Since Google open sourced its machine learning framework in 2015, Tensorflow has risen in popularity with more than 1500 projects mentions on Github. Riding on the back of its popularity, Mountain View company made a significant announcement earlier this year, updating TensorFlow to Version 1.0 that is packed with newer features and promises more performance improvements with high level APIs.
In the showdown of AI frameworks, TensorFlow emerges as the winner, thanks to its soaring popularity among developers, contributors and researchers in forums such as Github and Stack Overflow. Interestingly, TensorFlow is being used in over 6000 open source repositories online and is being used by a wide array of people from academia and coders for language translation and early detection of skin cancer among other cases. And it is also changing the way developers are interacting with machine learning technology.
According to Fei Fei Li, in the highly rarefied field of AI, TensorFlow is tearing down by the barriers by providing the infrastructure and hardware. “AI requires enormous computing and deep learning algorithms can easily boast of tens of millions of parameters and billons of connections. Training and using such models requires computational resource,” she said, adding the TensorFlow library allows one to focus on the creativity of their solution and leave the infrastructure aside.
TensorFlow: Making Machine Learning accessible for everyone
Building on the success of DistBelief, built by Google Brain team in 2012, machine learning system based on deep learning neural networks, TensorFlow is the second-generation machine learning system. Today, TensorFlow is used in image recognition wherein deep convolutional neural networks are leveraged to deliver images with great accuracy. A new popular augmented video messaging app Tribe uses TensorFlow machine learning API and natural language processing techniques to scan the audio in video messages and look for keywords.
On the other hand, TensorFlow has also been used to combat diabetic retinopathy, one of the leading causes of blindness across the globe with nearly 415 million at risk diabetic patients. The Journal of American Medical Association proclaimed that the computer vision model could be better than a median opthamologist. The project was carried out in association with doctors from India and US with a dataset of 128,000 images that were evaluated by 3-7 ophthalmologists from a panel of 54 ophthalmologists. Further, this dataset was used to train a deep neural network to detect referable diabetic retinopathy.
What’s behind the surge in popularity of TensorFlow
On the technical end, TensorFlow 1.0’s high -level APIs make deep learning accessible to everyone. Interestingly, there is a growing body of students and researchers from varied fields who are gravitating towards deep learning thanks to TensorFlow. There is a huge community engagement with TensorFlow being declared the #1 repository in “machine learning” on Github.
Megan Kacholia Engineering Director on TensorFlow/Brain, listed down some key advantages of TensorFlow at the recent TensorFlow Dev Summit 2017
- When it comes to Neural machine translation, TensorFlow reduces errors by 55%-85%.
- In Neural architecture search, one can figure out what is the right neural network to use for a problem
- TensorFlow allows coders to iterate quickly, train models faster and run more experiments
- On the production end— teams can run TensorFlow on large scale server farms embedded on devices, CPUs, GPUs, TPUs
TensorFlow Use Cases
In the context of deep learning and neural networks, TensorFlow has gained momentum as compared to Caffe, Theano, and Torch, other well-known deep learning frameworks. Some of the most popular TensorFlow use cases are in voice recognition, text based applications such as Google Translate, Image recognition and Video Detection. In fact, NASA is developing a predictive model of Near Earth Objects with TensorFlow and Deep Learning. According to NASA, with TensorFlow, the team can design a multilayered model, that shall recognize and classify the potential of NEOs. The popular computational graph system also offers faster solutions by computing in multiple GPUs.
TensorFlow vs popular Deep Learning Frameworks
Theano: This is one of the most well-known libraries with a lot of old practitioners using Theano. And while TensorFlow is definitely better marketed than the rest (Torch) Theano wins hands down in numerical computations optimization. It also provides better visualizing convolutional filters, images, and graphs. Theano is developed and maintained by University of Montreal’s MILA, which means it has support from AI’s leading think tank Yoshua Bengio.
Torch: Torch was used by DeepMind before it migrated to TensorFlow. It uses the programming language Lua, which means one has to learn the language first before working on Torch. Torch is also extensively used at Facebook, Twitter and NVIDIA and Facebook maintains the libraries that were open sourced recently.
TensorFlow is definitely better marketed, is accessible and has ease of use as compared to other deep learning framework. And it enjoys great community support, with forums abuzz how TensorFlow has won the “framework for numerical computation using data-flow graphs” as compared to other frameworks such as Caffe, Theano and Torch. Most of the TensorFlow adopters are Python programmers, leaning towards deep learning projects. Another key point is that Google is betting heavily on its Google Cloud ML engine and this is one of the ways to attract people to Google Cloud.
Try deep learning using MATLAB