Roundup Here’s a quick roundup to keep you updated on what’s been happening in AI, beyond what we've already covered, for your long weekend.
It includes news of Samsung and Qualcomm setting up new AI research teams, why human radiologists are still better than machines and support for Amazon’s Keras-MXNet backend.
Hold your horses AI radiologists People are quick to believe that machines will soon replace radiologists because they think computers are much better at spotting abnormalities like tumors or clots in medical scans.
But results reported by Stanford University shows that radiologists still trump AI. A group of researchers built a large convolutional neural network (CNN) with 169 layers to predict the probability of an abnormality appearing in a particular scan from the MURA (musculoskeletal radiographs) dataset.
AI and robotics legend Joanna Bryson to kick off MCubedREAD MORE
It collects 40,561 scans of the elbow, finger, forearm, hand, humerus, shoulder, and wrist of 12,173 patients. They compared the CNN’s results to three professional human radiologists, and the machine performed worse on every single category apart from the wrist scans.
The results have also been published in a paper on arXiv. On the plus side, the dataset has also been published on GitHub so other developers can build their own models and have a crack at improving the results.
Also, check out this Twitter thread by Lior Pachter, a professor of computational biology at the California Institute of Technology. He criticises AI evangelists like Andrew Ng for overhyping the technology to the public.
A few months ago @AndrewYNg tweeted that radiologists were on the verge of being obsolete because AI: https://t.co/rrHjShakdF. Andrew has >300K twitter followers so his tweet made the rounds (>2,000 likes) /1— Lior Pachter (@lpachter) May 24, 2018
Ng is listed as a co-author on the paper so he should know better really.
New Samsung AI labs in UK, Russia and Canada Samsung has opened a UK research hub in Cambridge.
It will be led by Andrew Blake, a professor specialising in computer vision at Cambridge University. Blake was previously a director at Microsoft Research Cambridge and the Alan Turing Institute.
Samsung is particularly interested in using AI with IoT for its products. “In the coming years, Internet of Things (IoT) devices embedded with AI will generate a vast array of data that can provide fascinating insights about our lives, analysing complex usage patterns and seamlessly enabling us to take advantage of intelligent services optimised for our own personal preferences and behaviours,” it said in a statement.
Samsung also announced it has launched research labs in Moscow, Russia and Toronto, Canada. It also has other research centers in Seoul, Korea and California.
OpenAI has emitted a full version of Gym Retro, a system for testing artificially intelligent agents within classic video games, following its challenge to developers to make bots to beat the original Sonic the Hedgehog. This framework is rather good for developing reinforcement-learning software, and evaluating it within computer game worlds.
Qualcomm gets a research lab too On the subject of new research organizations: Qualcomm is getting one too.
“Qualcomm Technologies began exploring fundamental AI research over a decade ago when it investigated spiking neuron networks for computer vision and motion control applications,” it said in a statement.
"Qualcomm AI Research’s work is diverse, spanning across power efficient AI, personalization, and data-efficient learning that builds on Qualcomm Technologies’ leadership in connectivity and computing."
Its main interest lies in hardware and building the necessary software platforms to support it for IoT devices, so it’ll probably be low powered accelerator chips.
Facebook language translation project Facebook have launched an open-source project based on its language translation model to help developers scale these tools for production.
It’s written in C++ and can be exported to Caffe 2 using ONXX, a platform developed in collaboration between many software and hardware companies to transfer models to different frameworks.
The code lets developers train a recurrent neural network - more specifically its a sequence-to-sequence model with attention that learns to map input words in one language to output ones in another language. It can also be used for other tasks such as text summarization, dialogue systems, and text generation.
You can download the model here.
MXNet + Keras Amazon’s AI framework MXNet has added backend support for Keras 2.0.
Keras is a popular TensorFlow API written in Python. It now means that developers can run large convolutional neural networks and recurrent neural networks using distributed training on MXNet.
“With an update of a few lines of code, Keras developers can increase training speed by using MXNet’s multi-GPU distributed training capabilities. Saving an MXNet model is another valuable feature of the release. You can design in Keras, train with Keras-MXNet, and run inference in production, at-scale with MXNet,” according to Amazon’s blog post.
There is also a step by step post on how to install and deploy Keras-MXNet here. ®