr=0.001, dropoup=0.5, epochs= 15 and my test accuracy was 86% with densenet121 as my feature extraction model. Pre-Trained Network The checkpoint.pth file contains the information of a network trained to recognise 102 different species of flowers. I has been trained with specific hyperparameters thus if you don't set them right the network will fail. In order to have a prediction for an image located in the path /path/to/image using my pretrained model you can simply type python predict.py /path/to/image checkpoint.pth Contributing Please read CONTRIBUTING.md for the process for submitting pull requests. Authors Shanmukha Mudigonda - Initial work Udacity - Final Project of the AI with Python Nanodegree

inference_withtorchtensorrt

Comparing Deep Learning Inference of Pytorch models running on CPU, CUDA and TensorRT

intel-edge-ai-for-iot-developers-nanodegree-program

Leverage the Intel® Distribution of OpenVINO™ Toolkit to fast-track development of high-performance computer vision and deep learning inference applications, and run pre-trained deep learning models for computer vision on-premise. You will identify key hardware specifications of various hardware types (CPU, VPU, FPGA, and Integrated GPU), and utilize the Intel® DevCloud for the Edge to test model performance on the various hardware types. Finally, you will use software tools to optimize deep learning models to improve performance of Edge AI systems. - Source

iron-man-jarvis

Jarvis is a fully accelerated, application framework for building multimodal conversational AI services that use an end-to-end deep learning pipeline (Figure 1). The Jarvis framework includes pretrained conversational AI models, tools, and optimized end-to-end services for speech, vision, and NLU tasks. In addition to AI services, Jarvis enables you to fuse vision, audio, and other sensor inputs simultaneously to deliver capabilities such as multi-user, multi-context conversations in applications such as virtual assistants, multi-user diarization, and call center assistants. Using Jarvis, you can easily fine-tune state-of-art-models on your data to achieve a deeper understanding of their specific contexts. Optimize for inference to offer real-time services that run in 150 ms compared to the 25 seconds required on CPU-only platforms.

jetson-inference

Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.

jshtcm

江苏省中医院自动预约挂号

large-scale-bayesian-neural-networks-on-hybrid-cmos-reram-architecture-for-energy-efficient-and-rob

This work presents an algorithm-hardware co- simulation approach to implement Bayesian inference with large- scale Convolutional Neural Networks (CNNs) on hardware in a highly resource-efficient manner. We replace the end layers of the CNN, including the output softmax, with a Bayesian Neural Network (BNN). We train this hybrid CNN+BNN model on software and deploy it on a hybrid CMOS-ReRAM based hardware architecture in an area and energy-efficient manner. Our approach introduces a novel way for deep learning algo- rithms to efficiently take advantage of Bayesian inference without the overhead of training a convolutional Bayesian model end- to-end as well as hardware complexity. We carry out analyses on VGG19 and ResNet18 networks using CIFAR10/CIFAR100 datasets and obtain > 3× energy-efficiency by deploying a hybrid VGG19+BNN architecture on hybrid CMOS-ReRAM hardware than a full-scale VGG19 BNN assumed to be implemented on a fully ReRAM-based hardware platform. Furthermore, we also find that our proposed hybrid Bayesian architecture is robust against adversarial and noisy inputs, i.e. it has the ability to identify such inputs and refuses to classify them.

machine-learning-systems-design

A booklet on machine learning systems design with exercises: https://huyenchip.com/machine-learning-systems-design/toc.html

models

Model Zoo for Intel® Architecture: contains Intel optimizations for running deep learning workloads on Intel® Xeon® Scalable processors

