AI: “It is an astonishing time to be a technologist.”
Want to master artificial intelligence (AI) techniques? A new book, The Supervised Learning Workshop, teaches you how to create machine-learning models using the Python programming language. A conversation with the co-author, Blaine Bateman.
Blaine Bateman is a business strategy consultant, helping companies identify growth strategies and opportunities.
Several years ago he decided to focus on data analysis or, more accurately, predictive analytics using machine learning.
“I started to see that clients had lots of data, frequently they didn’t know anything about it and they weren’t using it,” he says. “At the same time, I started to see that AI and machine learning were really on the uptick.”
Machine learning work is also rewarding, he says: “You build stuff and when you get it to work, you do something that helps someone.”
But it is not all fun: there is a lot of “data wrangling”, preparatory work to get the data ready for modelling.
First, the data may need to be integrated if it comes from several sources, and it may need to be scaled. It also pays to study the data, to discover as much as possible about it before modelling. All this takes time.
“Everyone likes the idea that you shovel data into a machine-learning black box and insights come out, but it is not that simple,” says Bateman.
Coming of age
AI and machine learning are terms commonly always mentioned together although machine learning is, in fact, a subset of AI.
There is also no real intelligence here, says Bateman. Machine learning, or what he calls predictive analytics, is the application of tools, algorithms and methodologies to train models.
“That is the learning part and it is using machine – a computer,” he says. “AI sounds a lot cooler but the vast majority of times you see the two, it is one and the same thing.”
AI is also not a new topic: neural networks, genetic algorithms and fuzzy logic were the subjects of much attention in the 1980s and ’90s. But developments in recent years has caused AI to finally hit its stride.
One factor is the maturity of silicon for AI, another is the advent of cloud computing. Bateman also highlights how the AI and machine-learning community embraced the open-source software movement. “It means there is a tremendous amount of commercial-scale work being done using open-source software,” he says.
Google’s TensorFlow, an example of open-source software, is one of the most used libraries for neural networks, while Keras is a software layer that sits on top, simplifying the use of TensorFlow in a Python environment.
“Coding languages such as Python and R have been around a long time and, with the open-source movement, these have grown and been extended into incredibly capable platforms,” says Bateman.
Another important catalyst for AI has been the development of e-learning, or massively open online courses (MOOC).
People used to go to college, learn a skill, enter industry and learn on the job before developing a career. “Now, people are jump-starting that,” he says.
There is an entire industry of people using online courses to learn as fast as possible. “Which is the reason for the books like the one I’ve participated on,” says Bateman.
Supervised learning
Supervised learning refers to data which has a characteristic or property that is known in advance, also referred to as labelled data.
Examples of such data could be numerous images of car registration plates to train an automated road-tolling system, or labelled images of lung abnormalities to train a medical scanning tool.
“We train a model by giving it a picture and telling it the answer,” says Bateman. “Then, once the model is built, you can translate a new picture into an answer.”
There is also unsupervised learning which refers to another aspect of machine learning. Here, data is applied to a clustering algorithm, for example, the MNIST handwritten digits database used to train algorithms to recognise ZIP or postcodes.
The MNIST database can be used for supervised learning, training a model to recognise each of the digits. But in unsupervised learning, the algorithm segregates the digits into clusters without being told what they are.
There are sophisticated clustering methods such as the uniform manifold approximation and projection (UMAP) approach that can reduce complex data sets into smaller dimensions. “It can take up to 80 dimensions and project them onto three and oftentimes find meaningful patterns,” he says.
Yet so far unsupervised learning is not used that much whereas supervised learning accounts for over 80 per cent of all machine learning applications used today, says Bateman.
Book
Packt Publishing wanted to issue a new edition of its machine learning book that included the latest supervised-learning practices. The publisher approached Bateman after seeing his work in the open-source community.
The resulting book – The Supervised Learning Workshop – is aimed at undergraduates and engineers. “Since it jumps into supervised learning, the expectation is that you have some coding skills and know enough Python to work through the exercises,” says Bateman.
The book uses Jupyter Notebooks, an open-source web application for developing Python code. “A lot of people use it to do small projects,” he says.
The topics addressed in the book can all be run using a decent laptop. “A huge amount of people working in AI are working on laptops; it is definitely doable with today’s technology,” he says.
And for larger data sets and bigger problems, there is always the cloud service providers such as Amazon Web Services, Google, Microsoft and others.
After a short introduction covering supervised and unsupervised learning, the book starts with linear regression, which remains an extremely important tool in machine learning.
His advice to students is to build a linear model first and see how well it performs. “That gives you a baseline,” says Bateman.
The topic of gradient-descent is then introduced, a technique used to train more sophisticated algorithms such as neural networks. Further into the book, more sophisticated techniques are introduced.
“The most sophisticated thing we talk about is ways to combine these algorithms into ensembles,” says Bateman.
Ensembling refers to using several less powerful models – what Bateman calls ‘weak learners’ – that are combined in some way, their results may be averaged or their outputs voted on.
Ensembling provides superior results compared with using a single model, even a sophisticated one
Bateman feels lucky to be working in the field of machine learning.
“We have this explosion of freely-available technology that you can use on a laptop to solve amazing problems,” he says. “It is an astonishing time to be a technologist.”
Using an open-source model to spur AI adoption
The Linux Foundation’s (LF) Deep Learning Foundation has set itself the ambitious goal of providing companies with all the necessary artificial intelligence (AI) software they will need.
