Machine learning, deep learning, cognitive computing, robotic process automation (RPA), natural language processing (NLP), machine perception, predictive APIs, image recognition, speech recognition, virtual agent, intelligent assistant, personal advisor, chatbot, semantic search.
Did I miss anything? I am sure I did. However, I guess I provide a good list for your next round of artificial intelligence (AI) buzzword bingo.
Oh, one last thing – machine reasoning! If you’ve never heard about this term before, just read until the end and you will get its idea and importance for AI.
AI hits puberty but gives enterprises a new hope
In 1955 Prof. John McCarthy already defined AI as the goal to develop machines that behave as though they were intelligent. However, according to a Forrester survey after 62 years, most enterprises worldwide are still in an early stage. Around 60 percent researches on AI including market, solutions, platforms, vendors, skills and techniques. Further, 39 percent are in the phase of identifying and designing AI capabilities they can deploy, and 36 percent are educating the business or building the business case. Only a fifth (19 percent) is testing AI capabilities in their own environment and 14 percent are already training their deployed AI system.
However, enterprises see lot of potential in AI and its technologies as part of a strategic benefit for their organization. Most of them (57 percent) believe that AI will improve the customer experience and support. However, the more interesting part is that 43 percent believe that AI provides them with the ability to disrupt their industry with new business models, products and services. Further 42 percent think, that AI allows them to develop new products and services. I can’t agree more on the last two results mentioned, since several customers of ours already have started their AI journey. In doing so, they have started building an AI-enabled Enterprise based on a semantic data graph and the data and knowledge they hold within their entire enterprise stack.
Artificial intelligence in a nutshell: about smart machines and teaching children
Following Prof. McCarthy’s AI definition above, we are talking about a vigorous system.
- A system which must be considered as a raw IQ container
- A system that needs unstructured input to train its sense
- A system that needs a semantic understanding of the world to be able to take further actions
- A system that needs a detailed map of its context to act independently and transfer experience from one context to another
- A system that is equipped with all the necessities to develop, foster and maintain knowledge
And it is our responsibility to share our knowledge with these machines as we would share it with our children, spouses or colleagues. This is the only way to transform these machines, made of hard- and software, into a status we would describe as “smart”, helping them to become more intelligent by learning on a daily basis, building the groundwork to create a self-learning system.
It is kind of rude to compare raising a child with teaching a machine. However, it follows basically the same principles. In 1950, Alan Turing in his paper “Computing Machinery and Intelligence” described the idea of teaching a machine with the essentials of teaching a child. He described three stages:
- The initial state of the mind (at birth)
- The education to which it has been subjected
- Other experience to which it has been subjected that are not to be described as education
Defining these steps of the process, Turing discussed whether it would be more reasonable to program a child’s mind and subject the child’s mind to a period of education afterwards. He compared a child to a brand-new notebook and thought that it would be much easier to program because of its simplicity.
Machine learning in a nutshell: jump into your data lake – again and again
Machine learning (ML) is a discipline where a program or system can dynamically alter its behavior based on the ever-changing data. Therefore, the system has the ability to learn without being explicitly programmed. In doing so, algorithms enable systems to make data-driven decisions or predictions by building a model from sample inputs. A system then simply does not just memorize the samples but recognizes patterns and regularities within.
The goal of ML algorithms is to find specific patterns in (large) data sets. However, the supreme discipline is to find the right patterns in all related data sources since random patterns can be simply found everywhere. According to Crisp Research analyst Bjoern Boettcher the most common used algorithms right now are:
- Regression algorithms
- Instance-based algorithms
- Decision tree algorithms
- Bayesian algorithms
- Clustering algorithms
- Artificial neural network algorithms
- Deep learning
- Dimensionality reduction
Once an algorithm has successfully identified a reasonable pattern, further algorithms respectively mathematic procedures can be used to create a new subset of data and identify new patterns. Thus, the entire system is optimizing the existing knowledge or “learning”. In general, four types of learning are distinguished:
- Supervised learning
- Unsupervised learning
- Reinforcement learning
- Semi-supervised learning
Facebook’s News Feed is a good example for machine learning to personalize each member’s feed. Meaning, a member who frequently stops scrolling to read or like a certain post of a friend will see more of that friend’s activity.
So far, the biggest market of the AI universe seems to be machine learning. At Arago we easily have identified over 100 companies offering solutions and services, including cloud companies like Amazon Web Services, Microsoft Azure or Google. But also, smaller companies as well as start-ups are going to try their luck. Ergo, what has started as a blue ocean has quickly turned into a red ocean where the differentiation just turns out in minor parts respectively in the hidden algorithms implemented in the back-ends.
Bottom line, machine learning helps to identify patterns within data sets and thus tries to make predictions based on the existing data. However, most important is to check the plausibility and correctness of the results since you can always find something in endless sets of data. And that’s also one of the drawbacks if you consider machine learning as a single concept. Machine learning needs lots of sample data or data in general to learn and be able to find valuable information respectively results in patterns. A fact, Jerry Kaplan highlights as one crucial drawback saying that machine learning is not useful in situations where “[…] there’s no data, just some initial conditions, a bunch of constrains, and one shot to get it right.”
So, machine learning is basically like jumping into your data lake of endless waters again and again fishing for the next big catch.
