Which President served as a lieutenant colonel in the Spanish-American war? Who was the first Democrat elected after the Civil War? Who introduced Social Security? If you’re looking to learn more about the past Presidents who have led our country, you’re in the right place. Take a look at our full set of biographies. Then, quiz your friends.
Take your machine learning to the next level with these artificial intelligence technologies.
Artificial
intelligence (AI) technologies are quickly transforming almost every sphere of
our lives. From how we communicate to the means we use for transportation, we
seem to be getting increasingly addicted to them.
Because
of these rapid advancements, massive amounts of talent and resources are
dedicated to accelerating the growth of the technologies.
Here
is a list of 8 best open source AI technologies you can use to take your
machine learning projects to the next level.
1. TensorFlow
Initially
released in 2015, TensorFlow is an open source machine
learning framework that is easy to use and deploy across a variety of
platforms. It is one of the most well-maintained and extensively used
frameworks for machine learning.
Created
by Google for supporting its research and production objectives, TensorFlow is
now widely used by several companies, including Dropbox, eBay, Intel, Twitter,
and Uber.
TensorFlow
is available in Python, C++, Haskell, Java, Go, Rust, and most recently,
JavaScript. You can also find third-party packages for other programming
languages.
The
framework allows you to develop neural networks (and even other computational
models) using flowgraphs.
2. Keras
Initially
released in 2015, Keras is an open source software library
designed to simplify the creation of deep learning models. It is written in
Python and can be deployed on top of other AI technologies such as TensorFlow,
Microsoft Cognitive Toolkit (CNTK), and Theano.
Keras
is known for its user-friendliness, modularity, and ease of extensibility. It
is suitable if you need a machine learning library that allows for easy and
fast prototyping, supports both convolutional and recurrent networks, and runs
optimally on both CPUs (central processing units) and GPUs (graphics processing
units).
3. Scikit-learn
Initially
released in 2007, scikit-learn is an open source
library developed for machine learning. This traditional framework is written
in Python and features several machine learning models including
classification, regression, clustering, and dimensionality reduction.
Scikit-learn
is designed on three other open source projects—Matplotlib, NumPy, and
SciPy—and it focuses on data mining and data analysis.
4. Microsoft Cognitive Toolkit
Initially
released in 2016, the Microsoft Cognitive Toolkit (previously
referred to as CNTK), is an AI solution that can empower you to take your
machine learning projects to the next level.
Microsoft
says that the open source framework is capable of "training deep learning
algorithms to function like the human brain."
Some
of the vital features of the Microsoft Cognitive Toolkit include highly
optimized components capable of handling data from Python, C++, or BrainScript,
ability to provide efficient resource usage, ease of integration with Microsoft
Azure, and interoperation with NumPy.
5. Theano
Initially
released in 2007, Theano is an open
source Python library that allows you to easily fashion various machine
learning models. Since it's one of the oldest libraries, it is regarded as an
industry standard that has inspired developments in deep learning.
At
its core, it enables you to simplify the process of defining, optimizing, and
assessing mathematical expressions.
Theano
is capable of taking your structures and transforming them into very efficient
code that integrates with NumPy, efficient native libraries such as BLAS,
and native code (C++).
Furthermore,
it is optimized for GPUs, provides efficient symbolic differentiation, and
comes with extensive code-testing capabilities.
6. Caffe
Initially
released in 2017, Caffe (Convolutional
Architecture for Fast Feature Embedding) is a machine learning framework that
focuses on expressiveness, speed, and modularity. The open source framework is
written in C++ and comes with a Python interface.
Caffe's
main features include an expressive architecture that inspires innovation,
extensive code that facilitates active development, fast performance that
accelerates industry deployment, and a vibrant community that stimulates
growth.
7. Torch
Initially
released in 2002, Torch is a machine learning library that
offers a wide array of algorithms for deep learning. The open source framework
provides you with optimized flexibility and speed when handling machine
learning projects—without causing unnecessary complexities in the process.
It
is written using the scripting language Lua and comes with an underlying C
implementation. Some of Torch's key features include N-dimensional arrays,
linear algebra routines, numeric optimization routines, efficient GPU support,
and support for iOS and Android platforms.
8. Accord.NET
Initially
released in 2010, Accord.NET is a machine learning
framework entirely written in C#.
The
open source framework is suitable for production-grade scientific computing.
With its extensive range of libraries, you can build various applications in
artificial neural networks, statistical data processing, image processing, and
many others.
Conclusion
Before
starting to build a machine learning application, selecting one technology from
the many options out there can be a difficult task. Therefore, it's important
to evaluate several options before making a final decision.
Apart
from the above-listed AI technologies in machine learning, which are you using
in your projects? Is there any other framework, library, or toolkit not
discussed? Please let us know in the comment section below.
have been talking and writing about AI since a while now, so I will not waste any time defining what it is and what is not (if you want to know more about it, you can check my explanation or a brief history of AI).
