Articles
Robert Lange wrote an excellent introduction article for Jax. A great article that outlines advantages of Jax and its flexibility through `vmap` and other concepts.
Google wrote about how they built a system for people who have low level vision to guide them on their running route. Beyond segmentation/filtering aspects, it is encouraging to see that machine learning can be leveraged to augment/improve human’s sensing capabilities.
Facebook wrote about mobile capabilities of PyTorchVideo. If you are interested in deploying torch models for various video tasks like action recognition, you might want to check out the post.
FastForwardLabs wrote about session based recommender systems in the context of ecommerce. If you want to learn about recommender systems and want to learn about content based and collaborative filtering approaches, this article gives a good overview for beginners.
This article gives a comprehensive comparison of ml platforms from various companies. The article has a very long list of references that gives various pointers blog posts, papers or videos to create the table shown above. *IH stands for in-house.
Pye.ai wrote an article that gives a comparison between different feature store solutions. If you want to use a feature store and want to see what other options are available in the industry, it is a definitely good start. However, if you are looking to a detailed comparison between these solutions in terms of advantages and disadvantages, you still need to do your own homework.
Libraries
Flops Counter PyTorch is a library to estimate the multiply-add operations in Convolutional Neural Networks. It also estimates the parameters per layer to estimate the cost of the neural network.
InvoiceNet is library to extract information from various invoices based on Tensorflow.
DataProfiler is a library to detect sensitive data(like Personal Identifiable Information) in the dataset.
Videos
In Google I/O, among many other topics, Responsible AI topic was also given a lot of importance. This talk gave a lot of good pointers to various libraries in TF ecosystem to integrate into a workflow to make sure the ML pipeline as well as model building is responsible and free from a variety of biases. Responsible AI website has also a ton of pointers to other tools that you want to use in your workflow if you are using TF. A detailed guidebook is also available in here. The main website for human centered AI is also in here, its accompanying video in Google I/O is here.
Model Optimization talks with a heavy emphasis on TFLite is here. In Google I/O, there was a lot of emphasis on “edge” and “tiny-ml” and this talk gave a lot of good pointers for a variety of model optimization techniques if you want to adopt model optimization library along with TFLite. TFLite’s main website is here, model optimization library is here. TFHub has now a separate section for models for TFLite separately if you want to pick up trained and optimized models for mobile/edge devices.
TFX was also highlighted in Google I/O with a number of Google Cloud offerings. Recently published paper gives an excellent historic overview on how TFX evolved over time and become what it is right now. I like this paper not because it provides a number of excellent reasons why TFX is needed but also how it evolved over time by accommodating a number of use cases that came up in Google over time.
Beidi Chen gave a presentation on LSH(Locality Sensitive Hashing) for Efficient Neural Network Training. The paper is here if you want to know more about.
Datasets
Kaggle has a nice datasets page where they display a number of different datasets through different categories. They also have a “usability” score of the datasets which signifies how easy and documented the dataset is.
Google I/O announced a new dataset tool called “Know Your Data” and as part of the that tool, there is a now new dataset page that works with Tensorflow Dataset. An example of one of these dataset is here.
Workshops
Sparsity in Neural Networks is a free workshop that will focus on how sparsity can be leveraged to improve efficiency of the neural network and accuracy of the model.