Gradient descent is inefficient to find saddle point (Nash equilibrium) for min-max games, because of spiralling behaviour. Beware when training your GANs … en.wikipedia.org/wiki/Nash_equi… en.wikipedia.org/wiki/Generativ…

See our paper validating risk models for bowel cancer just published nature.com/articles/bjc20…. Several models based on routinely collected data have good discrimination. Next step - do they improve efficiency of screening? PCU Prevention Group Primary Care Cancer Group

Will you enjoy this puzzle on Bayesian theory and probability? The odds are you will. buff.ly/2ETk01l

Pleased to report our new publication: Overtesting and undertesting in primary care: a systematic review and meta-analysis bmjopen.bmj.com/content/8/2/e0…

Computer scientist, neuroscientist, & author of the paper "Could a Neuroscientist Understand a Microprocessor?” Dr. Jonas (Eric Jonas) sits down with 𝒶𝓁𝓌𝒶𝓎𝓈 🄵🅁🄴🅂🄷 to talk about connectomes, brain lesions & neurons vs transistors in this #HTMSchool video. youtu.be/bLbeIAGEgpc

"Deep Reinforcement Learning Doesn't Work Yet" alexirpan.com/2018/02/14/rl-… great read, hits a lot of points I've also come to realize over last ~2 years. 70% is a vast understatement.

From @sciencemagazine: 'booming field of artificial intelligence (#AI) is grappling w/replication crisis” "AI researchers found it difficult to reproduce many key results” "esearchers often don't share source code” sciencemag.org/news/2018/02/m… #BigData #DigitalHealth #MachineLearning

In an effort to enhance rigor & reproducibility in research, the J K has created a clearinghouse website for training modules that includes the new Harvard Online course: Principles, Statistical and Computational Tools for Reproducible Science youtube.com/watch?v=GJ9nK3…

From the archives: Spiders appear to offload cognitive tasks to their webs, making them one of a number of species with a mind that isn’t fully confined within the head. buff.ly/2nXpkJR

Artificial brain: A tiny mesh of nanowires and synapselike switches remembers its past and solves simple problems. buff.ly/2EkPDkf

Conscious, preconscious, and subliminal processing: a testable taxonomy researchgate.net/profile/Claire… by Stanislas Dehaene @standehaene.bsky.social ‘s team in Trends in Cognitive Sciences

Thinking about RL sample efficiency. Imagine you apply a transformation to Atari pixels that makes it look like noise to humans. How long would a person need to solve it? Very long. They’d probably just memorize the actions and overfit. Why expect RL to do something different?

Few large high quality studies are available on robotics surgery: State of the Art Review looks at existing evidence on its feasibility, safety, efficacy, and costs bmj.com/content/360/bm…

How do you get classical data in and out of a quantum neural network? Solving that problem could usher in a new era of machine learning. buff.ly/2EK4LHO

Have you checked out the Society to Improve Diagnosis in Medicine's new website? Let us know what you think. improvediagnosis.org

Deep learning for biology, featuring Casey Greene (our deep learning review), Anne Carpenter, PhD (CellProfiler), Mark DePristo (DeepVariant) and more. In @NatureNews. HT Simina M. Boca nature.com/articles/d4158…

AI experts warn of new types of threats on.ft.com/2ok7eBa

ICYMI: Complex natural systems defy standard mathematical analysis, so one ecologist is throwing out the equations. buff.ly/2E1Wulz

Simulation study recommends considering prognostic covariates in regression models when estimating biomarker–treatment interactions, as the power for detecting true interactions can be increased - #Trialsjournal: trialsjournal.biomedcentral.com/articles/10.11…

“Proofs from THE BOOK” is a collection of some of the most beautiful and interesting proofs in mathematics. We spoke with Günter Ziegler, one of the book’s curators, about what makes the proofs in the book so special. quantamagazine.org/gunter-ziegler…