Artificial Intelligence: A Guide for Thinking Humans, Melanie Mitchell

Prologue; Part I:  BACKGROUND

The goal of AI is, supposedly, to solve intelligence and use it to solve everything else.

The main hypothesis behind the creation of AI is that any aspect of learning or intelligence can be simulated by a machine. Brain is composed of cells, neurons, and matter that obey laws of physics.

Two problems associated with the development of AI are: (1) people are rushing blindly and deliriously to create AI without giving a thought on its implications, and (2) the trivialization of humanity (emotions, society, music, art, games, movies, literature).

Technological singularity is the point at which machine intelligence surpasses human intelligence, which leads to an intelligence explosion. This happens because an intelligent being is capable of creating an even more-intelligent being. Ray Kurzweil is a proponent of technological singularity and forecasts that singularity is imminent in a few decades. Ray Kurzweil attributes his prediction to exponential growth, Moore's law, neuroscience, virtual reality. He also predicts that a computer will pass the Turing Test by 2029. The two main arguments of his sceptics are: (1) intelligence does not reside solely in the brain; that all the senses associated with the human body are necessary for intelligence to operate (social intelligence, emotional intelligence), and (2) software has not shown any exponential progress.

PART II: COMPUTER VISION

ImageNet (~WordNet) is a dataset of images built with the help of annotators recruited using Amazon Mechanical Turk. AlexNet solved the task of object identification in ImageNet using convolutional neural-networks (ConvNet) and GPUs.

Some technical challenges to AI are (1) the long tail, (2) easy to fool (adversarial examples), (3) bias, (4) supervised learning is not how humans learn; unsupervised learning is the dark matter of AI, (5) data-hungry systems "big data", need a lot of labelled data, (6) inability to interpret the results "show your work", (7) spurious correlation "clever Hans", (8) rely on human understanding of the task in model design, hyperparameter tuning.

We already use AI systems in our day-to-day lives: recommender systems (Amazon, YouTube), web search engines (Google Search, Microsoft Bing), translators (Google Translate, DeepL), personal digital assistants (Alexa, Cortana, Siri). However there are some issues of dispute and concern. How to enforce ethics in AI? How to regulate development and deployment of AI especially in domains such as mass surveillance, face recognition, and arms race? Do we need to teach morality to machines "moral machines"?

Part III: LEARNING TO PLAY

Reinforcement Learning (RL)

Reinforcement learning is a type of machine learning in which the agent learns by itself through reinforcement (rewards) from the environment. Humans learn to ride the bicycle through a similar process. 

The primary approach in RL is to learn the Q-table through multiple episodes (one episode corresponds to a complete game of Breakthrough) and multiple iterations (one iteration corresponds to a single action such as one move). An evaluation function evaluates the current states and predicts the next move by exploiting and exploring the set of values for each possible action. These value of an action indicates the agent's estimate of the reward at the end of the episode if that action is chosen at a given state.

For domains with simple state description, computing the Q-table is easier.  The objective then is to compute values of each action at every possible state. However for complex domains such as self-driving cars which can have inputs from thousands of sensors, neural networks can be used to process the input and generate the state description. These neural networks can then group similar states together through its state representation.

Game Engines

Traditionally, game engines such as IBM Deep Blue (vs. Garry Kasparov, 1997) relied on game trees and pruning algorithms such as minimax. However generating the game tree in real-time for games such as Go proved to be difficult even with sophisticated computing power.

With the advent of deep learning, it has been shown that it is possible to learn the Q-table through the use of neural networks. These Deep Q-Networks (DQN) employed a technique called temporal difference learning, which is based on the hypothesis that the evaluation function becomes more accurate as the games proceeds. 

DeepMind developed AI game engines that relied on deep Q-learning on Atari games and Go (AlphaGo vs Lee Sedol, 2016) Specifically, AlphaGo employed a combination of techniques: convolutional neural networks, deep Q-learning, Monte Carlo tree search and powerful hardware.

Simulation to Reality

However it is difficult to transfer the knowledge gained by the agent to other domains (due to lack of understanding or capabilities of abstraction or transfer learning) or into the real-world (due to ambiguity, noise, unpredictability). These models are also vulnerable to adversarial examples.

While board games or Atari games are easy to solve by current AI systems, these domains are not so challenging. The simple game of dumb charades will be much more harder because the game requires visual, linguistic and social understanding.

Moreover, though it is claimed that the RL agents are capable of learning by themselves without any specialized domain knowledge, the agent receive some amount of guidance through the choice of model architecture and hyperparameters.

PART IV: NATURAL LANGUAGE PROCESSING (NLP)

AI has been very successful in speech recognition.

The use of language is a major hallmark of human intelligence. Reading and writing are specific to humans. Humans all over the world talk. The main component of the Turing Test is language.

The recent advances in NLP can be attributed to  word embeddings (which are based on distributional semantics), recurrent neural networks and pre-trained language model (GPT-2, BERT).

Machine translation is an active research area. Though machines perform well, the evaluation method (BLUE score) is flawed. They make mistakes. They are vulnerable to adversarial examples or in the presence of noise (e.g. grammatical errors, SMS text). They do not work well on low-resource languages (e.g. Indian languages).

Question answering is another active research area. IBM Watson beat the world champion in the quiz show Jeopardy! Though digital personal assistants such as Siri, Cortana, Alexa perform question-answering, typically they perform mere answer retrieval. 

PART V: MEANING AND UNDERSTANDING 

The primary traits of human intelligence are the following: 

• intuitive physics (e.g. world is composed of objects, objects have weight), intuitive biology (e.g. that there are animate and inanimate objects), intuitive psychology (social intelligence), 
• knowledge of abstract and concrete concepts (metaphor), abstraction (ability to generalize between concepts) and analogy making (identify common essence between two concepts), 
• cause and effect (e.g. if we push an object, the objects moves unless it is heavy or pulled by somebody else), predict possible futures,  simulate counterfactual events (e.g. tsunami in Heidelberg), ability to reason (use knowledge), 
• self-awareness (consciousness)

Bongard problems are designed to test the abstraction and analogy-making skills of AI. Copycat was an approach that tested these skills in a microworld --- the letter-string domain.

Cyc is an ongoing project to teach machines common-sense.

Humans have the power of metacognition, in addition of cognition. We can think about, and reason on, what we think and what others think.

Q1: How soon are self-driving cars going to a reality? Ans: We're still far, far away from developing fully  autonomous vehicles.

Q2: Will AI replace all jobs? Ans: AI will likely create a new set of jobs. AI will replace low--pay, boring, degrading, dangerous jobs.

Q3: Should we be scared of AI? Ans: Humans need not be scared of super-intelligence; at least for the next couple of decades. However we should be worried about the reckless rush to deploy AI in all spheres of life without understanding intelligence, and capabilities and vulnerabilities of AI.

Q4: Can machines be creative? Ans: Most AI systems need a human in the loop. Seemingly creative systems such as AI relied on Cope's knowledge about music, algorithm design, and selection of music. More importantly, EMI was not capable enough to judge or appreciate its own music. EMI did not "understand" music, time signatures, rhythm, tune and many other concepts related to music and composition.

What makes us proud is not only the ability to produce music, but also the ability to appreciate it, the ability to comprehend what it communicates.