Machine Learning

A central task of developmental psychology and philosophy of science is to show how humans learn radically new concepts. Famously, Fodor has argued that such learning is impossible if concepts have definitional structure and all learning is hypothesis testing. We present several learning processes that can generate novel concepts. They yield transformations of the fundamental feature space, generating new similarity structures which can underlie conceptual change. This framework provides a tractable, empiricist-friendly account that unifies and shores up various strands of (...) the neo-Quinean approach to conceptual development. (shrink)

Artificial intelligence is now capable of generating sophisticated and compelling images from simple text prompts. In this paper, I focus specifically on how artists might make use of AI to create art. Most existing discourse analogizes AI to a tool or collaborator; this focuses our attention on AI’s contribution to the production of an artistically significant output. I propose an alternative approach, the exploration paradigm, which suggests that artists instead relate to AI as a participant: artists create a space for (...) interaction with the AI algorithm by way of their prompts, thereby allowing them to explore the way that the algorithm “sees” and “represents.” AI art practiced in this fashion bears a striking resemblance to contemporary conceptual and participatory art. Viewing AI art in this way has implications for the appreciation of style in AI art; concerns about novelty and originality; and the assignment of artistic credit and copyright. (shrink)

Ambiguity occurs insofar as a reasoner lacks information about the relevant physical probabilities. There are objections to the application of standard Bayesian inductive logic and decision theory in contexts of significant ambiguity. A variety of alternative frameworks for reasoning under ambiguity have been proposed. Two of the most prominent are Imprecise Bayesianism and Dempster–Shafer theory. We compare these inductive logics with respect to the Ambiguity Dilemma, which is a problem that has been raised for Imprecise Bayesianism. We develop an agent-based (...) model comparison that isolates the difference between the two inductive logics in their updating methods. We find that Dempster–Shafer theory does not avoid the Ambiguity Dilemma. We discuss the implications of this result. (shrink)

Descartes famously constructed a language test to determine the existence of other minds. The test made critical observations about how humans use language that purportedly distinguishes them from animals and machines. These observations were carried into the generative (and later biolinguistic) enterprise under what Chomsky in his Cartesian Linguistics, terms the “creative aspect of language use” (CALU). CALU refers to the stimulus-free, unbounded, yet appropriate use of language—a tripartite depiction whose function in biolinguistics is to highlight a species-specific form of (...) intellectual freedom. This paper argues that CALU provides a set of facts that have significant downstream effects on explanatory theory-construction. These include the internalist orientation of linguistics, the invocation of a competence-performance distinction, and the postulation of a generative language faculty that makes possible—but does not explain—CALU. It contrasts the biolinguistic approach to CALU with the recent wave of enthusiasm for the use of Transformer-based Large Language Models (LLMs) as tools, models, or theories of human language, arguing that such uses neglect these fundamental insights to their detriment. It argues that, in the absence of replication, identification, or accounting of CALU, LLMs do not match the explanatory depth of the biolinguistic framework, thereby limiting their theoretical usefulness. (shrink)

Despite an increasing role of machine learning in science, there is a lack of results on limits of empirical exploration aided by machine learning. In this paper, we construct one such limit by proving undecidability of learnability of state spaces of physical systems. We characterize state spaces as binary hypothesis classes of the computable Probably Approximately Correct learning framework. This leads to identifying the first limit for learnability of state spaces in the agnostic setting. Further, using the fact that finiteness (...) of the combinatorial dimension of hypothesis classes is undecidable, we derive undecidability for learnability of state spaces as well. Throughout the paper, we try to connect our formal results with modern neural networks. This allows us to bring the limits close to the current practice and make a first step in connecting scientific exploration aided by machine learning with results from learning theory. (shrink)

Recent breakthroughs in artificial intelligence have primarily resulted from training deep neural networks (DNNs) with vast numbers of adjustable parameters on enormous datasets. Due to their complex internal structure, DNNs are frequently characterized as inscrutable ``black boxes,'' making it challenging to interpret the mechanisms underlying their impressive performance. This opacity creates difficulties for explanation, safety assurance, trustworthiness, and comparisons to human cognition, leading to divergent perspectives on these systems. This chapter examines recent developments in interpretability methods for DNNs, with a (...) focus on interventionist approaches inspired by causal explanation in philosophy of science. We argue that these methods offer a promising avenue for understanding how DNNs process information compared to merely behavioral benchmarking and correlational probing. We review key interventionist methods and illustrate their application through practical case studies. These methods allow researchers to identify and manipulate specific computational components within DNNs, providing insights into their causal structure and internal representations. We situate these approaches within the broader framework of causal abstraction, which aims to align low-level neural computations with high-level interpretable models. While acknowledging current limitations, we contend that interventionist methods offer a path towards more rigorous and theoretically grounded interpretability research, potentially informing both AI development and computational cognitive neuroscience. (shrink)

A novel advantage of the use of machine learning (ML) systems in medicine is their potential to continue learning from new data after implementation in clinical practice. To date, considerations of the ethical questions raised by the design and use of adaptive machine learning systems in medicine have, for the most part, been confined to discussion of the so-called “update problem,” which concerns how regulators should approach systems whose performance and parameters continue to change even after they have received regulatory (...) approval. In this paper, we draw attention to a prior ethical question: whether the continuous learning that will occur in such systems after their initial deployment should be classified, and regulated, as medical research? We argue that there is a strong prima facie case that the use of continuous learning in medical ML systems should be categorized, and regulated, as research and that individuals whose treatment involves such systems should be treated as research subjects. (shrink)