neural2d

Neural net optimized for 2D image data

neurons

An ANN is a model based on a collection of connected units or nodes called "artificial neurons", which loosely model the neurons in a biological brain. Each connection, like the synapses in a biological brain, can transmit information, a "signal", from one artificial neuron to another. An artificial neuron that receives a signal can process it and then signal additional artificial neurons connected to it. In common ANN implementations, the signal at a connection between artificial neurons is a real number, and the output of each artificial neuron is computed by some non-linear function of the sum of its inputs. The connections between artificial neurons are called "edges". Artificial neurons and edges typically have a weight that adjusts as learning proceeds. The weight increases or decreases the strength of the signal at a connection. Artificial neurons may have a threshold such that the signal is only sent if the aggregate signal crosses that threshold. Typically, artificial neurons are aggregated into layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first layer (the input layer) to the last layer (the output layer), possibly after traversing the layers multiple times. The original goal of the ANN approach was to solve problems in the same way that a human brain would. However, over time, attention moved to performing specific tasks, leading to deviations from biology. Artificial neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games and medical diagnosis. Deep learning consists of multiple hidden layers in an artificial neural network. This approach tries to model the way the human brain processes light and sound into vision and hearing. Some successful applications of deep learning are computer vision and speech recognition.[68] Decision trees Main article: Decision tree learning Decision tree learning uses a decision tree as a predictive model to go from observations about an item (represented in the branches) to conclusions about the item's target value (represented in the leaves). It is one of the predictive modeling approaches used in statistics, data mining, and machine learning. Tree models where the target variable can take a discrete set of values are called classification trees; in these tree structures, leaves represent class labels and branches represent conjunctions of features that lead to those class labels. Decision trees where the target variable can take continuous values (typically real numbers) are called regression trees. In decision analysis, a decision tree can be used to visually and explicitly represent decisions and decision making. In data mining, a decision tree describes data, but the resulting classification tree can be an input for decision making. Support vector machines Main article: Support vector machines Support vector machines (SVMs), also known as support vector networks, are a set of related supervised learning methods used for classification and regression. Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that predicts whether a new example falls into one category or the other.[69] An SVM training algorithm is a non-probabilistic, binary, linear classifier, although methods such as Platt scaling exist to use SVM in a probabilistic classification setting. In addition to performing linear classification, SVMs can efficiently perform a non-linear classification using what is called the kernel trick, implicitly mapping their inputs into high-dimensional feature spaces. Illustration of linear regression on a data set. Regression analysis Main article: Regression analysis Regression analysis encompasses a large variety of statistical methods to estimate the relationship between input variables and their associated features. Its most common form is linear regression, where a single line is drawn to best fit the given data according to a mathematical criterion such as ordinary least squares. The latter is often extended by regularization (mathematics) methods to mitigate overfitting and bias, as in ridge regression. When dealing with non-linear problems, go-to models include polynomial regression (for example, used for trendline fitting in Microsoft Excel[70]), logistic regression (often used in statistical classification) or even kernel regression, which introduces non-linearity by taking advantage of the kernel trick to implicitly map input variables to higher-dimensional space. Bayesian networks Main article: Bayesian network A simple Bayesian network. Rain influences whether the sprinkler is activated, and both rain and the sprinkler influence whether the grass is wet. A Bayesian network, belief network, or directed acyclic graphical model is a probabilistic graphical model that represents a set of random variables and their conditional independence with a directed acyclic graph (DAG). For example, a Bayesian network could represent the probabilistic relationships between diseases and symptoms. Given symptoms, the network can be used to compute the probabilities of the presence of various diseases. Efficient algorithms exist that perform inference and learning. Bayesian networks that model sequences of variables, like speech signals or protein sequences, are called dynamic Bayesian networks. Generalizations of Bayesian networks that can represent and solve decision problems under uncertainty are called influence diagrams. Genetic algorithms Main article: Genetic algorithm A genetic algorithm (GA) is a search algorithm and heuristic technique that mimics the process of natural selection, using methods such as mutation and crossover to generate new genotypes in the hope of finding good solutions to a given problem. In machine learning, genetic algorithms were used in the 1980s and 1990s.[71][72] Conversely, machine learning techniques have been used to improve the performance of genetic and evolutionary algorithms.[73] Training models Usually, machine learning models require a lot of data in order for them to perform well. Usually, when training a machine learning model, one needs to collect a large, representative sample of data from a training set. Data from the training set can be as varied as a corpus of text, a collection of images, and data collected from individual users of a service. Overfitting is something to watch out for when training a machine learning model. Federated learning Main article: Federated learning Federated learning is an adapted form of distributed artificial intelligence to training machine learning models that decentralizes the training process, allowing for users' privacy to be maintained by not needing to send their data to a centralized server. This also increases efficiency by decentralizing the training process to many devices. For example, Gboard uses federated machine learning to train search query prediction models on users' mobile phones without having to send individual searches back to Google.