In the previous lectures, we looked at a variety of algorithms for DM and ML, eg. kNN, clustering, neural networks.

Now we'll examine how these have been/can be implemented - using tools/frameworks or APIs/languages/hardware. Note that you've already seen some of them, and even used some in HWs.

Note that in 'industry' (ie. outside academia and gov't), tools/APIs... are heavily used to build RL products - so you need to be aware of, and knowledgeable in, as many of them as possible.

ADVICE: you are valued (compensated!) for your knowledge of tools, as much as (if not more than) your knowledge of theory - so, MASTER USEFUL TOOLS!

These are the most heavily used:

• TensorFlow ('TF') [we already looked at this, in the ML lecture]
• Keras: https://keras.io/ [a higher level lib, compared to TF etc] - here is the latest
• Torch, PyTorch: https://pytorch.org, http://torch.ch/
• Spark MLlib: https://spark.apache.org/mllib/ and https://spark.apache.org/docs/2.2.0/ml-pipeline.html
• scikit-learn: https://scikit-learn.org/stable/
• Apache mxnet: https://mxnet.apache.org/ [multi-language APIs, GPU and cloud support...]
• Theano: http://www.deeplearning.net/software/theano/ [for multi-dim arrays]
• Caffe2, Caffe: https://caffe2.ai/, http://caffe.berkeleyvision.org/
• CNTK: https://docs.microsoft.com/en-us/cognitive-toolkit/

JAX is relatively new - it is high-performance numerical computing library written 'on top of' (using) NumPy. You can learn more here.

Upcoming/lesser-used/'internal'/specific:

• here is FBLearner Flow - Facebook's version of TensorFlow :)
• Apache Mahout - a collection of ML algorithms, in Java/Scala
• .NET ML: https://dotnet.microsoft.com/apps/machinelearning-ai/ml-dotnet
• fastai [on top of PyTorch]: https://github.com/fastai/fastai
• OpenVINO: https://software.intel.com/en-us/openvino-toolkit and https://www.youtube.com/watch?v=rUwayTZKnmA&t=1s [a tutorial]
• Turi: an alternative to Apple's CreateML: https://github.com/apple/turicreate
• LibSVM: https://www.csie.ntu.edu.tw/~cjlin/libsvm/
• LightGBM: https://github.com/Microsoft/LightGBM
• XGBoost: https://xgboost.ai/ [and, look at Tianqi's slides and talk]
• CatBoost: https://tech.yandex.com/catboost/

The virtually unlimited computing power and storage that a cloud offers, make it an ideal platform for data-heavy and computation-heavy applications such as ML.

Amazon: https://aws.amazon.com/machine-learning/

Google: https://cloud.google.com/products/ai/ [in addition, Colab is an awesome resource!]

Microsoft: https://azure.microsoft.com/en-us/services/machine-learning-studio/ [and AutoML]

IBM Cloud, Watson: https://www.ibm.com/cloud/ai [eg. look at https://www.ibm.com/cloud/watson-language-translator]

Others:

• h2o: https://www.h2o.ai/products/h2o/ [supports R, Python, Java, Scala, JSON, native Flow GUI [similar to Jupyter], REST...]
• BigML: https://bigml.com/features#platform
• FloydHub: https://www.floydhub.com/
• Paperspace: https://www.paperspace.com/ - GPU cloud (FREE for personal use)
• ...

A pre-trained model includes an architecture, and weights obtained by training the architecture on specific data (eg. flowers, typical objects in a room, etc) - ready to be deployed.

Eg. this is simple object detection in the browser! You can even run this detector on a command line.

Apple's CreateML is useful for creating a pre-trained model, which can then be deployed (eg. as an iPad app) using the companion CoreML product.

Several end-to-end applications exist, for DM/ML. Here are popular ones.

Weka is a Java-based collection of machine learning algorithms.

RapidMiner uses a dataflow ("blocks wiring") approach for building ML pipelines.

KNIME is another dataflow-based application.

Orange is yet another dataflow-based tool.

Analytic Solver Data Mining is an Excel add-on.

These languages are popular, for building ML applications (the APIs we saw earlier, are good examples):

• Python
• R
• Julia [a potential Python replacement!]
• Wolfram
• JavaScript - this is a good list of JS-based libraries [look at ConvnetJS for nice demos]
• Scala - a functional+OO language - here is a roundup of libraries [these are in addition to Spark's MLlib Scala API]
• Java - another robust language for building ML libs [we already saw WEKA] and apps
• Jupyter [an environment, not a language] (eg. here is a collection of ML notebooks - as an exercise, run them all in Colab!)
• ...

Because (supervised) ML is computationally intensive, and detection/inference needs to happen in real-time almost always, it makes sense to accelerate the calculations using hardware. Following are examples.

Google TPU: TF is in hardware! Google uses a specialized chip called a 'TPU', and documents TPUs' improved performance compared to GPUs. Here is a pop-sci writeup, and a Google blog post on it.

Amazon Inferentia: a chip, for accelerating inference (detection): https://aws.amazon.com/machine-learning/inferentia/

NVIDIA DGX-1: an 'ML supercomputer': https://www.nvidia.com/en-us/data-center/dgx-1/ [here is another writeup]

Intel's Movidius (VPU): https://www.movidius.com/ - on-device computer vision

In addition to chips and machines, there are also boards and devices:

• Pixy2: https://pixycam.com/ - camera + ML in a single board
• Coral: https://coral.withgoogle.com/
• Jetson Nano: https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/
• Movidius NCS: https://software.intel.com/en-us/movidius-ncs
• ...

We looked at a plethora of ways to 'do' ML. Pick a few, and master them - they complement your coursework-based (theoretical) knowledge, and, make you marketable to employers!