Background: Microbiome biomarker discovery for patient diagnosis, prognosis, and risk evaluation is attracting broad interest. Selected groups of microbial features provide signatures that characterize host disease states such as cancer or cardio-metabolic diseases. Yet, the current predictive models stemming from machine learning still behave as black boxes and seldom generalize well. Their interpretation is challenging for physicians and biologists, which makes them difficult to trust and use routinely in the physician-patient decision-making process. Novel methods that provide interpretability and biological insight (...) are needed. Here, we introduce "predomics", an original machine learning approach inspired by microbial ecosystem interactions that is tailored for metagenomics data. It discovers accurate predictive signatures and provides unprecedented interpretability. The decision provided by the predictive model is based on a simple, yet powerful score computed by adding, subtracting, or dividing cumulative abundance of microbiome measurements. -/- Results: Tested on >100 datasets, we demonstrate that predomics models are simple and highly interpretable. Even with such simplicity, they are at least as accurate as state-of-the-art methods. The family of best models, discovered during the learning process, offers the ability to distil biological information and to decipher the predictability signatures of the studied condition. In a proof-of-concept experiment, we successfully predicted body corpulence and metabolic improvement after bariatric surgery using pre-surgery microbiome data. -/- Conclusions: Predomics is a new algorithm that helps in providing reliable and trustworthy diagnostic decisions in the microbiome field. Predomics is in accord with societal and legal requirements that plead for an explainable artificial intelligence approach in the medical field. (shrink)

Consider a set of shuffled observations drawn from a fixed probability distribution over some instance domain. What enables learning of inductive generalizations which proceed from such a set of observations? The scenario is worthwhile because it epistemically characterizes most of machine learning. This kind of learning from observations is also inverse and ill-posed. What reduces the non-uniqueness of its result and, thus, its problematic epistemic justification, which stems from a one-to-many relation between the observations and many learnable generalizations? The paper (...) argues that this role belongs to any complexity regularization which satisfies Norton’s Material Theory of Induction (MTI) by localizing the inductive risk to facts in the given domain. A prime example of the localization is the Lottery Ticket Hypothesis (LTH) about overparameterized neural networks. The explanation of MTI’s role in complexity regularization of neural networks is provided by analyzing the stability of Empirical Risk Minimization (ERM), an inductive rule that controls the learning process and leads to an inductive generalization on the given set of observations. In cases where ERM might become asymptotically unstable, making the justification of the generalization by uniform convergence unavailable, LTH and MTI can be used to define a local stability. A priori, overparameterized neural networks are such cases and the combination of LTH and MTI can block ERM’s trivialization caused by equalizing the strengths of its inductive support for risk minimization. We bring closer the investigation of generalization in artificial neural networks and the study of inductive inference and show the division of labor between MTI and the optimality justifications (developed by Gerhard Schurz) in machine learning. (shrink)

The concept of Generative Artificial Intelligence (GenAI) is ubiquitous in the public and semi-technical domain, yet rarely defined precisely. We clarify main concepts that are usually discussed in connection to GenAI and argue that one ought to distinguish between the technical and the public discourse. In order to show its complex development and associated conceptual ambiguities, we offer a historical-systematic reconstruction of GenAI and explicitly discuss two exemplary cases: the generative status of the Large Language Model BERT and the differences (...) between protein structure predictions from AlphaFold 2 and 3. Our analysis shows that there is no unique and unambiguous definition of GenAI based on a purely technical account of the term. Following this conclusion, we argue that the public discourse is not simply a less complex way of speaking, but instead transcends its technical basis. As a means to structure this newly emerging discussion landscape we introduce a non-exhaustive list of four central aspects of GenAI: (multi-)modality, interaction, flexibility, and productivity. These dimensions constitute a first step towards defining GenAI beyond its technical basis. (shrink)

This paper analyses the phenomenology and epistemology of chatbots such as ChatGPT and Bard. The computational architecture underpinning these chatbots are large language models (LLMs), which are generative artificial intelligence (AI) systems trained on a massive dataset of text extracted from the Web. We conceptualise these LLMs as multifunctional computational cognitive artifacts, used for various cognitive tasks such as translating, summarizing, answering questions, information-seeking, and much more. Phenomenologically, LLMs can be experienced as a “quasi-other”; when that happens, users anthropomorphise them. (...) For most users, current LLMs are black boxes, i.e., for the most part, they lack data transparency and algorithmic transparency. They can, however, be phenomenologically and informationally transparent, in which case there is an interactional flow. Anthropomorphising and interactional flow can, in some users, create an attitude of (unwarranted) trust towards the output LLMs generate. We conclude this paper by drawing on the epistemology of trust and testimony to examine the epistemic implications of these dimensions. Whilst LLMs generally generate accurate responses, we observe two epistemic pitfalls. Ideally, users should be able to match the level of trust that they place in LLMs to the degree that LLMs are trustworthy. However, both their data and algorithmic opacity and their phenomenological and informational transparency can make it difficult for users to calibrate their trust correctly. The effects of these limitations are twofold: users may adopt unwarranted attitudes of trust towards the outputs of LLMs (which is particularly problematic when LLMs hallucinate), and the trustworthiness of LLMs may be undermined. (shrink)

