Ten top frameworks for deep learning

By the end of 2015, it was time to look at 2016 years of technological trends, especially in the area of deep learning. At the end of 2015, the new unit sent an article, "will deep learning make machine learning engineers unemployed?" ", causing great repercussions. Indeed, in the past year, deep learning is changing more and more AI domains. Google DeepMind engineer Jack Rae predicts that those models that used to be considered the best predictive algorithm for medium-to-large datasets (such as raising decision trees (Boosted decision Trees) and random forests) would become irrelevant.

Deep learning, or more broadly, a machine learning algorithm that uses a linked architecture, may turn machine learning algorithms into the past, because deep learning algorithms are far from saturated. Over the next few years, there is a good chance that some deep neural networks will be trained to significantly improve performance. There are also some breakout spaces between the optimization method, the activation function, the junction structure, and the initialization steps.

This is likely to allow many machine learning algorithms to approach the edge of the exit.

So, can 2016 be seen as a year of deep learning to formally govern AI? If so, what can we do to prepare for it? New intelligence has compiled the industry's vision of deep learning technology for the 2016 and the 10 most popular frameworks for deep learning in the 2015.

Ten technology outlook of deep learning

Head of the research department of Iiya Sutskever:openai

We expect to see deeper models (deeper Models), which can learn from less data than today's models, especially in unsupervised learning. We can also expect to see more accurate and useful results in the areas of speech recognition and image recognition.

Sven Behnke: Full-time professor at the University of Bonn, director of Intelligent Systems Group

I look forward to deep learning techniques that will be applied to the growing number of multi-structured data issues. This brings new areas of application to deep learning, including robotics, data mining, and knowledge discovery.

Christian Szegedy:google Senior Engineer

The current deep learning algorithms and neural networks are far from the theoretical possible performance. Compared to a year ago, we now have a visual neural network model, its price is 5 to 10 times times cheaper, processing parameters less than 15 times times, but the performance is also better. Behind this is a better network structure and better training methods. I believe this is just the beginning, the deep learning algorithm will be so cheap, it can run on cheap mobile devices, and no more hardware device support, and no need for additional memory.

Andrej karpathy: Ph. D. In computer science, Stanford University, OpenAI research engineer

I see a trend, the structure tends to be bigger and more complex. We will build a very large neural network that can exchange neural network components, train some networks in advance, add new modules, and adjust all components. For example, convolutional neural networks were once the largest deep neural networks, but today they are separated as part of a new large neural network. Similarly, now these neural networks will also be part of a larger neural network in the new Year. We are learning the spelling of Lego toys and learn how to stitch them together efficiently.

Pieter Abbeel:uc, assistant professor at the University of Berkeley, Gradescope co-founder

Deep learning verticals that rely on supervisory technology require new methods (NLP) to exceed existing technical performance. We will see the outstanding performance of deep learning in non-supervised learning and enhanced learning.

Eli David:deep Instinct CTO

In the past two years, we have seen deep learning in various fields to achieve a great breakthrough. But even so, within 5 years it will not reach the holy grail of the human level (but I think this will happen in the end of life). We see great breakthroughs in all areas. In particular, I think the most promising areas come from unsupervised learning, and most of the world's data is untagged, and our brains themselves are also very good unsupervised learning boxes.

When deep instinct becomes the first company to use in-depth learning in the security field, it can be expected that more companies will also use deep learning to deploy. But the threshold for deep learning is still very high, especially for Internet security companies, where they don't actually use AI tools (only a few solutions use traditional machine learning techniques). So deep learning is going to be a big application in the security world, and it's going to take years.

Daniel Mcduff:affectiva Research Director

Deep learning has become a dominant form of machine learning in computer vision, speech analysis, and other fields. I want to use an accurate identification system that can be deployed from 1 to 2 GPUs, enabling developers to deploy new software to the real world. I hope more focus will be placed on unsupervised training or semi-supervised training algorithms.

