Machine Learning for Monsters
Workshop
A critical and constructive exploration of machine learning. 10-day intensive workshop taught at DeepMay's Summer 2023 camp.
Course Overview
In this course, we'll use natural language processing (NLP) as a context to explore the state of machine learning and AI from many perspectives: critical and constructive, philosophical and practical.
Critique, Combat, Construct: Three Perspectives on Machine Learning
AI genealogies: How are logics of capture and (mis)representation encoded into machine learning models, and how are these models currently employed within apparatuses of control?
AI adversaries: How can the function of ML models be exploited to get them to behave in unintended ways? How can we resist the push to be desired users?
AI poetics: Can we imagine use ML in ways that step outside the paradigm of representation, prediction, and control? What does partisan machine learning look like?
Theory, Practice, Context: Three Levels of Machine Learning Inquiry
To construct these perspectives, the course exposes the historical situatedness of formal abstraction.
mathematical: build familiarity and comfort with formalisms and concepts from statistics, linear algebra and probability theory (information theory and probability, loss functions, backpropagation)
operational: develop ability to use tools in the ecosystem that implement these formalisms (text processing, language modeling, classification)
socio-historical: understand how foundational formalisms came to be important in the field and how commercial interests and ideologies of ML developers affect the way problems are defined and approached (history, political theory)
These levels correspond to degrees of contextualization. In order to engage ML through any of these lenses or at any of these levels, and do it well, it is absolutely necessary to engage with them all. To understand the math, it helps to understand the social contexts that led to its development, and to use tools that allow us to abstract away from the implementation details once we understand them. To understand and critique the current sociocultural landscapes of AI, we have to know how to interpret the methods used by applications, at the formal and toolchain level. To intervene in this landscape through misuse of models, deception of models, or application of ML to our own curiosities and problems, we need to be comfortable using the tools. To confidently apply tools to a variety of situations and new kinds of data, we need to develop firm intuitions about the underlying math. Conversely, thinking through sociological and epistemological questions raised by different modeling techniques will enable us to recognize what possibilities the ecosystem of tools open up and what possibilities they foreclose on.