Richard Evans

In this paper, I attempt to address a fundamental challenge for machine intelligence: to understand whether and how a machine’s internal states and external outputs can exhibit original non-derivative intentionality. This question has three aspects. First, what does it take for a machine to exhibit original de dicto intentionality? Second, what does it take to exhibit original de re intentionality? Third, what is required for the machine to defer to the external objective world by respecting the word-to-world direction of fit? I attempt to answer the first challenge by providing a constitutive counts-as understanding of de dicto intentionality. This analysis involves repurposing Kant’s vision of a self-legislating agent as a specification of a machine that reprograms itself. I attempt to answer the second and third challenges by extending Kant’s synchronic model of de dicto intentionality with Brandom’s interpretation of Hegel’s diachronic model of de re intentionality, using Hegel’s notion of recollection to provide an understanding of what is involved in achieving deference to the external world.

This paper describes an attempt to repurpose Kant’s a priori psychology as the architectural blueprint for a machine learning system. First, it describes the conditions that must be satisfied for the agent to achieve unity of experience: the intuitions must be connected, via binary relations, so as to satisfy various unity conditions. Second, it shows how the categories are derived within this model: the categories are pure unary predicates that are derived from the pure binary relations. Third, I describe how Kant’s cognitive architecture has been implemented in a computer system (the Apperception Engine) and show in detail what it is like for the system to construct a unified experience from a sequence of raw sensory input.

How should a machine intelligence perform unsupervised structure discovery over streams of sensory input? One approach to this problem is to cast it as an apperception task [1]. Here, the task is to construct an explicit interpretable theory that both explains the sensory sequence and also satisfies a set of unity conditions, designed to ensure that the constituents of the theory are connected in a relational structure. However, the original formulation of the apperception task had one fundamental limitation: it assumed the raw sensory input had already been parsed using a set of discrete categories, so that all the system had to do was receive this already-digested symbolic input, and make sense of it. But what if we don't have access to pre-parsed input? What if our sensory sequence is raw unprocessed information? The central contribution of this paper is a neuro-symbolic framework for distilling interpretable theories out of streams of raw, unprocessed sensory experience. First, we extend the definition of the apperception task to include ambiguous (but still symbolic) input: sequences of sets of disjunctions. Next, we use a neural network to map raw sensory input to disjunctive input. Our binary neural network is encoded as a logic program, so the weights of the network and the rules of the theory can be solved jointly as a single SAT problem. This way, we are able to jointly learn how to perceive (mapping raw sensory information to concepts) and apperceive (combining concepts into declarative rules).

This paper attempts to answer a central question in unsupervised learning: what does it mean to “make sense” of a sensory sequence? In our formalization, making sense involves constructing a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the causal theory – objects, properties, and laws – must be integrated into a coherent whole. On our account, making sense of sensory input is a type of program synthesis, but it is unsupervised program synthesis. Our second contribution is a computer implementation, the Apperception Engine, that was designed to satisfy the above requirements. Our system is able to produce interpretable human-readable causal theories from very small amounts of data, because of the strong inductive bias provided by the unity conditions. A causal theory produced by our system is able to predict future sensor readings, as well as retrodict earlier readings, and impute (fill in the blanks of) missing sensory readings, in any combination. In fact, it is able to do all three tasks simultaneously. We tested the engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The Apperception Engine performs well in all these domains, significantly out-performing neural net baselines. We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.

This paper formalizes part of the cognitive architecture that Kant develops in the Critique of Pure Reason. The central Kantian notion that we formalize is the rule. As we interpret Kant, a rule is not a declarative conditional stating what would be true if such and such conditions hold. Rather, a Kantian rule is a general procedure, represented by a conditional imperative or permissive, indicating which acts must or may be performed, given certain acts that are already being performed. These acts are not propositions; they do not have truth-values. Our formalization is related to the input/ output logics, a family of logics designed to capture relations between elements that need not have truth-values. In this paper, we introduce KL3 as a formalization of Kant’s conception of rules as conditional imperatives and permissives. We explain how it differs from standard input/output logics, geometric logic, and first-order logic, as well as how it translates natural language sentences not well captured by first-order logic. Finally, we show how the various distinctions in Kant’s much-maligned Table of Judgements emerge as the most natural way of dividing up the various types and sub-types of rule in KL3. Our analysis sheds new light on the way in which normative notions play a fundamental role in the conception of logic at the heart of Kant’s theoretical philosophy.