In artificial intelligence (AI), responses generated by machine-learning models (most often large language models) may be unfactual information presented as a fact. For example, a chatbot might state that the Mona Lisa was painted in 1815. Such phenomenon is called AI hallucinations, seeking inspiration from human psychology, with a great difference of AI ones being connected to unjustified beliefs (that is, AI “beliefs”) rather than perceptual failures). -/- AI hallucinations may have their source in the data itself, that is, the (...) source content, or in the training procedure, i.e. the way the knowledge was encoded in the model’s parameters, so that errors in encoding and decoding textual and non-textual representations can cause hallucinations. In this paper, we will observe how such errors come to life and how they might be mitigated. For this purpose, we will analyze the usability of justification logics, to behave as a proof checker for validating the correctness of large language models’ (LLM) responses. Justification logic was developed by S. Artemov, and later on mostly by Artemov and M. Fitting, deriving its main idea from the logic of proofs (LP): knowledge and belief modalities are seen as justification terms, i.e. t:X stands for t is a (proper) justification for X. Justification logic originated from attempts to create semantics for intuitionistic logic where proofs were the most proper justifications, but in further development, justification logic could be applied to different kinds of justifications). With the recent attempts to mitigate incorrect LLM responses, we will analyze various guardrails that are currently used for LLM responses, and see how the logic of justification may provide its benefits as an AI safety layer against false data. (shrink)

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

The problem considered in this chapter emerges from the tension we find when looking at the design and architecture of chatbots on the one hand and their conversational aptitude on the other. In the way that LLM chatbots are designed and built, we have good reason to suppose they don't possess second-order capacities such as intention, belief or knowledge. Yet theories of conversation make great use of second-order capacities of speakers and their audiences to explain how aspects of interaction succeed. (...) As we can all bear witness to now though, at the point of use chatbots appear capable of performing language tasks at a level close to that of humans. This creates a tension when we consider something like, for example, the classic Gricean theory of implicature. On a broad summary of this type of account, to utter p and implicate q requires the reflexive occurrence of an audience supposing a speaker believes that q, and the speaker believing that their audience can determine they believe it when they utter p. So taken at face value, if a chatbot doesn’t have the capacity for belief, then either in their role as speaker or audience, they would not seem capable of either generating or comprehending implicatures. Yet on the surface it does seem that chatbots are capable of dealing with (some) implicatures, and as such it raises questions about how we should then correlate this with what we think occurs in cases of implicature with chatbots. (shrink)

Issues of reliability are claiming center-stage in the epistemology of machine learning. This paper unifies different branches in the literature and points to promising research directions, whilst also providing an accessible introduction to key concepts in statistics and machine learning – as far as they are concerned with reliability.

Explainable AI (xAI) methods are important for establishing trust in using black-box models. However, recent criticism has mounted against current xAI methods that they disagree, are necessarily false, and can be manipulated, which has started to undermine the deployment of black-box models. Rudin (2019) goes so far as to say that we should stop using black-box models altogether in high-stakes cases because xAI explanations ‘must be wrong’. However, strict fidelity to the truth is historically not a desideratum in science. Idealizations (...) – the intentional distortions introduced to scientific theories and models – are commonplace in the natural sciences and are seen as a successful scientific tool. Thus, it is not falsehood qua falsehood that is the issue. In this paper, I outline the need for xAI research to engage in idealization evaluation. Drawing on the use of idealizations in the natural sciences and philosophy of science, I introduce a novel framework for evaluating whether xAI methods engage in successful idealizations or deceptive explanations (SIDEs). SIDEs evaluates whether the limitations of xAI methods, and the distortions that they introduce, can be part of a successful idealization or are indeed deceptive distortions as critics suggest. I discuss the role that existing research can play in idealization evaluation and where innovation is necessary. Through a qualitative analysis we find that leading feature importance methods and counterfactual explanations are subject to idealization failure and suggest remedies for ameliorating idealization failure. (shrink)

Modern life sciences research is increasingly relying on artificial intelligence (AI) approaches to model biological systems, primarily centered around the use of machine learning (ML) models. Although ML is undeniably useful for identifying patterns in large, complex data sets, its widespread application in biological sciences represents a significant deviation from traditional methods of scientific inquiry. As such, the interplay between these models and scientific understanding in biology is a topic with important implications for the future of scientific research, yet it (...) is a subject that has received little attention. Here, we draw from an epistemological toolkit to contextualize recent applications of ML in biological sciences under modern philosophical theories of understanding, identifying general principles that can guide the design and application of ML systems to model biological phenomena and advance scientific knowledge. We propose that conceptions of scientific understanding as information compression, qualitative intelligibility, and dependency relation modelling provide a useful framework for interpreting ML-mediated understanding of biological systems. Through a detailed analysis of two key application areas of ML in modern biological research – protein structure prediction and single cell RNA-sequencing – we explore how these features have thus far enabled ML systems to advance scientific understanding of their target phenomena, how they may guide the development of future ML models, and the key obstacles that remain in preventing ML from achieving its potential as a tool for biological discovery. Consideration of the epistemological features of ML applications in biology will improve the prospects of these methods to solve important problems and advance scientific understanding of living systems. (shrink)

