Artificial Intelligence and Machine Learning — What’s the difference?
So what’s the difference between Artificial Intelligence (AI) and Machine Learning (ML)? Everyone seems to be using these two phrases interchangeably. However in popular media, AI is being used synonymously with ML. Are they actually the same concept? Do they refer to the same thing? Let me explain.
Technically, the answer is no. Let’s go back a few decades to get a bit of context. Artificial Intelligence concerns itself with all things that has got to do with making computing systems intelligent. Since the early days of AI (say 1950s), it has comprised of many tasks— recognising images, understanding language, solving theorems, playing games like chess, providing expert advice to name a few. AI, in those days, was defined as intelligent behavior that was easy for humans to perform but hard for computers to emulate. In order to solve these tasks, two major approaches evolved — rule based approach and machine learning approach.
Rule based approach was concerned with making machines intelligent using what is called as symbolic logic and search algorithms. In this approach, experts believed that intelligent behaviour can be created artificially by representing the objects in the world as symbols and manipulating them computationally. For instance, there were systems called Expert Systems that emulated and provided expert advice on advanced subjects like medicine. The knowledge required to provide advice were encoded in the form of rules. The following image shows examples of rules used in an expert system called MYCIN.
Rule based approaches were used in other domains of AI such as Natural Language Processing (NLP) to understand language. Two popular NLP systems that used rules and really showed promise were ELIZA, a chatbot that pretended to be a therapist and SHRDLU, a virtual robot that can take instructions in natural language and move objects in a virtual toy world. However they did not scale to the complexities of the real world.
On the other hand, machine learning approaches to AI were also explored during the same time but early approaches were very simplistic and did not perform as well as rule based methods. Early machine learning approaches included the development of artificial neural networks (ANNs) inspired by biological neural networks and algorithms like Perceptron that were used to train ANNs. These approaches were called the Connectionist approach to AI.
While early on, between 1950s and 1980s, symbolic AI approaches had an upper hand, they did not scale as promised. There was a period of time when government sponsored AI funding ceased due to lack of progress. This period was notoriously called the AI Winter. Rule based approaches could not scale up to real world problems. Cost of authoring rules went up as they involved significant manual effort. Even so, they were very brittle as they predominantly used binary logic (e.g. something is either true or false and the AI system cannot see shades of grey in the real world). On the other hand, Connectionist approaches could not perform well due to lack of computational power.
Early 1990s saw the revival of AI. As computational power of machines went up along with lowering cost of hardware, AI tasks became popular again. Deep Blue, the IBM supercomputer that defeated the world champion Gary Kasparov in a game of Chess could in 1997 evaluate 200 million chess positions per second. It used a search algorithm to search through possibilities of game states and decided the next best move. This was an example of GOFAI. However, other AI domains like NLP, image recognitions, etc predominantly started using machine learning techniques. Because rule based methods became old fashioned, they are now called GOFAI — Good Old Fashioned AI.
With availability of large amounts of data and computational power (i.e cloud computing, Graphical Processing Units (GPUs), etc), machine learning approches have come back into vogue. Development of deep neural networks and other deep learning architectures, that use more than one hidden layer of neurons, in particular have propelled the state of the art in AI tasks such as speech recognition, image recognition and recently many NLP and image recognition tasks as well.
- DeepMind’s AlphaGo, an AI game playing system defeated the human champion Lee Sedol in the game of Go using a deep learning technique called Deep Reinforcement Learning.
- Google’s Neural Machine Translation system reduced translation errors by about 55–85% compared to an earlier approach called Phrase based translation.
- DeepMind’s WaveNet uses Deep Neural Net to generate speech from text. It has been ranked as the most natural compared to earlier approaches.
- Google Voice Search is powered by a deep learning method called Long Short Term Memory (LSTM) which has shown to deliver the state of the art performance in speech recognition tasks.
- Algorithms like word2vec have been used to generate word embeddings that represented words in contextual vectors. These have significantly improved performance of a number of NLP tasks.
So, to get back to our question — are AI and ML different? Yes. They still are. While in 1950s — 1990s, AI tasks were predominantly rule based, AI and ML seemed to be two very different concepts. However these days, modern machine learning methods have shown to perform many times better than rule based methods in many AI tasks. Most AI systems built these days do comprise an element of ML. This is why AI and ML seem synonomous now.
Even so, AI is still considered to be an umbrella term comprising all sorts of intelligent behaviour that a computing system can emulate such as learning, planning, reasoning, etc. On the other hand, ML is considered to be a set of techniques/algorithms that enable machines to learn from data and experience. Let me leave you with a simple Venn diagram to show how they are related.
