Researchers at Microsoft developed an artificial intelligence (AI) algorithm that can achieve the maximum score on Ms. Pac-Man, 999,999, four times greater than the highest human score. After recovering from your wave of relief at the news that we've solved the Ms. Pac-Man problem, you might wonder why our greatest minds were spending their days chasing that particular goal. As it turns out, this accomplishment is significant because the "divide-and-conquer" method used can be applied to teach AI agents to complete other complex tasks. The system, according to Microsoft's blog, was developed by a Maluuba, a deep learning startup company which was acquired by Microsoft earlier in the year. The divide-and-conquer method assigns individual AI agents different tasks but also allows them to work together collaboratively through a "top manager."

This paper introduces Dex, a reinforcement learning environment toolkit specialized for training and evaluation of continual learning methods as well as general reinforcement learning problems. We also present the novel continual learning method of incremental learning, where a challenging environment is solved using optimal weight initialization learned from first solving a similar easier environment. We show that incremental learning can produce vastly superior results than standard methods by providing a strong baseline method across ten Dex environments. We finally develop a saliency method for qualitative analysis of reinforcement learning, which shows the impact incremental learning has on network attention.

Traditional recommendation systems rely on past usage data in order to generate new recommendations. Those approaches fail to generate sensible recommendations for new users and items into the system due to missing information about their past interactions. In this paper, we propose a solution for successfully addressing item-cold start problem which uses model-based approach and recent advances in deep learning. In particular, we use latent factor model for recommendation, and predict the latent factors from item's descriptions using convolutional neural network when they cannot be obtained from usage data. Latent factors obtained by applying matrix factorization to the available usage data are used as ground truth to train the convolutional neural network. To create latent factor representations for the new items, the convolutional neural network uses their textual description. The results from the experiments reveal that the proposed approach significantly outperforms several baseline estimators.

We propose Sparse Neural Network architectures that are based on random or structured bipartite graph topologies. Sparse architectures provide compression of the models learned and speed-ups of computations, they can also surpass their unstructured or fully connected counterparts. As we show, even more compact topologies of the so-called SNN (Sparse Neural Network) can be achieved with the use of structured graphs of connections between consecutive layers of neurons. In this paper, we investigate how the accuracy and training speed of the models depend on the topology and sparsity of the neural network. Previous approaches using sparcity are all based on fully connected neural network models and create sparcity during training phase, instead we explicitly define a sparse architectures of connections before the training. Building compact neural network models is coherent with empirical observations showing that there is much redundancy in learned neural network models. We show experimentally that the accuracy of the models learned with neural networks depends on expander-like properties of the underlying topologies such as the spectral gap and algebraic connectivity rather than the density of the graphs of connections.

The early layers of a deep neural net have the fewest parameters, but take up the most computation. In this extended abstract, we propose to only train the hidden layers for a set portion of the training run, freezing them out one-by-one and excluding them from the backward pass. Through experiments on CIFAR, we empirically demonstrate that FreezeOut yields savings of up to 20% wall-clock time during training with 3% loss in accuracy for DenseNets, a 20% speedup without loss of accuracy for ResNets, and no improvement for VGG networks. Our code is publicly available at https://github.com/ajbrock/FreezeOut

We propose a new approach to the problem of neural network expressivity, which seeks to characterize how structural properties of a neural network family affect the functions it is able to compute. Our approach is based on an interrelated set of measures of expressivity, unified by the novel notion of trajectory length, which measures how the output of a network changes as the input sweeps along a one-dimensional path. Our findings can be summarized as follows: (1) The complexity of the computed function grows exponentially with depth.

This blog was coauthored by Lewis Fishgold and Rob Emanuele. Aerial and satellite imagery gives us the unique ability to look down and see the earth from above. It is being used to measure deforestation, map damaged areas after natural disasters, spot looted archaeological sites, and has many more current and untapped use cases. At Azavea, we understand the potential impact that imagery can have on our understanding of the world. We also understand that the enormous and ever-growing amount of imagery presents a significant challenge: how can we derive value and insights from all of this data? There are not enough people to look at all of the images all of the time.

Basically, each X(t n) consists of a full set of connections that are input at that particular timestep of the sequence. Also not shown are the fact that each gate and cell has it's own set of weights and biases for both the input and recurrent connections. Thus, an LSTM actually has four sets of input and recurrent weight and bias parameters. In practice this means that usually the input is represented as a tensor with three dimensions (batch, timestep, input).

It also supports a rich set of higher-level tools including Spark SQL for SQL and structured data processing, MLlib for machine learning, GraphX for graph processing, and Spark Streaming. Scikit-learn (formerly scikits.learn) is a free software machine learning library for the Python programming language. Torch is a scientific computing framework with wide support for machine learning algorithms that puts GPUs first. Machine Learning for Language Toolkit (MALLET) is a Java toolkit fro statistical natural language processing, document classification, clustering, topic modeling and information extraction.

Thirty five years ago having a PhD in computer vision was considered the height of unfashion, as artificial intelligence languished at the bottom of the trough of disillusionment. Back then it could take a day for a computer vision algorithm to process a single image. "The competition for talent at the moment is absolutely ferocious," agrees Professor Andrew Blake, whose computer vision PhD was obtained in 1983, but who is now, among other things, a scientific advisor to UK-based autonomous vehicle software startup, FiveAI, which is aiming to trial driverless cars on London's roads in 2019. Blake founded Microsoft's computer vision group, and was managing director of Microsoft Research, Cambridge, where he was involved in the development of the Kinect sensor -- which was something of an augur for computer vision's rising star (even if Kinect itself did not achieve the kind of consumer success Microsoft might have hoped). He's now research director at the Alan Turing Institute in the UK, which aims to support data science research, which of course means machine learning and AI, and includes probing the ethics and societal implications of AI and big data. So how can a startup like FiveAI hope to compete with tech giants like Uber and Google, which are also of course working on autonomous vehicle projects, in this fierce fight for AI expertise?