In Philosophy and Technology 36, David Watson discusses the epistemological and metaphysical implications of unsupervised machine learning (ML) algorithms. Watson is sympathetic to the epistemological comparison of unsupervised clustering, abstraction and generative algorithms to human cognition and sceptical about ML’s mechanisms having ontological implications. His epistemological commitments are that we learn to identify “natural kinds through clustering algorithms”, “essential properties via abstraction algorithms”, and “unrealized possibilities via generative models” “or something very much like them.” The same issue contains a commentary (...) on Watson’s paper in which Tom Sterkenburg fiercely opposes the epistemological claim that clustering algorithms can identify natural kinds. Sterkenburg argues there’s nothing about clustering arguments themselves that enables natural kind identification. What’s at stake in the Watson-Sterkenburg debate is whether it is possible for clustering algorithms to identify natural kinds and whether they underly human identification of natural kinds. My contribution to the Watson-Sterkenburg debate is twofold. First, I argue Sterkenburg’s criticism of Watson’s claim is too severe because it denies any clustering algorithm to identify any natural kind. Secondly, I argue Watson’s reply overestimates both clustering algorithms’ and humans’ access to natural kinds and his claim has to be restricted accordingly to ‘We identify kinds via clustering algorithms, or something very much like them’. (shrink)

Bayesian classifiers perform well when each of the features is completely independent of the other which is not always valid in real world applications. The aim of this study is to implement and compare the performances of each variant of the Bayesian classifier (Multinomial, Bernoulli, and Gaussian) on anomaly detection in network intrusion, and to investigate whether there is any association between each variant’s assumption and their performance. Our investigation showed that each variant of the Bayesian algorithm blindly follows its (...) assumption regardless of feature property, and that the assumption is the single most important factor that influences their accuracy. Experimental results show that Bernoulli has accuracy of 69.9% test (71% train), Multinomial has accuracy of 31.2% test (31.2% train), while Gaussian has an accuracy of 81.69% test (82.84% train). Going deeper, we investigated and found that each Naïve Bayes variants performances and accuracy is largely due to each classifier assumption, Gaussian classifier performed best on anomaly detection due to its assumption that features follow normal distributions which are continuous, while multinomial classifier have a dismal performance as it simply assumes discreet and multinomial distribution. (shrink)

As systems based on opaque Artificial Intelligence (AI) continue to flourish in diverse real-world applications, understanding these black box models has become paramount. In response, Explainable AI (XAI) has emerged as a field of research with practical and ethical benefits across various domains. This paper not only highlights the advancements in XAI and its application in real-world scenarios but also addresses the ongoing challenges within XAI, emphasizing the need for broader perspectives and collaborative efforts. We bring together experts from diverse (...) fields to identify open problems, striving to synchronize research agendas and accelerate XAI in practical applications. By fostering collaborative discussion and interdisciplinary cooperation, we aim to propel XAI forward, contributing to its continued success. Our goal is to put forward a comprehensive proposal for advancing XAI. To achieve this goal, we present a manifesto of 27 open problems categorized into nine categories. These challenges encapsulate the complexities and nuances of XAI and offer a road map for future research. For each problem, we provide promising research directions in the hope of harnessing the collective intelligence of interested stakeholders. (shrink)

Despite its centrality in the philosophy of cognitive science, there has been little prior philosophical work engaging with the notion of representation in contemporary NLP practice. This paper attempts to fill that lacuna: drawing on ideas from cognitive science, I introduce a framework for evaluating the representational claims made about components of neural NLP models, proposing three criteria with which to evaluate whether a component of a model represents a property and operationalising these criteria using probing classifiers, a popular analysis (...) technique in NLP (and deep learning more broadly). The project of operationalising a philosophically-informed notion of representation should be of interest to both philosophers of science and NLP practitioners. It affords philosophers a novel testing-ground for claims about the nature of representation, and helps NLPers organise the large literature on probing experiments, suggesting novel avenues for empirical research. (shrink)

This work demonstrates the implementation and use of an encoder-decoder model to perform a many-to-many mapping of video data to text captions. The many-to-many mapping occurs via an input temporal sequence of video frames to an output sequence of words to form a caption sentence. Data preprocessing, model construction, and model training are discussed. Caption correctness is evaluated using 2-gram BLEU scores across the different splits of the dataset. Specific examples of output captions were shown to demonstrate model generality over (...) the video temporal dimension. Predicted captions were shown to generalize over video action, even in instances where the video scene changed dramatically. Model architecture changes are discussed to improve sentence grammar and correctness. (shrink)

The irrational exuberance associated with contemporary artificial intelligence (AI) reminds me of Charles Dickens: "it was the age of foolishness, it was the epoch of belief" (cf. Nature Editorial, 2016; to get a feel for the vanity fair that is AI, see Mitchell and Krakauer, 2023; Stilgoe, 2023). It is particularly distressing—feels like yet another rerun of Seinfeld, which is all about nothing (pun intended); we have seen it in the 60s and again in the 90s. AI might have had (...) an inauspicious beginning, which may very well be the original sin, and our current predicament is the attendant divine condemnation. Being unable to stand it, however divine it may be, without further ado, I get to the genesis (instead of getting lost in the here-and-now of this or that so-called breakthrough of AI). Here's my unvarnished critique! (shrink)

