At this year's SXSW the theme for our conference tech coverage is "AI for good." Technologists will be discussing how they are using new developments in artificial intelligence to solve complex, long-standing social problems, including crimes against children. Federico Gomez Suarez (pictured), senior technical program manager at Microsoft, spoke to John Furrier (@furrier), host of theCUBE, SiliconANGLE Media's mobile live streaming studio, during SXSW about his involvement in Thorn, a special outfit that works with law enforcement to rescue traffic or exploited children. "The fact that we're making a difference in those lives is extremely encouraging," Suarez said. He credits Microsoft's Hack for Good program for paving his avenue to Thorn.

For Silicon Valley, the headline was sweet nectar: Google DeepMind, the world's hottest artificial intelligence lab, embraces the blockchain, the endlessly fascinating idea at the heart of the bitcoin digital currency. The lab's re-imagining of the blockchain has very little to do with AI--or the blockchain, for that matter. If you want AI crossed with the blockchain, try wrapping your head around Numerai, the world's strangest hedge fund. To DeepMind's credit, its new project depends less on trendy ideas than an apparent desire to solve a real problem in the real world--one that involves the most private and personal information. DeepMind is building an auditing system for healthcare data. That may not sound sexy, but it matters.

The Indus script has long challenged epigraphists because of the difficulty in reading and classifying text and symbols on the artefacts. Now, a Chennai-based team of scientists has built a programme which eases the process. Ronojoy Adhikari of The Institute of Mathematical Sciences and Satish Palaniappan, who is at Sri Sivasubramaniya Nadar College of Engineering, have developed a "deep-learning" algorithm that can read the Indus script from images of artefacts such as a seal or pottery that contain Indus writing. Scanning the image, the algorithm smartly "recognises" the region of the image that contains the script, breaks it up into individual graphemes (the term in linguistics for the smallest unit of the script) and finally identifies these using data from a standard corpus. In linguistics the term corpus is used to describe a large collection of texts which, among other things, are used to carry out statistical analyses of languages.

DeepMind, the Google-owned artificial intelligence company, is developing a new technology similar to blockchain for secure tracking of patient health data. In a blog post, London-based DeepMind said its new Verifiable Data Audit project could be the first steps toward a "a service that could give mathematical assurance about what is happening with each individual piece of personal data, without possibility of falsification or omission." The aim is to enable hospitals, and eventually patients, to gain real-time insight into where and how data is being used. "For example, an organization holding health data can't simply decide to start carrying out research on patient records being used to provide care, or repurpose a research dataset for some other unapproved use," according to DeepMind. "It's not just where the data is stored, it's what's being done with it that counts. We want to make that verifiable and auditable, in real-time, for the first time."

Now seems like as good of a time as any to post some of the talks that I've done in the last year. I was originally going to give a talk I had done before (hence the same title), but I had some recent inspiration about some issues in the field that I wanted to share. This was a last minute whim, and after 24 hours of neither eating nor sleeping, I finished my slides with 30 minutes to spare. The talk went great and immediately afterwards, I even had the opportunity to do a TWiML podcast (thanks, Sam!). Both were surprisingly well-received, and writing the talk really helped me refine my thoughts on the field and what direction I wanted to take my work.

A key challenge in scaling up robot learning to many skills and environments is removing the need for human supervision, so that robots can collect their own data and improve their own performance without being limited by the cost of requesting human feedback. Model-based reinforcement learning holds the promise of enabling an agent to learn to predict the effects of its actions, which could provide flexible predictive models for a wide range of tasks and environments, without detailed human supervision. We develop a method for combining deep action-conditioned video prediction models with model-predictive control that uses entirely unlabeled training data. Our approach does not require a calibrated camera, an instrumented training set-up, nor precise sensing and actuation. Our results show that our method enables a real robot to perform nonprehensile manipulation -- pushing objects -- and can handle novel objects not seen during training.

A novel deep learning architecture (XmasNet) based on convolutional neural networks was developed for the classification of prostate cancer lesions, using the 3D multiparametric MRI data provided by the PROSTATEx challenge. End-to-end training was performed for XmasNet, with data augmentation done through 3D rotation and slicing, in order to incorporate the 3D information of the lesion. XmasNet outperformed traditional machine learning models based on engineered features, for both train and test data. For the test data, XmasNet outperformed 69 methods from 33 participating groups and achieved the second highest AUC (0.84) in the PROSTATEx challenge. This study shows the great potential of deep learning for cancer imaging.

Deep neural network algorithms are difficult to analyze because they lack structure allowing to understand the properties of underlying transforms and invariants. Multiscale hierarchical convolutional networks are structured deep convolutional networks where layers are indexed by progressively higher dimensional attributes, which are learned from training data. Each new layer is computed with multidimensional convolutions along spatial and attribute variables. We introduce an efficient implementation of such networks where the dimensionality is progressively reduced by averaging intermediate layers along attribute indices. Hierarchical networks are tested on CIFAR image data bases where they obtain comparable precisions to state of the art networks, with much fewer parameters. We study some properties of the attributes learned from these databases.

How spiking networks are able to perform probabilistic inference is an intriguing question, not only for understanding information processing in the brain, but also for transferring these computational principles to neuromorphic silicon circuits. A number of computationally powerful spiking network models have been proposed, but most of them have only been tested, under ideal conditions, in software simulations. Any implementation in an analog, physical system, be it in vivo or in silico, will generally lead to distorted dynamics due to the physical properties of the underlying substrate. In this paper, we discuss several such distortive effects that are difficult or impossible to remove by classical calibration routines or parameter training. We then argue that hierarchical networks of leaky integrate-and-fire neurons can offer the required robustness for physical implementation and demonstrate this with both software simulations and emulation on an accelerated analog neuromorphic device.

The weight of the automotive and tech industries is fully behind the move toward self-driving cars. Cars with "limited autonomy"--i.e., the ability to drive themselves under certain conditions (level 3) or within certain geofenced locations (level 4)--should be on our roads within the next five years. But a completely autonomous vehicle--capable of driving anywhere, any time, with human input limited to telling it just a destination--remains a more distant goal. To make that happen, cars are going to need to know exactly where they are in the world with far greater precision than currently possible with technology like GPS. And that means new maps that are far more accurate than anything you could buy at the next gas station--not that a human would be able to read them anyway.