What’s New in the Latest TensorFlow 2.10?

Updated November 2026. This was first published on 28 November 2022. The idea is the same. Any number below is what the original piece said then. It is not a new count.

Key takeaways

TensorFlowis a free and open-source software library for artificial intelligence and machine learning.
Although it can be applied to many different tasks, deep neural network training and inference are given special emphasis.
The Google Brain team created TensorFlow for use in internal Google research and production.

TensorFlowis a free and open-source software library for artificial intelligence and machine learning. Although it can be applied to many different tasks, deep neural network training and inference are given special emphasis.

The Google Brain team created TensorFlow for use in internal Google research and production. 2015 saw the maiden release under the Apache License 2.0. In September 2019, TensorFlow 2.0, the upgraded version of TensorFlow, was released by Google.

Python, JavaScript, C++, and Java are just a few of the many programming languages to which TensorFlow can be applied. This adaptability allows for many applications across numerous industries.

What is Tensorflow? #

The second-generation system of Google Brain is called TensorFlow. The 1.0.0 version was made available on February 11, 2017. TensorFlow can run on several CPUs and GPUs, unlike the reference setup. only uses a single device (with optional CUDA and SYCL extensions for general-purpose computing on graphics processing units). On 64-bit Linux, macOS, Windows, and mobile operating systems like Android and iOS, TensorFlow is accessible.

Because of its adaptable design, computing may be easily deployed across a range of platforms (CPUs, GPUs, and TPUs), from desktop computers to server clusters to mobile and edge devices.

TensorFlow calculations are represented using stateful dataflow graphs. These neural networks’ actions on multidimensional data arrays, known as tensors, are where the name TensorFlow originates. Jeff Dean revealed at the Google I/O Conference in June 2016 that just 5 of the 1,500 GitHub projects that cited TensorFlow were from Google.

Update in Tensorflow #

Google released TensorFlow.js version 1.0 for JavaScript machine learning in March 2018. Google unveiled TensorFlow 2.0 in January 2019. In September 2019, it was made formally accessible. Google released TensorFlow Graphics for deep learning in computer graphics in May 2019.

Edge TPU #

The Edge TPU was announced in July 2018. Edge TPU is Google’s custom-built ASIC chip created to execute TensorFlow Lite machine learning (ML) models on edge computing clients. include small client computing devices like smartphones.

TensorFlow Lite #

Google unveiled TensorFlow Lite, a software stack designed exclusively for mobile development, in May 2017. Google’s TensorFlow Lite Micro, also known as TensorFlow Lite for Microcontrollers, and ARM’s uTensor were to merge, according to a May 2019 announcement.

Pixel Visual Core (PVC) #

Google announced the Pixel Visual Core (PVC), a completely programmable image, vision, and AI engine for mobile devices, in October 2017 with the Google Pixel 2. The PVC for machine learning supports TensorFlow.

TensorFlow 2.0 #

The TensorFlow team announced the release of a new major version of the library in September 2019 as TensorFlow’s market share among research papers was dropping to PyTorch’s advantage. The most significant of the many changes made by TensorFlow 2.0 was the introduction of TensorFlow Eager. replaced the static computational graph-based automatic differentiation scheme with the “Define-by-Run” scheme first made well-known by Chainer and later PyTorch. Other key changes included eliminating outdated libraries, cross-compatibility across trained models on various TensorFlow iterations, and a notable boost in GPU performance.

What is New in New Tensorflow 2.10? #

New tools for loading audio data Improvements to the optimizers’ API Deterministic and stateless initializers User-friendly Keras what it can do to aid in the development of transformers.

TensorFlow Decision Forests 1.0 is also included in this release!

Updates to Keras #

To support a unified and extended portfolio of built-in and customized optimizers, TensorFlow also published an experimental new version of the Keras Optimizer API for version 2.9. The then-introduced experimental version of the optimizers will replace the current version namespace in the following release (TensorFlow v2.11). The team has started exporting current Keras optimizers under tf.keras.optimizers.legacy in version 2.10 to get ready for the official transfer of the optimizer namespace to the new API.

Note that TensorFlow offered an API to make operations deterministic in version 2.8; the team attempted to enhance its performance in version 2.9. The team has now created stateless, deterministic Keras initializers on top of stateless TF random operations. Both seeded and unseeded initializers will produce identical values starting with the new version each time they are invoked. Keras will be able to implement additional features like multi-client model training with DTensor if it has a stateless initializer (a TensorFlow extension for synchronous distributed computing).

oneDNN #

To improve performance on Intel CPUs, TensorFlow v2.9 included the oneDNN performance package. OneDNN optimizations would be enabled by default in this version on Linux x86 packages and for CPUs having hardware features geared toward neural networks, such as those found in Intel Cascade Lake and newer CPUs.

to integrate the Compute Library for the Arm Architecture with oneDNN moving forward, TensorFlow has partnered with Arm AWS and Linaro in version 2.10. On aarch64 CPUs, this integration would improve performance.

TensorFlow Decision Forests #

Modern algorithms for training and analyzing Decision Forest models may be found in TensorFlow Decision Forests. It is a set of Keras models that supports regression, classification, and ranking.

With the release of TensorFlow 2.10, the Decision Forests are v1.0, creating a more reliable and developed library. According to the manufacturer, TensorFlow Decision Tree is prepared for usage in business settings with updated documentation and more thorough testing. The new release provides Javascript and Go APIs for inference of the Decision Tree models. The team is looking for input on these APIs, which are still in beta.

By making splits based on the linear combinations of characteristics, oblique splits, which recursively partition feature space, perform better with TensorFlow Decision Tree 1.0. They enable decision trees to describe complicated patterns by simultaneously conditioning them on a number of factors. On most datasets, it has been demonstrated that oblique splits perform better than axis-aligned divides.

Conclusion #

TensorFlow is currently the most popular machine learning platform. We have millions of developers using it. On GitHub, it has received the third-most stars (after Vue and React), and on PyPI, it is the most-downloaded machine learning package. TensorFlow underpins nearly all production machine learning in all of Google’s products, including Search, Gmail, YouTube, Maps, Play, Ads, Photos, and many others. Other Alphabet companies use TensorFlow and Keras than Google to power the artificial intelligence in Waymo’s self-driving cars.

Many of the biggest machine learning users in the world, such as Apple, ByteDance, Netflix, Tencent, Twitter, and countless others, rely on TensorFlow to run their machine learning systems. Also, Google Scholar indexes around 3,000 new scientific publications each month that reference TensorFlow or Keras.

Questions and answers #

What is the main point? #

TensorFlowis a free and open-source software library for artificial intelligence and machine learning.

What else did the original piece say? #

Although it can be applied to many different tasks, deep neural network training and inference are given special emphasis.