The present paper looks at one of the most thorough articles on the intelligence of GPT, research conducted by engineers at Microsoft. Although there is a great deal of value in their work, I will argue that, for familiar philosophical reasons, their methodology, ‘Black-box Interpretability’ is wrongheaded. But there is a better way. There is an exciting and emerging discipline of ‘Inner Interpretability’ (also sometimes called ‘White-box Interpretability’) that aims to uncover the internal activations and weights of models in order (...) to understand what they represent and the algorithms they implement. In my view, a crucial mistake in Black-box Interpretability is the failure to appreciate that how processes are carried out matters when it comes to intelligence and understanding. I can’t pretend to have a full story that provides both necessary and sufficient conditions for being intelligent, but I do think that Inner Interpretability dovetails nicely with plausible philosophical views of what intelligence requires. So the conclusion is modest, but the important point in my view is seeing how to get the research on the right track. Towards the end of the paper, I will show how some of the philosophical concepts can be used to further refine how Inner Interpretability is approached, so the paper helps draw out a profitable, future two-way exchange between Philosophers and Computer Scientists. (shrink)

We propose the concepts of philomatics and psychomatics as hybrid combinations of philosophy and psychology with mathematics. We explain four motivations for this combination which are fulfilling the desire of analytical philosophy, proposing science of philosophy, justifying mathematical algorithms by philosophy, and abstraction in both philosophy and mathematics. We enumerate various examples for philomatics and psychomatics, some of which are explained in more depth. The first example is the analysis of relation between the context principle, semantic holism, and the usage (...) theory of meaning with the attention mechanism in mathematics. The other example is on the relations of Plato's theory of forms in philosophy with the holographic principle in string theory, object-oriented programming, and machine learning. Finally, the relation between Wittgenstein's family resemblance and clustering in mathematics is explained. This paper opens the door of research for combining philosophy and psychology with mathematics. (shrink)

The use of machine learning, or “artificial intelligence” (AI) in medicine is widespread and growing. In this paper, I focus on a specific proposed clinical application of AI: using models to predict incapacitated patients’ treatment preferences. Drawing on results from machine learning, I argue this proposal faces a special moral problem. Machine learning researchers owe us assurance on this front before experimental research can proceed. In my conclusion I connect this concern to broader issues in AI safety.

We propose a quantum curiosity algorithm as a means to implement quantum thinking into AI, and we illustrate 5 new quantum curiosity types. We then introduce 6 new hybrid quantum curiosity types combining animal and plant curiosity elements with biomimicry beyond human sensing. We then introduce 4 specialized quantum curiosity types, which incorporate quantum thinking into coding frameworks to radically transform problem-solving and discovery in science, medicine, and systems analysis. We conclude with a forecasting of the future of quantum thinking (...) in AI and illustrate an example of how to apply the new curiosity types: General Collaborative Networks. (shrink)

Current methods of operationalizing concepts of misinformation in machine learning are often problematic given idiosyncrasies in their success conditions compared to other models employed in the natural and social sciences. The intrinsic value-ladenness of misinformation and the dynamic relationship between citizens' and social scientists' concepts of misinformation jointly suggest that both the construct legitimacy and the construct validity of these models needs to be assessed via more democratic criteria than has previously been recognized.

Recent advances in natural language processing have given rise to a new kind of AI architecture: the language agent. By repeatedly calling an LLM to perform a variety of cognitive tasks, language agents are able to function autonomously to pursue goals specified in natural language and stored in a human-readable format. Because of their architecture, language agents exhibit behavior that is predictable according to the laws of folk psychology: they function as though they have desires and beliefs, and then make (...) and update plans to pursue their desires given their beliefs. We argue that the rise of language agents significantly reduces the probability of an existential catastrophe due to loss of control over an AGI. This is because the probability of such an existential catastrophe is proportional to the difficulty of aligning AGI systems, and language agents significantly reduce that difficulty. In particular, language agents help to resolve three important issues related to aligning AIs: reward misspecification, goal misgeneralization, and uninterpretability. (shrink)

Some physicians, in their care of patients at risk of misusing opioids, use machine learning (ML)-based prediction drug monitoring programmes (PDMPs) to guide their decision making in the prescription of opioids. This can cause a conflict: a PDMP Score can indicate a patient is at a high risk of opioid abuse while a patient expressly reports oppositely. The prescriber is then left to balance the credibility and trust of the patient with the PDMP Score. Pozzi1 argues that a prescriber who (...) downgrades the credibility of a patient’s testimony based on a low PDMP Score is epistemically and morally unjustified and contributes to a form of testimonial injustice. This results in patients being silenced, excluded from decision-making processes and subjected to structural injustices. Additionally, the use of ML systems in medical practices raises concerns about perpetuating existing inequalities, overestimating their capabilities and displacing human authority. However, almost the very same critiques apply to human-based systems. Formalisation, ML systems included, should instead be viewed positively,2 and precisely as a powerful means to begin eroding these and other problems in ethically sensitive domains. In this case, the epistemic virtues of formalisation include promoting transparency, consistency and replicability in decision making. Rigorous ML systems can also help ensure that models …. (shrink)

