60.010 Minds, Machines, and Design

Our goal in this course is to explore popular types of AI systems from three perspectives. From a design perspective, we’ll learn about the algorithms on which they run. From a psychological perspective, we’ll consider how well these algorithms model human cognition. From a philosophical perspective, we’ll grapple with the metaphysical and ethical questions they raise.

Our approach to this course will be a blend of hands-on experimentation and theoretical reflection. We’ll build spam classifiers and language models, then consider how well they model the way we recognise spam and understand language. We’ll perform adversarial and privacy attacks, then reflect on what the weaknesses of AI systems mean for artificial cognition and AI ethics. Through our explorations, we’ll learn to analyse AI systems from both design and human angles.

Learning objectives

The main goal of this course is to equip students with an ability to analyse AI systems from design, psychological, and philosophical perspectives.

After the course, the student will be able to:

• Understand, at a high level, the algorithmic design of machine learning models, reinforcement learning models, neural networks, and language models.
• Explain how selected AI models relate to human learning, decision-making, understanding, and language acquisition.
• Identify relevant similarities and differences between human and machine cognition.
• Describe the simulation argument, the Turing test, Newcomb’s problem, philosophical theories of explanation, theories of normative ethics, and explain their implications for the design of AI systems.
• Explain the implications of good word attacks, adversarial attacks, and membership inference attacks for the design and interpretation of AI systems.

Measurable outcomes

Students’ attainment of learning objectives will be measured by:

• Completion of problem sets, demonstrating ability to apply philosophical techniques introduced in class.
• Participation in seminar discussions, demonstrating understanding of the algorithmic design of AI models and engagement with the relevant philosophical and psychological ideas.
• Discussion and comments on peers’ papers, demonstrating engagement with peers’ views and understanding of seminar content.
• Writing and presentation of a paper, demonstrating abilities to describe and to explain AI-relevant psychological and philosophical ideas, as well as their implications, and to make psychologically and philosophically informed recommendations regarding the design of AI systems.

Course requirements

(WEC: Writing, Expression, Communication)

Weekly schedule

Week 1 – Simulation argument

We discuss Nick Bostrom’s Simulation Argument from three angles. By considering the design of present AI systems, we estimate the probability that AI systems can run large-scale simulations. Psychological theories shed light on assumptions made about the simulation of human minds. Philosophical analysis helps us to consider the argument’s implications, if it is sound. Our goals in this course will be to broaden our understanding of AI systems from design, psychological, and philosophical perspectives.

Week 2 – Learning

We give a brief refresher to machine learning concepts and algorithms that students have been exposed to in freshmore core courses. We consider psychological theories of inductive learning in humans, and how the Naïve Bayes approach to supervised classification is intended to mirror human inductive learning, with a Naïve Bayes spam classifier as an example. Our concepts of learning as it applies to machines and humans are compared.

Week 3 – Intelligence

Can machine learning models genuinely understand the data they process? According to John Searle, a purely algorithmic process does not possess genuine understanding of the data it processes, even if it can produce output indistinguishable from that of humans. But according to Alan Turing, if a computer is behaviourally indistinguishable from a human being, we should attribute cognition to that computer. Our concepts of intelligence as applied to humans and computers are compared. We reflect on what the limits of a spam classifier imply for the possibility of its understanding the concept of spam.

Week 4 – Reinforcement

According to reinforcement theory, humans learn to make decisions about their environment when behaviours are reinforced or punished. We introduce the key concepts associated with reinforcement learning, which models this approach to learning. We simulate the multi-armed bandit scenario and build a reinforcement learning system to explore it. The concepts of reinforcement as it applies to humans and machines are compared.

Week 5 – Decision theory

Reinforcement learning assumes that optimal (or sufficiently good) decision making can be modelled algorithmically, but is that so? We introduce decision theory, a formal framework for modelling decision making. This principle delivers intuitive results in most common cases, but potentially runs into difficulties with cases involving infinity like the St. Petersburg game and Pascal’s Wager.

Week 6 – Newcomb’s problem

The principle of expected utility maximisation says that the rational decision is one that maximises the expected reward. The dominance principle says that it is never rational to make a decision that is inferior to another in any possible world-state. Both principles sound intuitively acceptable, but Newcomb’s problem puts them at tension. We discuss possible responses to Newcomb’s problem.

Week 7: Recess Week

Week 8 – Round-table discussion 1

We discuss drafts of our term papers in groups and give feedback on each other’s ideas.

Week 9 – Cognition

Two approaches to modelling cognition are classicism (on which human cognition is like symbolic manipulation) and connectionism (on which the human brain stores information as connections in a network). We introduce the basic concepts associated with neural networks, which seem initially to be a reason to favour connectionism over classicism, but can be interpreted through a classicist lens.

Week 10 – Explanation

It’s commonly suggested that we want AI to be explainable, but what does explainability mean? We survey some popular philosophical and psychological theories of explanation and see that there are perfectly good senses in which even some complex AI systems can be said to be explainable. Adversarial attacks potentially challenge our understanding of explanation in AI systems, since they represent cases in which AI decisions change as a result of seemly irrelevant input changes.

Week 11 – Language

There are at least three possible objections to the genuine possession of language in language models. First, the nativist theory of language acquisition appears to suggest that artificial systems cannot master language. Second, language models appear to lack the communicative intentions typical of language use. Third, language models appear to lack internal world-models that allow linguistic terms to play a representational role. We consider ways of defending a positive view of understanding in language models against these objections.

Week 12 – Ethics

Value-sensitive design is an approach to AI ethics that seeks to incorporate ethical values into the AI design process: we identify the ethical principles that AI systems should uphold (like privacy) and aim to develop AI in line with those principles. One weakness of principles-based approaches to AI ethics is that there could be tensions between plausible-sounding principles. Membership inference attacks materialise this tension: transparency recommends that AI systems signal uncertainty clearly to avoid misleading users, but such signals can be exploited to compromise privacy.

Week 13 – Round-table discussion 2

We present our term papers in groups and give feedback on each other’s ideas.

Week 14 – Review

Mutually exclusive subject

• 02.184DH Minds, Machine and Design

Instructor

Nathaniel Gan