However, I never touched upon blockchain and cryptocurrencies so far, so I will dedicate this first block to describe what it is and how it works.
A blockchain is a secure distributed immutable database shared by all parties in a distributed network where transaction data can be recorded (either on-chain for basic information or off-chain in case of extra attachments) and easily audited.
A blockchain is a secure distributed immutable database shared by all parties in a distributed network
Put simply (with Bank of England’s words), the blockchain is “a technology that allows people who don’t know each other to trust a shared record of events”.
The data are stored in rigid structures called blocks, which are connected to each other in a chain through a hash (each block also includes a timestampand a link to the previous block via its hash). The blocks have a header, which includes metadata, and a content, which includes the real transaction data. Since every block is connected to the previous one, as the number of participants and blocks grow, it is extremely hard to modify any information without having the network consensus.
The network can validate the transaction through different mechanisms, but mainly through either a “proof-of-work” or a “proof-of-stake”. A proof-of-work(Nakamoto, 2008) asks the participants (called “miners”) to solve complex mathematical problems in order to add a block, which in turn require a ton of energy and hardware capacity to be decoded. A proof-of-stake (Vasin, 2014) instead tries to solve this energy efficiency issue attributing (roughly) more mining power to participants who own more coins (there are many variations of it and some skepticism around its famous “nothing at stake” problem — see Buterin’s blog post to know more on this).
Additional mechanisms are the Byzantine-fault-tolerant algorithm (Castro and Liskov, 2002), the Quorum slicing (Mazieres, 2016), as well as variations of the Proof-of-stake (Mingxiao et al., 2017), but we will not get into those now.
The final characteristic that needs to be explained is the category of blockchain based on the different network access permission, i.e., whether it is free for anyone to view it (permissionless vs permissioned) or to participate in the consensus formation (public vs private). In the former case, anyone can access and read or write data from the ledger, while in the latter one predetermined participants have the power to join the network (and of course only in the public permissionless case a reward structure for miners has been designed).
It should be clear by now the intrinsic power of this technology, which is not simply a disruptive innovation but rather a foundational technology that aims to “change the scope of intermediation” (Catalini and Gans, 2017). Distributed ledger technologies will indeed reduce both the costs of verification and networking, influencing then the market structure and eventually allowing the creation of new marketplaces. Iansiti and Lakhani (2017) also drew a brilliant parallel between blockchain and TCP/IP in a recent work (which I highly recommend), showing how blockchain is slowly going through the four phases that identify previous foundational technologies such as the TCP/IP, i.e., single-use, localized use, substitution, and transformation. As they explained, the “novelty” of such a technology makes it harder for people to understand the solution domain, while its “complexity” requires a larger institutional change to foster an easy adoption.
However, it is also true that the blockchain is shifting the traditional business models distributing value in an opposite way with respect to previous stacks: if it made more sense to invest in applications rather than protocol technologies fifteen years ago, in a blockchain world the value is concentrated in the shared protocol layer and only marginally at the application level (see the “Fat Protocol” theory by Joel Monegro).
It’s a stack with “fat” protocols and “thin” applications (Joel Monegro).
To conclude this introductory section, I will just mention on the fly the possibility for the blockchain to not simply allow for transactions but also the possibility to create (smart) contracts that are triggered by specific events and threshold and that are traceable and auditable without effort.
Bonus Paragraph: Initial Coin Offerings (ICOs)
A big hype is nowadays surrounding this new phenomenon of the Initial Coin Offerings (ICOs). Even if many people are pouring money into that because of its resemblance to the most common (and valuable) Initial Public Offerings (IPOs), an ICO is nothing more than a token sale, where a token is the smallest functional unit of a specific network (or application).
ICOs experts (if any) will forgive my approximate definition, but an ICO is a hybrid concept that has elements of a shares allocation, a pre-sales/crowdfunding campaign, and a currency with a limited power and application’s domain.
It is definitely an interesting innovation that introduces new unregulated ways to raise capitals, but it also poses several issues to an unprepared community. I am happy to receive feedback on this, but I would distill the key points of an ICO evaluation in what follows:
a token has an additional utility with respect to the exchange of value and companies selling token with the only goal of raising capital are sending a bad signal to the market. Tokens are needed to create a users’ base and to incentivize stakeholders to participate in the ecosystem at the earliest stage. A good white paper is not enough;
Be wary of token sales that are uncapped;
Be wary of token sales that have no time limit;
Be wary of token sales that do not clearly state the (present and future) number as well as the value of the token (it could sound absurd, but you may be surprised of how non-transparent an ICO can look like).