Eyal Felstaine“Everything AI, we want you to take from open source,” says Eyal Felstaine, a member of the LF Deep Learning governing board and also the CTO of Amdocs. “We intend to have the entire [software] stack.”
The Deep Learning Foundation is attracting telecom, large-scale data centre operators and other players. Orange, Ciena, Red Hat, Chinese ride-sharing firm, Didi, and Intel are the latest companies to join the initiative.
The Deep Learning Foundation’s first project is Acumos, a platform for developers to build, share and deploy AI applications. Two further projects have since been added: Angel and Elastic Deep Learning.
Goal
The 'democratisation of data' is what has motivated the founding of the deep-learning initiative, says Felstaine.
A company using a commercial AI platform must put its data in a single repository. “You are then stuck [in that environment],” says Felstaine. Furthermore, fusing data from multiple sources exacerbates the issue in that the various datasets must be uploaded to the one platform.
Using an open-source approach will result in AI software that companies can download for free. “You can run it at your own place and you are not locked into any one vendor,” says Felstaine.
Everything AI, we want you to take from open source
Deep learning, machine learning and AI
Deep learning is associated with artificial neural networks which is one way to perform machine learning. And just as deep learning is a subset of machine learning, machine learning is a subset of AI, albeit the predominant way AI is undertaken today.
“Forty years ago if you had computer chess, the program’s developers had to know how to play chess,” says Felstaine. “That is AI but it is not machine learning.”
With machine learning, a developer need not know the rules of chess. “The software developer just needs to get the machine to see enough games of chess such that the machine will know how to play,” says Felstaine.
A neural network is composed of interconnected processing units or neurons. Like AI, it is a decades-old computer science concept. But an issue has been the efficient execution of a neural network when shared across processors due to input-output constraints. Now, with the advent of the internet content providers and cloud, not only can huge datasets be used to train neural networks but the ‘hyper-connectivity’ between the servers’ virtual machines or containers means large-scale neural networks can be used.
Containers offer a more efficient way to run many elements on a server. “The numbers of virtual machines on a CPU is maybe 12 if you are lucky; with containers, it is several hundred,” says Felstaine. Another benefit of developing an application using containers is that it can be ported across different platforms.
“This [cloud clustering] is a quantitative jump in the enabling technology for traditional neural networks because you can now have thousands and even tens of thousands of nodes [neurons] that are interconnected,” says Felstaine. Running the same algorithms on much larger neural networks has only become possible in the last five years, he says.
Felstaine cites as an example the analysis of X-ray images. Typically, X-rays are examined by a specialist. For AI, the images are sent to a firm for parsing where the images are assessed and given a ‘label’. Millions of X-ray images can be labelled before being fed to a machine-learning application such as Tensorflow or H2O. Tensorflow, for example, is open-source software that is readily accessible.
The resulting trained software, referred to as a predictor, is then capable of analysing an X-ray picture and give a prognosis based on what it has learnt from the dataset of X-rays and labels created by experts. “This is pure machine learning because the person who defined Tensorflow doesn’t know anything about human anatomy,” says Felstaine. Using the software creates a model. “It’s an empty hollow brain that needs to be taught.”
Moreover, the X-ray data could be part of a superset of data from several providers such as life habits from a fitness watch, the results of a blood test, and heart data to create a more complex model. And this is where an open-source framework that avoids vendor lock-in has an advantage.
Acumos
Acumos started as a collaboration between AT&T and the Indian IT firm, Tech Mahindra, and was contributed to the LF Deep Learning Foundation.
Felstaine describes Acumos as a way to combine, or federate, different AI tools that will enable users to fuse data from various sources "and make one whole out of it”.
There is already an alpha release of Acumos and the goal, like other open-source projects, is to issue two new software releases a year.
How will such tools benefit telecom operators? Felstaine says AT&T is already using AI to save costs by helping field engineers maintain its cell towers. The field engineer uses a drone to inspect the operator’s cell towers, and employing AI to analyse the drone’s images, it guides the field engineer as to what maintenance, if any, is needed.
One North American operator has said it has over 30 AI projects including one that is guiding the operator as to how to upgrade a part of its network to minimise the project's duration and the disruption.
One goal for Acumos is to benefit the Open Networking Automation Platform (ONAP) that oversees Network Functions Virtualisation (NFV)-based networks. ONAP is an open-source project that is managed by the Linux Foundation Networking Fund.
NFV is being adopted by operators to help them lunch and scale services more efficiently and deliver operational and capital expenditure savings. But the operation and management of NFV across a complex telecom network is a challenge to achieving such benefits, says Felstaine.
ONAP already has a Data Collection, Analytics, and Events (DCAE) subsystem which collects data regarding the network’s status. Adding Acumos to ONAP promises a way for machine learning to understand the network’s workings and provide guidance when faults occur, such as the freezing of a virtual machine running a networking function.
With such a fault, the AI could guide the network operations engineer, pointing out that humans take this action next and that the action has an 85 percent success rate. It then gives the staff member the option to proceed or not. Ultimately, AI will control the networking actions and humans will be cut out of the loop. “AI as part of ONAP? That is in the future,” says Felstaine.
The two new framework projects - Angel and Elastic Deep Learning - have been contributed to the Foundation from the Chinese internet content providers, Tencent and Baidu, respectively.
Both projects address scale and how to do clustering. “They are not AI, more ways to distribute and scale neural networks,” says Felstaine.
The Deep Learning Foundation was launched in March by the firms Amdocs, AT&T, B.Yond, Baidu, Huawei, Nokia, Tech Mahindra, Tencent, Univa, and ZTE.