[74] Applications There are many applications for machine learning, including: Agriculture Anatomy Adaptive websites Affective computing Banking Bioinformatics Brain–machine interfaces Cheminformatics Citizen science Computer networks Computer vision Credit-card fraud detection Data quality DNA sequence classification Economics Financial market analysis[75] General game playing Handwriting recognition Information retrieval Insurance Internet fraud detection Linguistics Machine learning control Machine perception Machine translation Marketing Medical diagnosis Natural language processing Natural language understanding Online advertising Optimization Recommender systems Robot locomotion Search engines Sentiment analysis Sequence mining Software engineering Speech recognition Structural health monitoring Syntactic pattern recognition Telecommunication Theorem proving Time series forecasting User behavior analytics In 2006, the media-services provider Netflix held the first "Netflix Prize" competition to find a program to better predict user preferences and improve the accuracy of its existing Cinematch movie recommendation algorithm by at least 10%. A joint team made up of researchers from AT&T Labs-Research in collaboration with the teams Big Chaos and Pragmatic Theory built an ensemble model to win the Grand Prize in 2009 for $1 million.[76] Shortly after the prize was awarded, Netflix realized that viewers' ratings were not the best indicators of their viewing patterns ("everything is a recommendation") and they changed their recommendation engine accordingly.[77] In 2010 The Wall Street Journal wrote about the firm Rebellion Research and their use of machine learning to predict the financial crisis.[78] In 2012, co-founder of Sun Microsystems, Vinod Khosla, predicted that 80% of medical doctors' jobs would be lost in the next two decades to automated machine learning medical diagnostic software.[79] In 2014, it was reported that a machine learning algorithm had been applied in the field of art history to study fine art paintings and that it may have revealed previously unrecognized influences among artists.[80] In 2019 Springer Nature published the first research book created using machine learning.[81] Limitations Although machine learning has been transformative in some fields, machine-learning programs often fail to deliver expected results.[82][83][84] Reasons for this are numerous: lack of (suitable) data, lack of access to the data, data bias, privacy problems, badly chosen tasks and algorithms, wrong tools and people, lack of resources, and evaluation problems.[85] In 2018, a self-driving car from Uber failed to detect a pedestrian, who was killed after a collision.[86] Attempts to use machine learning in healthcare with the IBM Watson system failed to deliver even after years of time and billions of dollars invested.[87][88] Bias Main article: Algorithmic bias Machine learning approaches in particular can suffer from different data biases. A machine learning system trained on current customers only may not be able to predict the needs of new customer groups that are not represented in the training data. When trained on man-made data, machine learning is likely to pick up the same constitutional and unconscious biases already present in society.[89] Language models learned from data have been shown to contain human-like biases.[90][91] Machine learning systems used for criminal risk assessment have been found to be biased against black people.[92][93] In 2015, Google photos would often tag black people as gorillas,[94] and in 2018 this still was not well resolved, but Google reportedly was still using the workaround to remove all gorillas from the training data, and thus was not able to recognize real gorillas at all.[95] Similar issues with recognizing non-white people have been found in many other systems.[96] In 2016, Microsoft tested a chatbot that learned from Twitter, and it quickly picked up racist and sexist language.[97] Because of such challenges, the effective use of machine learning may take longer to be adopted in other domains.[98] Concern for fairness in machine learning, that is, reducing bias in machine learning and propelling its use for human good is increasingly expressed by artificial intelligence scientists, including Fei-Fei Li, who reminds engineers that "There’s nothing artificial about AI...It’s inspired by people, it’s created by people, and—most importantly—it impacts people. It is a powerful tool we are only just beginning to understand, and that is a profound responsibility.”[99] Model assessments Classification of machine learning models can be validated by accuracy estimation techniques like the holdout method, which splits the data in a training and test set (conventionally 2/3 training set and 1/3 test set designation) and evaluates the performance of the training model on the test set. In comparison, the K-fold-cross-validation method randomly partitions the data into K subsets and then K experiments are performed each respectively considering 1 subset for evaluation and the remaining K-1 subsets for training the model. In addition to the holdout and cross-validation methods, bootstrap, which samples n instances with replacement from the dataset, can be used to assess model accuracy.[100] In addition to overall accuracy, investigators frequently report sensitivity and specificity meaning True Positive Rate (TPR) and True Negative Rate (TNR) respectively. Similarly, investigators sometimes report the false positive rate (FPR) as well as the false negative rate (FNR). However, these rates are ratios that fail to reveal their numerators and denominators. The total operating characteristic (TOC) is an effective method to express a model's diagnostic ability. TOC shows the numerators and denominators of the previously mentioned rates, thus TOC provides more information than the commonly used receiver operating characteristic (ROC) and ROC's associated area under the curve (AUC).[101] Ethics Machine learning poses a host of ethical questions. Systems which are trained on datasets collected with biases may exhibit these biases upon use (algorithmic bias), thus digitizing cultural prejudices.[102] For example, using job hiring data from a firm with racist hiring policies may lead to a machine learning system duplicating the bias by scoring job applicants against similarity to previous successful applicants.[103][104] Responsible collection of data and documentation of algorithmic rules used by a system thus is a critical part of machine learning. Because human languages contain biases, machines trained on language corpora will necessarily also learn these biases.[105][106] Other forms of ethical challenges, not related to personal biases, are more seen in health care. There are concerns among health care professionals that these systems might not be designed in the public's interest but as income-generating machines. This is especially true in the United States where there is a long-standing ethical dilemma of improving health care, but also increasing profits. For example, the algorithms could be designed to provide patients with unnecessary tests or medication in which the algorithm's proprietary owners hold stakes. There is huge potential for machine learning in health care to provide professionals a great tool to diagnose, medicate, and even plan recovery paths for patients, but this will not happen until the personal biases mentioned previously, and these "greed" biases are addressed.[107] Hardware Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks (a particular narrow subdomain of machine learning) that contain many layers of non-linear hidden units.[108] By 2019, graphic processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method of training large-scale commercial cloud AI.[109] OpenAI estimated the hardware compute used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017), and found a 300,000-fold increase in the amount of compute required, with a doubling-time trendline of 3.4 months.[110][111] Software Software suites containing a variety of machine learning algorithms include the following: Free and open-source so