Abstract This chapter is interested in the epistemology of algorithms. As I intend to approach the topic, this is an issue about epistemic justification. Current approaches to justification emphasize the transparency of algorithms, which entails elucidating their internal mechanisms –such as functions and variables– and demonstrating how (or that) these produce outputs. Thus, the mode of justification through transparency is contingent on what can be shown about the algorithm and, in this sense, is internal to the algorithm. In contrast, I (...) advocate for an externalist epistemology of algorithms that I term computational reliabilism (CR). While I have previously introduced and examined CR in the field of computer simulations ([42, 53, 4]), this chapter extends this reliabilist epistemology to encompass a broader spectrum of algorithms utilized in various scientific disciplines, with a particular emphasis on machine learning applications. At its core, CR posits that an algorithm’s output is justified if it is produced by a reliable algorithm. A reliable algorithm is one that has been specified, coded, used, and maintained utilizing reliability indicators. These reliability indicators stem from formal methods, algorithmic metrics, expert competencies, cultures of research, and other scientific endeavors. The primary aim of this chapter is to delineate the foundations of CR, explicate its operational mechanisms, and outline its potential as an externalist epistemology of algorithms. (shrink)

Definition of the problem The umbrella term “explicability” refers to the reduction of opacity of artificial intelligence (AI) systems. These efforts are challenging for medical AI applications because higher accuracy often comes at the cost of increased opacity. This entails ethical tensions because physicians and patients desire to trace how results are produced without compromising the performance of AI systems. The centrality of explicability within the informed consent process for medical AI systems compels an ethical reflection on the trade-offs. Which (...) levels of explicability are needed to obtain informed consent when utilizing medical AI? Arguments We proceed in five steps: First, we map the terms commonly associated with explicability as described in the ethics and computer science literature, i.e., disclosure, intelligibility, interpretability, and explainability. Second, we conduct a conceptual analysis of the ethical requirements for explicability when it comes to informed consent. Third, we distinguish hurdles for explicability in terms of epistemic and explanatory opacity. Fourth, this then allows to conclude the level of explicability physicians must reach and what patients can expect. In a final step, we show how the identified levels of explicability can technically be met from the perspective of computer science. Throughout our work, we take diagnostic AI systems in radiology as an example. Conclusion We determined four levels of explicability that need to be distinguished for ethically defensible informed consent processes and showed how developers of medical AI can technically meet these requirements. (shrink)

If Large Language Models can make real scientific contributions, then they can genuinely use language, be systematically wrong, and be held responsible for their errors. AI models which can make scientific contributions thereby meet the criteria for scientific authorship.

Inspired by recent progress in multi-agent Reinforcement Learning (RL), in this work we examine the collective intelligent behaviour of theoretical universal agents by introducing a weighted mixture operation. Given a weighted set of agents, their weighted mixture is a new agent whose expected total reward in any environment is the corresponding weighted average of the original agents' expected total rewards in that environment. Thus, if RL agent intelligence is quantified in terms of performance across environments, the weighted mixture's intelligence is (...) the weighted average of the original agents' intelligences. This operation enables various interesting new theorems that shed light on the geometry of RL agent intelligence, namely: results about symmetries, convex agent-sets, and local extrema. We also show that any RL agent intelligence measure based on average performance across environments, subject to certain weak technical conditions, is identical (up to a constant factor) to performance within a single environment dependent on said intelligence measure. (shrink)

This paper presents a new trilemma with respect to resolving the control and alignment problems in machine ethics. Section 1 outlines three possible types of artificial moral agents (AMAs): (1) 'Inhuman AMAs' programmed to learn or execute moral rules or principles without understanding them in anything like the way that we do; (2) 'Better-Human AMAs' programmed to learn, execute, and understand moral rules or principles somewhat like we do, but correcting for various sources of human moral error; and (3) 'Human-Like (...) AMAs' programmed to understand and apply moral values in broadly the same way that we do, with a human-like moral psychology. Sections 2–4 then argue that each type of AMA generates unique control and alignment problems that have not been fully appreciated. Section 2 argues that Inhuman AMAs are likely to behave in inhumane ways that pose serious existential risks. Section 3 then contends that Better-Human AMAs run a serious risk of magnifying some sources of human moral error by reducing or eliminating others. Section 4 then argues that Human-Like AMAs would not only likely reproduce human moral failures, but also plausibly be highly intelligent, conscious beings with interests and wills of their own who should therefore be entitled to similar moral rights and freedoms as us. This generates what I call the New Control Problem: ensuring that humans and Human-Like AMAs exert a morally appropriate amount of control over each other. Finally, Section 5 argues that resolving the New Control Problem would, at a minimum, plausibly require ensuring what Hume and Rawls term ‘circumstances of justice’ between humans and Human-Like AMAs. But, I argue, there are grounds for thinking this will be profoundly difficult to achieve. I thus conclude on a skeptical note. Different approaches to developing ‘safe, ethical AI’ generate subtly different control and alignment problems that we do not currently know how to adequately resolve, and which may or may not be ultimately surmountable. (shrink)