In this paper, I reinterpret Kant’s Transcendental Analytic as a description of a cognitive architecture. I describe a computer implementation of this architecture, and show how it has been applied to two unsupervised learning tasks. The resulting program is very data efficient, able to learn from a tiny handful of examples. I show how the program achieves data-efficiency: the constraints described in the Analytic of Principles are reinterpreted as strong prior knowledge, constraining the set of possible solutions.

We are able to participate in countless different sorts of social practice. This indefinite set of capacities must be explainable in terms of a finite stock of capacities. This paper compares and contrasts two different explanations. A standard decomposition of the capacity to participate in social practices goes something like this: the interpreter arrives on the scene with a stock of generic practice-types. He looks at the current scene to fill-in the current tokens of these types. He looks at the current state of these practice tokens to see what actions are available to him. He uses his current desires to choose between these various possible actions. I argue that this standard explanation is defective, drawing on arguments by Searle and Wittgenstein and Garfinkel. I propose an alternative explanation, in which the participants must continually show each other the state of the scene in order to maintain the scene’s intelligibility. I provide a simple formal language in which to describe this alternative approach, in which we can state quite precisely what someone is doing when they participate in a practice. This language is related to both deontic and epistemic logics, but it is much simpler – it does not include the classic propositional connectives, and it is driven by a very different set of assumptions. The inspirations for this formal language are Searle’s analysis of directions of fit, Wittgenstein’s remarks on rule-following and Garfinkel’s ethnomethodology.

Cathoristic logic is a multi-modal logic where negation is replaced by a novel operator allowing the expression of incompatible sentences. We present the syntax and semantics of the logic including complete proof rules, and establish a number of results such as compactness, a semantic characterisa- tion of elementary equivalence, the existence of a quadratic-time decision pro- cedure, and Brandom’s incompatibility semantics property. We demonstrate the usefulness of the logic as a language for knowledge representation.

Research in multi-agent systems typically assumes a regulative model of social practice. This model starts with agents who are already capable of acting autonomously to further their individual ends. A social practice, according to this view, is a way of achieving coordination between multiple agents by restricting the set of actions available. For example, in a world containing cars but no driving regulations, agents are free to drive on either side of the road. To prevent collisions, we introduce driving regulations, insisting that everyone drives on the left hand side of the road. We accept this limitation on our freedom because it lowers the probability of a collision. This paper describes AI systems that are based on the constitutive view of social practice. According to this alternative view, there are certain actions that are only available to the agent because he is participating in a practice of a certain sort. For example, you can swing a peculiarly shaped piece of wood without participating in any particular practice - but this action will only constitute a strike if you are participating in a game of baseball. I can raise my hand whenever I like, but this only counts as voting for the motion within the institution of voting. But not all the examples are from frivolous, optional or institutional activities. Sellars and Brandom argue that even the fundamental act of assertion is only possible within the context of a particular social practice: the Game of Giving and Asking for Reasons. The constitutive view, which is familiar to philosophers, is relatively unknown within the AI community. The central aim of this paper is to show how the constitutive view of practice can be put to use in AI applications.

The central claim of normative pragmatism is that intentional states can be explained in terms of participation in practices. My aim in this paper is not so much to defend this claim as to rearticulate it in a different medium: the medium of computation. I describe two computer programs in which this claim is re-expressed. The first is the latest version of THE SIMS, in which participation in practices enables the Sims to do and understand more. The second is a prototype simulation of philosophical debate. In this second program, normative pragmatism is expressed at two different levels: once as the theory powering the implementation, and once as a set of claims which are debated by the simulated philosophers. Here, a philosophical theory is used to implement a system in which that very same theory is articulated, challenged, and justified.