nlp-

星空智能对话机器人的Gavin认为Transformer是拥抱数据不确定性的艺术。 Transformer的架构、训练及推理等都是在Bayesian神经网络不确定性数学思维下来完成的。Encoder-Decoder架构、Multi-head注意力机制、Dropout和残差网络等都是Bayesian神经网络的具体实现；基于Transformer各种模型变种及实践也都是基于Bayesian**指导下来应对数据的不确定性；混合使用各种类型的Embeddings来提供更好Prior信息其实是应用Bayesian**来集成处理信息表达的不确定性、各种现代NLP比赛中高分的作品也大多是通过集成RoBERTa、GPT、ELECTRA、XLNET等Transformer模型等来尽力从最大程度来对抗模型信息表示和推理的不确定性。 从数学原理的角度来说，传统Machine Learning及Deep learning算法训练的目标函数一般是基于Naive Bayes数学原理下的最大似然估计MLE和最大后验概率MAP来实现，其核心是寻找出最佳的模型参数；而Bayesian的核心是通过计算后验概率Posterior的predictive distribution，其通过提供模型的不确定来更好的表达信息及应对不确定性。对于Bayesian架构而言，多视角的先验概率Prior知识是基础，在只有小数据甚至没有数据的时候是主要依赖模型Prior概率分布(例如经典的高斯分布)来进行模型推理，随着数据的增加，多个模型会不断更新每个模型的参数来更加趋近真实数据的模型概率分布；与此同时，由于（理论上）集成所有的模型参数来进行Inference，所以Bayesian神经网络能够基于概率对结果的提供基于置信度Confidence的分布区间，从而在各种推理任务中更好的掌握数据的不确定性。 当然，由于Bayesian模型因为昂贵的CPU、Memory及Network的使用，在实际工程实践中计算Bayesian神经网络中所有概率模型分布P(B)是棘手的甚至是Intractable的几乎不能实现事情，所以在工程落地的时候会采用Sampling技术例如MCMC的Collapsed Gibbs Sampling、Metropolis Hastings、Rejection Sampling及Variational Inference的Mean Field及Stochastic等方法来降低训练和推理的成本。Transformer落地Bayesian**的时候权衡多种因素而实现最大程度的近似估计Approximation，例如使用了计算上相对CNN、RNN等具有更高CPU和内存使用性价比的Multi-head self-attention机制来完成更多视角信息集成的表达，在Decoder端训练时候一般也会使用多维度的Prior信息完成更快的训练速度及更高质量的模型训练，在正常的工程落地中Transformer一般也会集成不同来源的Embeddings，例如星空智能对话机器人的Transformer实现中就把One-hot encoding、Word2vec、fastText、GRU、BERT等encoding集成来更多层级和更多视角的表达信息。 拥抱数据不确定性的Transformer基于Bayesian下共轭先验分布conjugate prior distribution等特性形成了能够整合各种Prior知识及多元化进行信息表达、及廉价训练和推理的理想架构。理论上讲Transformer能够更好的处理一切以 “set of units” 存在的数据，而计算机视觉、语音、自然语言处理等属于这种类型的数据，所以理论上讲Transformer会在接下来数十年对这些领域形成主导性的**力。

okhttp

Square’s meticulous HTTP client for the JVM, Android, and GraalVM.

parallelize-ml-inference

Sample code for parallelizing across multiple CPU/GPUs on a single machine to speed up deep learning inference

pmls-caffe

PMLS-Caffe: Distributed Deep Learning Framework for Parallel ML System

practical-machine-learning-with-python

Master the essential skills needed to recognize and solve complex real-world problems with Machine Learning and Deep Learning by leveraging the highly popular Python Machine Learning Eco-system.

qix

Machine Learning、Deep Learning、PostgreSQL、Distributed System、Node.Js、Golang

querymq

江苏省人民医院

serving

A flexible, high-performance serving system for machine learning models

shadowsocks-windows

If you want to keep a secret, you must also hide it from yourself.

signal-android

A private messenger for Android.