We argue that machine learning algorithms can inflict microaggressions on members of marginalized groups and that recognizing these harms as instances of microaggressions is key to effectively addressing the problem. The concept of microaggression is also illuminated by being studied in algorithmic contexts. We contribute to the microaggression literature by expanding the category of environmental microaggressions and highlighting the unique issues of moral responsibility that arise when we focus on this category. We theorize two kinds of algorithmic microaggression, stereotyping and (...) erasure microaggressions, and argue that corporations are responsible for the microaggressions their algorithms create. As a case study, we look at the problems faced by Google’s autocomplete prediction and at the insufficiency of their solutions. The case study of autocomplete demonstrates our core claim that microaggressions constitute a distinct form of algorithmic bias and that identifying them as such allows us to avoid seeming solutions that recreate the same kinds of harms. Google has a responsibility to make information freely available, without exposing users to degradation. To fulfill its duties to marginalized groups, Google must abandon the fiction of neutral prediction and instead embrace the liberatory power of suggestion. (shrink)

We can illustrate the phenomenon of model-induced escape by examining the phenomenon of spam filters. Spam filter A is, we can assume, very effective at blocking spam. Indeed it is so effective that it motivates the authors of spam to invent new types of spam that will beat the filters of spam filter A. -/- An example of this phenomenon in the realm of philosophy is illustrated in the work of Nyíri on Wittgenstein's political beliefs. Nyíri writes a paper demonstrating (...) convincingly that there are strong signals of a conservative strain of thought in the writings of Wittgenstein. This has initially a tiny effect; but then a more significant effect sets in as the authors of Wittgenstein secondary literature, consciously or unconsciously, draw attention to features of Wittgenstein which cast the conservatism thesis in a negative light. This is an example of what I call model-induced escape: the Nyíri model of Wittgenstein initiates a process which undermines the Nyíri model of Wittgenstein. -/- Wittgenstein, as obiectum philosophiae, is reduced to a constantly mutating set of interpretations of a certain body of work. -/- Perhaps this is what makes philosophy so problematic when viewed from the perspective of results, or in other words of signs of progress commonly accepted across the discipline. There are just too many ways of inducing escape from any given putative discovery; too many dimensions along which an interpretative or definitional arms race can be triggered. (shrink)

This is the proceedings of the "1st Turkish Conference on AI and ANNs," K. Oflazer, V. Akman, H. A. Guvenir, and U. Halici (editors). The conference was held at Bilkent University, Bilkent, Ankara on 25-26 June 1992. -/- Language of contributions: English and Turkish.

Should we fear a future in which the already tricky world of academic publishing is increasingly crowded out by super-intelligent artificial general intelligence (AGI) systems writing papers on phenomenology and ethics? What are the chances that AGI advances to a stage where a human philosophy instructor is similarly removed from the equation? If Jobst Landgrebe and Barry Smith are correct, we have nothing to fear.

We present a game mechanic called pseudo-visibility for games inhabited by non-player characters (NPCs) driven by reinforcement learning (RL). NPCs are incentivized to pretend they cannot see pseudo-visible players: the training environment simulates an NPC to determine how the NPC would act if the pseudo-visible player were invisible, and penalizes the NPC for acting differently. NPCs are thereby trained to selectively ignore pseudo-visible players, except when they judge that the reaction penalty is an acceptable tradeoff (e.g., a guard might accept (...) the penalty in order to protect a treasure because losing the treasure would hurt even more). We describe an RL agent transformation which allows RL agents that would not otherwise do so to perform some limited self-reflection to learn the training environments in question. (shrink)

The desirable gambles framework offers the most comprehensive foundations for the theory of lower pre- visions, which in turn affords the most general ac- count of imprecise probabilities. Nevertheless, for all its generality, the theory of lower previsions rests on the notion of linear utility. This commitment to linearity is clearest in the coherence axioms for sets of desirable gambles. This paper considers two routes to relaxing this commitment. The first preserves the additive structure of the desirable gambles framework and (...) the machinery for coherent inference but detaches the interpretation of desirability from the multiplicative scale invariance axiom. The second strays from the additive combination axiom to accommodate repeated gambles that return rewards by a non-stationary processes that is not necessarily additive. Unlike the first approach, which is a conservative amendment to the desirable gambles framework, the second is a rad- ical departure. Yet, common to both is a method for describing rewards called discounted utility. (shrink)

Ever since the era of internet had ushered in cloud computing, there had been increase in the demand for the unlimited data available through cloud computing for data analysis, pattern recognition and technology advancement. With this also bring the problem of scalability, efficiency and security threat. This research paper focuses on how data can be dynamically mine in real time for pattern detection in a secure cloud computing environment using combination of decision tree algorithm and Random Forest over a restful (...) Application Programming Interface (API). We are able to successfully Implement data mining on cloud computing bypassing or avoiding direct interaction with data warehouse and without any terminal involve by using combination of IBM Cloud storage facility, Amazing Web Service, Application Programming Interface and Window service along with a decision tree and Random Forest algorithm for our classifier. We were able to successfully bypass direct connection with the data warehouse and cloud terminal with 94% accuracy in our model. (shrink)

Unless and until our society recognizes cyber bullying for what it is, the suffering of thousands of silent victims will continue.” ~ Anna Maria Chavez. There had been series of research on cyber bullying which are unable to provide reliable solution to cyber bullying. In this research work, we were able to provide a permanent solution to this by developing a model capable of detecting and intercepting bullying incoming and outgoing messages with 92% accuracy. We also developed a chatbot automation (...) messaging system to test our model leading to the development of Artificial Intelligence powered anti-cyber bullying system using machine learning algorithm of Multinomial Naïve Bayes (MNB) and optimized linear Support Vector Machine(SVM). Our model is able to detect and intercept bullying outgoing and incoming bullying messages and take immediate action. (shrink)