Machine reasoning in a nutshell: teaching the machine with human experience
Machine reasoning (MR) systems generate conclusions from available knowledge by using logical techniques like deduction and induction. Thus, machine reasoning systems build the foundation for knowledge-based environments. Reasoning expert Léon Bottou defines [machine] reasoning as an “algebraically manipulating previously acquired knowledge in order to answer a new question”. However, reasoning systems come in different approaches that vary in expressive power, in predictive abilities as well as computational requirements. Bottou classifies seven types of approaches:
- First order logic reasoning
- Probabilistic reasoning
- Causal reasoning
- Newtonian mechanics
- Spatial reasoning
- Social reasoning
- Non-falsifiable reasoning
Everyone who wants to get a scientific perspective on Machine Reasoning I recommend reading the Léon Bottou’s paper “From Machine Learning to Machine Reasoning”.
Kaplan describes reasoning systems as a concept that deconstructs “[…] tasks requiring expertise into two components: “knowledge base” – a collection of facts, rules and relationships about a specific domain of interest represented in symbolic form – and a general-purpose “inference engine” that described how to manipulate and combine these symbols.” As one of the biggest advantages of reasoning systems Kaplan states that based on facts and rules those kinds of systems can be modified more easily since new facts and knowledge are incorporated. In doing so, reasoning systems are taught by “knowledge engineers” who interview practitioners and “[…] incrementally incorporating their expertise into computer programs […]”. This structure makes it also much more convenient to explain the reasoning to the system.
How does a sophisticated machine reasoning system look like today?
Talking reasoning systems today, the abilities and thus requirements differ from the ones described by Bottou and Kaplan above. Today, an AI technology based on a sophisticated machine reasoning system has the characteristics to empower a system
- to learn on its own
- to find solutions on its own
- to discover the world on its own
- to understand the world based on concepts (ontology)
The ontology can be explained by how children learn a language. They do learn by listening and then being taught sentences in school together with the right grammar. The ontology is taught by people. People define things for the ontology that should define a common language. And thus, the machine can work with that language.
To create a knowledge pool for an AI system, experts need to teach the AI with their contextual knowledge that includes the what, when, where and why. They must teach the AI with atomic pieces that can be prioritized by the AI. Context and indexing enable these atomic pieces to be combined to form many solutions afterwards.
To achieve the three steps above, a today’s sophisticated machine reasoning system is built on four pillars:
- Learning: First, a system has to be taught. This can be done by single experts or a community is used where people teach the machine bits of knowledge. This is what the machine uses to be able to learn on its own. You might think this way it doesn’t learn on its own, but it does. Consider how a child learns. It learns by being taught by his parents, teacher, other children or anyone else teaching things and it just copies and pastes everything with its “sensors” like ears and eyes. Thus, the AI learns best practices and reasoning from experts. Knowledge is taught in atomic pieces of information that represent individual steps of a process.
- Semantic Graph: The taught knowledge has to be stored, which is done within a data store. The store is used to supply information for the understanding of the world doing semantic reasoning. Like: I know that my mom is connected to dad. And I am connected to my sister. And my sister is connected to her work colleagues. And she works in this city in that building. This is a semantic map of the world that we know. That is part of our memory – a semantic graph. By creating a semantic data map, the AI understands the world in which it operates.
- Process Engine: The engine is the central back-end service that puts everything together and thus delivers a solution to a certain problem. The engine knows the map of the world where a system is acting in. In doing so, the engine takes everything it knows and finds the correct solution to a specific problem on its own, step by step based on the knowledge it has.
- Problem Solving: Problem solving also known as machine reasoning (MR) is the ability to dynamically react to change and by doing this, reusing existing knowledge for new and unknown problems. With machine reasoning, problems are solved in ambiguous and changing environments. The AI dynamically reacts to the ever-changing context, selecting the best course of action. Thus, machine reasoning is the basis for a general artificial intelligence (General AI).
Best of both worlds: machine reasoning optimized by machine learning
So, after all, why is machine learning just a fancy plugin that helps you to get results out of tons of data but also lets you jump into it again and again?
With machine learning you will never be able to adapt to change, which is what every company is looking for. Because change equals innovation! Thus, we consider machine learning as a mathematic optimization technique, which is fully optional. Talking about a decision-making process, everything works correctly without machine learning. Thus, the machine will find a solution on its own. Machine learning can be used to make the way to the solution shorter or more efficient by applying or selecting better knowledge. That’s what machine learning is used for.
In our case, machine learning classifies the atomic knowledge pieces in the situation of a certain problem and prioritizes and chooses the better suited pieces to provide the best solution. Thus, machine learning helps to select the best knowledge to a specific state of a problem.
Thus, machine learning as well as deep learning never tells you what, when, where and why a system has solved a problem or has done the decision the way it did. The technologies and algorithms behind are like a black box and you will never get the reason, just a result.
Jerry Kaplan summarizes the pro and cons of machine reasoning vs. machine learning as “[…] symbolic reasoning is more appropriate for problems that require abstract reasoning, while machine learning is better for situations that require sensory perception or extracting patterns from noisy data.”
Of course, you must identify which approach fits best for your specific situation. Or in Jerry Kaplan’s words “[…] if you have to stare at a problem and think about it, a symbolic reasoning approach is probably more appropriate. If you look at lots of examples or play around with the issues to get a “feel” for It, machine learning is likely to be more effective.”
By the way, if you want to read probably the best book on artificial intelligence on the market right now, get Jerry Kaplan’s “Artificial Intelligence: What everyone needs to know.”