Computations requiring high performance computing (HPC) power may soon be done in the palm of your hand thanks to work done this summer by IBM Research in Dublin, Ireland. While scientists have come a long away in teaching machines how to process images for facial recognition and understand language to translate texts, IBM researchers focused on a different problem: how to use artificial intelligence (AI) techniques to forecast a physical process. In this case, the focus was on ocean waves, using traditional physics-based models driven by external forces, such as the rise and fall of tides, winds blowing in different directions, the depth and physical properties of water influence the speed and height of the waves. HPC is normally essential to resolve the differential equations that encapsulate these physical processes and their relationships, and the expense often limits the spatial resolution, physical processes and time-scales that can be investigated by a real-time forecasting platform. In an interview with EE Times, IBM Research Senior Research Manager Sean McKenna said an HPC cluster using Big Iron has generally been the solution to dealing with the heavy computational load.

A year after it open sourced its PaddlePaddle deep learning suite, Baidu has dropped another piece of AI tech into the public domain – a project to put AI on smartphones. Mobile Deep Learning (MDL) landed at GitHub under the MIT license a day ago, along with the exhortation "Be all eagerness to see it". MDL is a convolution-based neural network designed to fit on a mobile device. Baidu said it is suitable for applications such as recognising objects in an image using a smartphone's camera. The neural network's calculations are offloaded to a phone's GPU, the company says at its repo, with high speed and low complexity.

Creating a program brings with it an amazing feeling of accomplishment. But ask an engineer or software developer what it's like to create artificial intelligence and… well, that's a whole different level of building for the web. Even though machine learning may feel more like a sci-fi movie than reality, it's a lot closer -- and a lot more doable -- than you might think. You'll learn how possible it is for even novice web creators to get a firm grip on AI with the Deep Learning and Artificial Intelligence Introductory Bundle. But as you work through these four courses, you'll not only understand the principles top-flight engineers use to create thinking machines, but you'll start putting that knowledge into practice, building your own mini-AI projects as you progress toward greater glory. With this training, you'll learn how to create the model for a machine that can learn from multiple inputs; go inside logistic regression (a key pillar in the architecture of Deep Learning); build your very first neural network, and utilize the powerful Theano and TensorFlow Python libraries.

Fast forward to 2017 I have spent 100's of hours working on Deep learning projects and the technology has become more and more accessible due to several advancements in software(ease of usage -- Keras, PyTorch), hardware(GPU becoming commercially viable for someone like me sitting in India -Not still cheap), availability of data, good books and MOOCS. After completing the 3 most popular MOOCS in deep learning from Fast.ai, deeplearning.ai/Coursera In this post I talk about 5 aspects of each course which will help you decide. I came across this course when reading an article in kddnudgets . For the first time I heard about Jeremy Howard, searched about him in Wikipedia and was impressed .

If it wasn't for a mysterious hot pepper allergy, Harini Suresh might never have found a way to improve patient care in intensive care units. Suresh, a doctoral student at MIT, wants to use AI to help critical care doctors choose the best treatment for each patient. That's not easy when patients are sick with dire conditions like heart failure or stroke, and doctors must quickly weigh vast and varied patient data that may range from simple demographics to complex lab tests. "The ICU is a high-stakes, high-demand environment, and doctors can spend only a limited amount of time with each patient," said Suresh. "When doctors are dealing with many data sources and data types, computational tools can make a difference."

If you're playing around with Google's new Home Max or Mini smart speakers, or if you're just using an Android phone such as the new Pixel 2, you may be familiar with the Google Assistant virtual helper. And if you've done so in the last couple days, you may have noticed that the virtual assistant's voice is sounding more realistic than before. That's because Alphabet's Google has started using a cutting-edge piece of technology called WaveNet--developed by its DeepMind "artificial intelligence" division--in Google Assistant. Synthesized speech is traditionally created by gluing together bits of recorded speech, in a technique known as "concatenative text-to-speech." The result does not sound natural, although some versions of the technique are better than others.

Last year, artificial intelligence (AI) research company DeepMind shared details on WaveNet, a deep neural network used to synthesize realistic human speech. Now, an improved version of the technology is being rolled out for use with Google Assistant. A system for speech synthesis -- otherwise known as text-to-speech (TTS) -- typically utilizes one of two techniques. Concatenative TTS involves the piecing together of chunks of recordings from a voice actor. The drawback of this method is that audio libraries must be replaced whenever upgrades or changes are made.

When I started learning deep learning I spent two weeks researching. I selected tools, compared cloud services, and researched online courses. In retrospect, I wish I could have built neural networks from day one. That's what this article is set out to do. You don't need any prerequisites.

Google acquired the British artificial-intelligence startup DeepMind in 2014 for a reported £400 million (roughly $525 million), a company its cofounder Demis Hassabis once described as aiming at "solving intelligence, and then using that to solve everything else." Since then, the company's researchers have built a system that beat humans at one of the most complicated board games ever, and is now trying to beat humans at complex video games. It's building AI that's learning to navigate 3D spaces as we do, and is training other systems on British medical data, theoretically to spot illness more quickly. It's also started to integrate with teams in the US to bring its work to Google products where they might be useful. All of this research comes at a price.

Feature representations from pre-trained deep neural networks have been known to exhibit excellent generalization and utility across a variety of related tasks. Fine-tuning is by far the simplest and most widely used approach that seeks to exploit and adapt these feature representations to novel tasks with limited data. Despite the effectiveness of fine-tuning, itis often sub-optimal and requires very careful optimization to prevent severe over-fitting to small datasets. The problem of sub-optimality and over-fitting, is due in part to the large number of parameters used in a typical deep convolutional neural network. To address these problems, we propose a simple yet effective regularization method for fine-tuning pre-trained deep networks for the task of k-shot learning. To prevent overfitting, our key strategy is to cluster the model parameters while ensuring intra-cluster similarity and inter-cluster diversity of the parameters, effectively regularizing the dimensionality of the parameter search space. In particular, we identify groups of neurons within each layer of a deep network that shares similar activation patterns. When the network is to be fine-tuned for a classification task using only k examples, we propagate a single gradient to all of the neuron parameters that belong to the same group. The grouping of neurons is non-trivial as neuron activations depend on the distribution of the input data. To efficiently search for optimal groupings conditioned on the input data, we propose a reinforcement learning search strategy using recurrent networks to learn the optimal group assignments for each network layer. Experimental results show that our method can be easily applied to several popular convolutional neural networks and improve upon other state-of-the-art fine-tuning based k-shot learning strategies by more than10%