There is a common misconception across the lntelligence Community (IC) to the effect that information trapped within multiple heterogeneous data silos can be semantically integrated by the sorts of meaning-blind statistical methods employed in much of artificial intelligence (Al) and natural language processlng (NLP). This leads to the misconception that incoming data can be analysed coherently by relying exclusively on the use of statistical algorithms and thus without any shared framework for classifying what the data are about. Unfortunately, such approaches (...) do not yield sustainable results where we are dealing with widely distributed, highly heterogeneous and often changing bodies of data. We argue here that the needed integration requires the use of what we call an lntegrating Semantic Framework (ISF), which provldes a consistent set of categories and relationships that can be reused over and over again to tag successive bodies of data in ways which foster more coherent representation and reasoning. (shrink)

Le déficit d’explicabilité des techniques d’apprentissage machine (AM) pose des problèmes opérationnels, juridiques et éthiques. Un des principaux objectifs de notre projet est de fournir des explications éthiques des sorties générées par une application fondée sur de l’AM, considérée comme une boîte noire. La première étape de ce projet, présentée dans cet article, consiste à montrer que la validation de ces boîtes noires diffère épistémologiquement de celle mise en place dans le cadre d’une modélisation mathéma- tique et causale d’un phénomène (...) physique. La différence majeure est qu’une méthode d’AM ne prétend pas représenter une causalité entre les paramètres d’entrées et ceux de sortie. Après avoir proposé une clarification et une adaptation des notions d’inter- prétabilité et d’explicabilité telles qu’on les rencontre dans la littérature déjà abondante sur le sujet, nous montrons dans cet article l’intérêt de mettre en œuvre les distinctions épistémologiques entre les différentes fonctions épistémiques d’un modèle, d’une part, et entre la fonction épistémique et l’usage d’un modèle, d’autre part. Enfin, la dernière partie de cet article présente nos travaux en cours sur l’évaluation d’une explication, qui peut être plus persuasive qu’informative, ce qui peut ainsi causer des problèmes d’ordre éthique. (shrink)

Under what conditions does machine learning (ML) model opacity inhibit the possibility of explaining and understanding phenomena? In this article, I argue that nonepistemic values give shape to the ML opacity problem even if we keep researcher interests fixed. Treating ML models as an instance of doing model-based science to explain and understand phenomena reveals that there is (i) an external opacity problem, where the presence of inductive risk imposes higher standards on externally validating models, and (ii) an internal opacity (...) problem, where greater inductive risk demands a higher level of transparency regarding the inferences the model makes. (shrink)

In a recent article, Erasmus, Brunet, and Fisher (2021) argue that Artificial Neural Networks (ANNs) are explainable. They survey four influential accounts of explanation: the Deductive-Nomological model, the Inductive-Statistical model, the Causal-Mechanical model, and the New-Mechanist model. They argue that, on each of these accounts, the features that make something an explanation is invariant with regard to the complexity of the explanans and the explanandum. Therefore, they conclude, the complexity of ANNs (and other Machine Learning models) does not make them (...) less explainable. In this reply, it is argued that Erasmus et al. left out one influential account of explanation from their discussion: the Unificationist model. It is argued that, on the Unificationist model, the features that makes something an explanation is sensitive to complexity. Therefore, on the Unificationist model, ANNs (and other Machine Learning models) are not explainable. It is emphasized that Erasmus et al.’s general strategy is correct. The literature on explainable Artificial Intelligence can benefit by drawing from philosophical accounts of explanation. However, philosophical accounts of explanation do not settle the problem of whether ANNs are explainable because they do not unanimously declare that explanation is invariant with regard to complexity. (shrink)

Simple idealized models seem to provide more understanding than opaque, complex, and hyper-realistic models. However, an increasing number of scientists are going in the opposite direction by utilizing opaque machine learning models to make predictions and draw inferences, suggesting that scientists are opting for models that have less potential for understanding. Are scientists trading understanding for some other epistemic or pragmatic good when they choose a machine learning model? Or are the assumptions behind why minimal models provide understanding misguided? In (...) this paper, using the case of deep neural networks, I argue that it is not the complexity or black box nature of a model that limits how much understanding the model provides. Instead, it is a lack of scientific and empirical evidence supporting the link that connects a model to the target phenomenon that primarily prohibits understanding. (shrink)

We introduce and discuss a knowledge-driven distillation approach to explaining black-box models by means of two kinds of interpretable models. The first is perceptron (or threshold) connectives, which enrich knowledge representation languages such as Description Logics with linear operators that serve as a bridge between statistical learning and logical reasoning. The second is Trepan Reloaded, an ap- proach that builds post-hoc explanations of black-box classifiers in the form of decision trees enhanced by domain knowledge. Our aim is, firstly, to target (...) a model-agnostic distillation approach exemplified with these two frameworks, secondly, to study how these two frameworks interact on a theoretical level, and, thirdly, to investigate use-cases in ML and AI in a comparative manner. Specifically, we envision that user-studies will help determine human understandability of explanations generated using these two frameworks. (shrink)