Jörg Bornschein:google Scholar, Institute for Advanced Technology (CIFAR), Canada

It's always hard to predict the future. When we consider machine learning in large-scale systems, non-supervised, semi-supervised, and intensive learning plays an increasingly important role in the field of robotic control systems, or in the brain systems of large-scale systems. It is obvious that the simple method of supervised learning is too restrictive in theory and it is difficult to solve practical problems.

Ian Goodfellow:google Senior Research Engineer

I predict that over the next 5 years, our neural networks can summarize what's going on in the video and have the ability to generate short videos. Neural networks have become a standard solution for visual tasks. I predict that neural networks will be a standard solution for NLP and robotic tasks. I also predict that neural networks will play important tools in other areas of science, such as gene behavior prediction, drugs, proteins, new medical programs, and so on.

Koray Kavukcuoglu & Alex graves:google DeepMind Research Engineer

Many things will happen in the next 5 years. We anticipate that non-supervised learning and enhanced learning are increasingly important. We also predict the rise of multi-mode learning (multimodal learning) and will go beyond individual data sets for learning.

2015 deep Learning ten top-level frameworks

1.Keras

Keras is a very minimalist, highly modular neural network library, written in Python and can run on top of TensorFlow and Thenao. It is designed to achieve faster experimentation, with as little time as possible from idea to result, which is the key to doing research.

2.MXNet

A lightweight, portable, and flexible distributed/mobile deep learning system that can dispatch dynamic, mutated data streams. MXNet supports programming languages such as Python, R, Julia, Go, Javascript, and is a deep learning framework designed for efficiency and flexibility. It can add a little seasoning to the deep learning program and maximize product efficiency.

3.Chainer

Neural network flexible framework for deep learning. Chainer supports a variety of network architectures, including Feed-forward Nets, convnets, recurrent Nets, and Recursive Nets. It also supports the Per-batch architecture. Chainer supports CUDA computing, which requires only a few lines of code to drive the GPU. It can also be run in a multi-GPUs architecture through some effort.

4.sickit-neuralnetwork

Deep neural networks are implemented, and there is no learning cliff (learning Cliff). The library is capable of performing multilayer perceptron, automatic encoders and recurrent neural networks that run on a stable future Proof interface and are compatible with the user-friendly Scikit-learn and Python interface.

5.theano-lights

Theano-lights is a Theano-based research architecture that provides the implementation of some recent deep learning models and facilitates training and testing capabilities. These models are not hidden, but in the process of research and learning, there is a lot of transparency and flexibility.

6.Deeppy

Deep learning framework based on Theano's highly scalable nature.

7.Idlf

Intel's deep learning framework.

The Intel deep learning Framework (IDLF) is an SDK library that provides training and execution for deepin neural networks.

It includes APIs that enable the construction of neural network topologies as computational workflows, function graphics optimization, and execution to hardware. Our initial focus was to drive object recognition (ImageNet topologies) on neural networks deployed on the CPU (Xeon) and GPU (Gen).

This API is designed so that we can easily support more devices in the future. Our key principle is to achieve maximum performance on every Intel-supported platform.

8.Reinforcejs

Reinforcejs is an enhanced learning library capable of performing common enhancement learning algorithms, and can do Web-side Demos. This library now includes:

Dynamic planning method (programming Methods)

Time difference Learning (temporal difference Learning) (sarsa/q-learning)

Deep q-learning

Stochastic/deterministic Policy gradients and Actor critic architecture

9.OpenDeep

Opendeep is a deep learning framework for Python, based on Theano, focused on flexibility and ease of use, serving industry data scientists and leading researchers. Opendeep is a modular, extensible architecture that can be used to build almost any neural network framework to solve your problem.

10.MXNetJS

Mxnetjs is a dmlc/mxnet javasript bag. Mxnetjs can bring the latest level of deep learning prediction API to the browser. It runs through Emscripten and amalgamation. Mxnetjs allows you to run the latest level of deep learning predictions in a variety of computing images and to bring deep learning to the client.

Ten top frameworks for deep learning