II. How AI can change Blockchain
A
lthough extremely powerful, a blockchain has its own limitations as well. Some of them are technology-related while others come from the old-minded culture inherited from the financial services sector, but all of them can be affected by AI in a way or another:
Energy consumption: mining is an incredibly hard task that requires a ton of energy (and then money) to be completed (O’Dwyer and David Malone, 2014). AI has already proven to be very efficient in optimizing energy consumption, so I believe similar results can be achieved for the blockchain as well. This would probably also result in lower investments in mining hardware;
Scalability: the blockchain is growing at a steady pace of 1MB every 10 minutes and it already adds up to 85GB. Satoshi (2008) first mentioned “blockchain pruning” (i.e., deleting unnecessary data about fully spent transactions in order to not hold the entire blockchain on a single laptop) as a possible solution but AI can introduce new decentralized learning systems such as federated learning, for example, or new data sharding techniques to make the system more efficient;
Security: even if the blockchain is almost impossible to hack, its further layers and applications are not so secure (e.g., the DAO, Mt Gox, Bitfinex, etc.). The incredible progress made by machine learning in the last two years makes AI a fantastic ally for the blockchain to guarantee a secure applications deployment, especially given the fixed structure of the system;
Privacy: the privacy issue of owning personal data raises regulatory and strategic concerns for competitive advantages (Unicredit, 2016). Homomorphic encryption (performing operations directly on encrypted data), the Enigma project (Zyskind et al., 2015) or the Zerocash project(Sasson et al., 2014), are definitely potential solutions, but I see this problem as closely connected to the previous two, i.e., scalability and security, and I think they will go pari passu;
Efficiency: Deloitte (2016) estimated the total running costs associated with validating and sharing transactions on the blockchain to be as much as $600 million a year. An intelligent system might be eventually able to compute on the fly the likelihood for specific nodes to be the first performing a certain task, giving the possibility to other miners to shut down their efforts for that specific transaction and cut down the total costs. Furthermore, even if some structural constraints are present, a better efficiency and a lower energy consumption may reduce the network latency allowing then faster transactions;
Hardware: miners (and not necessarily companies but also individuals) poured an incredible amount of money into specialized hardware components. Since energy consumption has always been a key issue, many solutions have been proposed and much more will be introduced in the future. As soon as the system becomes more efficient, some piece of hardware might be converted (sometimes partially) for neural nets use (the mining colossus Bitmain is doing exactly this);
Lack of talent: this is leap of faith, but in the same way we are trying to automate data science itself (unsuccessfully, to my current knowledge), I don’t see why we couldn’t create virtual agents that can create new ledgers themselves (and even interact on it and maintain it);
Data gates: in a future where all our data will be available on a blockchain and companies will be able to directly buy them from us, we will need help to grant access, track data usage, and generally make sense of what happens to our personal information at a computer speed. This is a job for (intelligent) machines.
Artificial Intelligence Exactly
III. How Blockchain can change AI
In
the previous section, we quickly touched upon the effects that AI might eventually have on the blockchain. Now instead, we will make the opposite exercise understanding what impact can the blockchain have on the development of machine learning systems. More in details, blockchain could:
Help AI explaining itself (and making us believe it): the AI black-box suffers from an explainability problem. Having a clear audit trail can not only improve the trustworthiness of the data as well as of the models but also provide a clear route to trace back the machine decision process;
Increase AI effectiveness: a secure data sharing means more data (and more training data), and then better models, better actions, better results…and better new data. Network effect is all that matter at the end of the day;
Lower the market barriers to entry: let’s go step by step. Blockchain technologies can secure your data. So why won’t you store all your data privately and maybe sell it? Well, you probably will. So first of all, blockchain will foster the creation of cleaner and more organized personal data. Second, it will allow the emergence of new marketplaces: a data marketplace (low-hanging fruit); a models marketplace (much more interesting); and finally even an AI marketplace (see what Ben Goertzel is trying to do with SingularityNET). Hence, easy data-sharing and new marketplaces, jointly with blockchain data verification, will provide a more fluid integration that lowers the barrier to entry for smaller players and shrinks the competitive advantage of tech giants. In the effort of lowering the barriers to entry, we are then actually solving two problems, i.e., providing a wider data access and a more efficient data monetization mechanism;
Increase artificial trust: as soon as part of our tasks will be managed by autonomous virtual agents, having a clear audit trail will help bots to trust each other (and us to trust them). It will also eventually increase every machine-to-machine interaction (Outlier Ventures, 2017) and transaction providing a secure way to share data and coordinate decisions, as well as a robust mechanism to reach a quorum (extremely relevant for swarm robotics and multiple agents scenarios). Rob May expressed a similar concept in one of his last newsletters
Reduce catastrophic risks scenario: an AI coded in a DAO with specific smart contracts will be able to only perform those actions, and nothing more (it will have a limited action space then).
In spite of all the benefits that AI will receive from an interaction with blockchain technologies, I do have one big question with no answer whatsoever.
