The AI News Roundup provides you with our take on the coolest and most interesting Artificial Intelligence (AI) news and developments each month. Stay tuned and feel free to comment with any stories you think we missed!
_________________________________________________________________
About one in every eight men will be diagnosed with prostate cancer in his lifetime, and after lung cancer, prostate cancer is the second most prevalent cause of cancer death in American men. Traditional diagnostic tests for prostate cancer only have an accuracy of about 30%, thereby requiring additional invasive testing in most patients just to diagnose the disease.
This need for invasive biopsy may be going away soon, however, as the Korea Institute of Science and Technology (KIST) has recently developed a much less invasive urine test that is nearly 100% accurate. This technique uses “an ultrasensitive semiconductor sensor system” for detecting tiny quantities of certain cancer factors in urine, combined with an AI algorithm that can identify complex signals and correlations in the sensor data to diagnose prostate cancer with high accuracy.
We’ve discussed Google’s DeepMind project several times before on the blog, and we’re talking about it this month once again. DeepMind’s newest program, called MuZero, has beaten many of the same games that we’ve discussed in previous blogs, like chess and Go. However, MuZero has achieved what previous DeepMind AI agents never have: it has figured out the rules of the games for itself.
According to DeepMind scientist David Silver, “For the first time, we actually have a system which is able to build its own understanding of how the world works.” With this development, AI technology is getting closer and closer to achieving what we in the biz call “general artificial intelligence” — the ability of a computer to learn and understand anything that a human can. For more on how MuZero is being put to use, check out this article from the BBC.
For all their positive characteristics, like superhuman performance on important tasks and a broad scope of applicability, the neural network has been plagued with justifiable complaints of inscrutability. The black-box approach has left many scratching their heads as to why certain correct and incorrect predictions are made (adversarial examples are interesting in this regard), or why and how certain concepts are learned in general.
Researchers at Duke University have recently come up with the notion of concept whitening, which bakes concepts into the networks themselves in each layer, tuning specific neurons to focus on a single output class. The focus here is on convolutional neural networks, and is implemented in training in a manner similar to batch normalization, making it relatively plug-and-play with existing models.
OpenAI, best known for their GPT-3 text generation model, has boggled minds again with their latest creation: Dall-E. This new project of theirs, however, doesn’t create humanlike text like GPT-3 does — it’s an image generator, prompting users to supply a text description that it uses as the basis for its computational imagination.
Though built in part based on GPT-3, Dall-E’s training naturally also included copious images, organized into pairs with appropriate text descriptions. There is currently no publicly available version of the model and no available paper describing it in full, there is a substantial blog post available here.
On planetary bodies, impact craters resulting from large collisions can explain a lot about the impacted object’s character and composition. Old layers of material are exposed for observation and sampling, providing a glimpse into geological history. Naturally, planetary scientists, astrophysicists, and people in related fields are deeply interested in cataloging impact craters for photographic documentation and scientific analysis.
Researchers at the NASA Jet Propulsion Laboratory have taken many pictures of Mars, with the Mars Reconnaissance Orbiter transmitting images of the red planet for about fifteen years. These troves of image data, while immensely valuable, are tough to parse by eye. Enter one of JPL’s latest creations: an AI that detects undiscovered craters from MRO images. Instead of using humans to comb through photos spending “…three-quarters of an hour for a single image,” the AI can expedite the crater-discovery process immensely.
That’s it for this month! In case you missed it, here’s our previous roundup with even more cool AI news. Check back in February for more of the most interesting developments in the AI community (from our point of view, of course).
_________________________________________________________________
If you liked this blog post, check out more of our work and follow us on social media (Twitter, LinkedIn, and Facebook).