This post is authored by Erika Menezes, Software Engineer at Microsoft. Using deep learning to learn feature representations from near-raw input has been shown to outperform traditional task-specific feature engineering in multiple domains in several situations, including in object recognition, speech recognition and text classification. With the recent advancements in neural networks, deep learning has been gaining popularity in computational creativity tasks such as music generation. There has been great progress in this field via projects such as Magenta, an open-source project focused on creating machine learning projects for art and music, from the Google Brain team, and Flow Machines, who have released an entire AI generated pop album. For those of you who are curious about music generation, you can find additional resources here.

The history of artificial intelligence is a procession of one-trick ponies. Over decades researchers have crafted a series of super-specialized programs to beat humans at tougher and tougher games. Most recently, Alphabet's DeepMind research group shocked the world with a program called AlphaGo that mastered the Chinese board game Go. But each of these artificial champions could play only the game it was painstakingly designed to play. DeepMind has now revealed the first multi-skilled AI board-game champ. A paper posted late Tuesday describes software called AlphaZero that can teach itself to be super-human in any of three challenging games: chess, Go, or Shogi--a game sometimes dubbed Japanese chess.

Our vast experience with planning data science conferences across a multitude of industries has enabled us to host, listen and learn valuable insights into the industry's most ambitious goals and research advancements. As the data science community heads towards 2018, we asked our top speakers to comment on 2017's most impactful achievements in Artificial Intelligence and make a few predictions for 2018. We summarize the most notable insights in this post, and offer expert commentary on the advancements, predictions and lessons learned regarding machine learning algorithms and deep learning systems. Daniel Monistere, SVP-Client Solutions at Nielsen points out the technology advancement electronic devices have met and the increase in their storage and data gathering capabilities. Also, applications have become intelligent being able to collect user data.

Governments can have a pretty dismal track record when it comes to predicting the next big thing. Tax dollars spent on visionary projects are often, it seems, tax dollars thrown away. But, this past spring, Ottawa might have made its best bet yet with the $125 million it has set aside over the next five years for a Pan-Canadian Artificial Intelligence Strategy. That money will go to three academic centres: the Montreal Institute for Learning Algorithms (MILA), the Alberta Machine Intelligence Institute (AMII) in Edmonton, and the new Vector Institute for Artificial Intelligence, based in Toronto. In return, the three organizations are to hire more scientists, do more research, train more students and – the important bit – nourish a growing ecosystem that will provide Canadian jobs, products and services based on artificial intelligence, or AI.

IBM has unveiled next-generation Power Systems Servers incorporating its newly designed Power9 processor, built specifically for compute-intensive AI workloads. Tthe new Power9 systems are capable of improving the training times of deep learning frameworks by nearly 4-times, allowing enterprises to build more accurate AI applications, faster. The new Power9 -based AC922 Power Systems are the first to embed PCI-Express 4.0, next-generation NVIDIA NVLink and OpenCAPI, which combined can accelerate data The system was designed to drive demonstrable performance improvements across popular AI frameworks such as Chainer, TensorFlow and Caffe, as well as accelerated databases such as Kinetica. As a result, data scientists can build applications faster, ranging from deep learning insights in scientific research, real-time fraud detection and credit risk analysis. Power9 is at the heart of the soon-to-be most powerful data-intensive supercomputers in the world, the US Department of Energy's "Summit" and "Sierra" supercomputers, and has been tapped by Google.

Researchers from Purdue University have developed a method using functional magnetic resonance imaging (fMRI) and computer algorithms to map the neural networks of the visual cortex and to build a model of their visual experience as it occurs. As recently reported in Science, the researchers have made possible technology that once belonged in the realm of science fiction. During the "training" phase of the research project, women were shown video clips of people, animals, or scenes from nature. Each video clip was shown multiple times to enable the research team to collect data on the neural activity of the visual cortex of the brain as it responded to aspects of each clip such colour, spatial orientation, or size. Researchers then used this data to make predictions regarding which areas would be stimulated when that same person watched a specific video clip.

Summary: As a profession we do a pretty poor job of agreeing on good naming conventions for really important parts of our professional lives. "Machine Learning" is just the most recent case in point. It's had a perfectly good definition for a very long time, but now the deep learning folks are trying to hijack the term. Let's make up our minds. As a profession we do a pretty poor job of agreeing on good naming conventions for really important parts of our professional lives.

It's an extreme rhetorical, but it illustrates one of the biggest challenges facing machine learning today. Neural networks are only as good as the information they're trained on, which had led to high-profile examples of how susceptible they are to bad data riddled with bias. But these technologies are also vulnerable to another kind of weakness known as "adversarial examples." An adversarial example occurs when a neural net identifies an image as one thing–while any person looking at it sees something else. The phenomenon was discovered in 2013, when a group of researchers from Google and OpenAI realized they could slightly shift the pixels in an image so that it would appear the same to the human eye, but a machine learning algorithm would classify it as something else entirely.

In this paper, we propose novel generative models for creating adversarial examples, slightly perturbed images resembling natural images but maliciously crafted to fool pre-trained models. We present trainable deep neural networks for transforming images to adversarial perturbations. Our proposed models can produce image-agnostic and image-dependent perturbations for both targeted and non-targeted attacks. We also demonstrate that similar architectures can achieve impressive results in fooling classification and semantic segmentation models, obviating the need for hand-crafting attack methods for each task. Using extensive experiments on challenging high-resolution datasets such as ImageNet and Cityscapes, we show that our perturbations achieve high fooling rates with small perturbation norms. Moreover, our attacks are considerably faster than current iterative methods at inference time.

The Generative Adversarial Networks (GANs) have demonstrated impressive performance for data synthesis, and are now used in a wide range of computer vision tasks. In spite of this success, they gained a reputation for being difficult to train, what results in a time-consuming and human-involved development process to use them. We consider an alternative training process, named SGAN, in which several adversarial "local" pairs of networks are trained independently so that a "global" supervising pair of networks can be trained against them. The goal is to train the global pair with the corresponding ensemble opponent for improved performances in terms of mode coverage. This approach aims at increasing the chances that learning will not stop for the global pair, preventing both to be trapped in an unsatisfactory local minimum, or to face oscillations often observed in practice. To guarantee the latter, the global pair never affects the local ones. The rules of SGAN training are thus as follows: the global generator and discriminator are trained using the local discriminators and generators, respectively, whereas the local networks are trained with their fixed local opponent. Experimental results on both toy and real-world problems demonstrate that this approach outperforms standard training in terms of better mitigating mode collapse, stability while converging and that it surprisingly, increases the convergence speed as well.