In a recent TED talk titled One Thing to Teach in the Age of AI, innovation practitioner Bryan Cassady shared a story about an exam created by one of his colleagues. The exam question was deceptively simple: “Prove you’ve learned the material in this course by giving me five good questions.” Many students struggled with that, and certainly More What Makes a Good Question in the Age of AI
If AI really turns out to be normal technology, an increasingly persuasive case Arvind Narayanan and Sayash Kapoor have continued to advance, what happens to the competitive dynamics and industry structures for the organisations operating in that space? Using a synthesis of Porter s Five Forces Framework, Helmer s Seven Powers and Oberholzer s Value Stick that I More If AI really turns out to be a…
A significantly expanded version of a paper I shared earlier on Privacy Technologies for Financial Intelligence is now published in the Journal of Economic Criminology s Special Issue on Economic Crime in a Changing World: Emerging Threats and Strategic Responses. The paper is open-access and written for a broad non-technical audience. It can be found here: More JEC Special Issue on Economic Crime…
We have substantially revised the factored conditional filtering paper to frame it more generally as a solution for belief acquisition in AI agents. Here s the revised paper on arXiv. https://arxiv.org/abs/2206.02178v3 Here s the abstract: This paper studies how belief acquisition can be accomplished using stochastic filtering. First, a theoretical foundation for empirical beliefs is outlined.…
Using sequential hypothesis testing techniques to check the modelling assumptions of Bayesian mixture estimators is a promising way of getting value out of combining the Bayesian and frequentist approaches to probability. Here s a paper to show how that can be done for Context Tree Weighting and related methods. Paper Abstract: Universal Bayesian sequence predictors like More Martingale Tests for…
All my life I have been searching for simple and effective methods for constructing prediction intervals for different AI/ML models. I don t know why I never encountered Conformal Prediction until recently, but I suppose it is better late than never. Conformal prediction is (arguably) the most elegant and practical technique for improving the robustness in More Notes on Conformal Prediction and…
Threat modelling is now considered a best practice in comprehensive technical approaches to dealing with AI safety issues [S+25]. Threat modeling [S14] is a structured, proactive process used to identify potential threats and vulnerabilities in a system. While the traditional focus is on cyber-security and privacy issues, threat modelling has been extended for AI systems More AI Risk Assessment…
AI governance and AI assurance are sometimes conflated in conversations but they are not the same thing. In simple terms, In more details, the AI Governance Framework for an organisation is the strategic blueprint that defines the organisation’s overarching approach to developing, deploying, and managing AI systems. Its purpose is to define the roles, responsibilities, More AI Governance vs AI…
I read a couple of really good books on China’s economic system this year that I can recommend for those who want to know the country better (from a distance). The New China Playbook: Beyond Socialism and Capitalism by Keyu Jin taught me about mayor economies, and how China’s centrally planned economy is actually executed More Book Recommendations on China’s Economic System
Like many difficult cyber security problems, prompt-injection attacks is likely to become an ongoing issue that shifts and turns with the continual discovery of new attacks and new defences going forward. Instead of responding in natural language given a prompt, the best current defence I know involves always generating code, say, in a safe interpreted More Thoughts on Prompt Injection Attacks
To support public-service agencies in the implementation of their own responsible use of AI policies, the Australian Government s Digital Transformation Agency (DTA) has made publicly available its AI Fundamentals training course in the form of a SCORM package, a commonly used technical standard for putting together content for Learning Management Systems (LMS). The DTA training More Customising…
Differential Privacy (DP) is one of the most widely adopted formal model of privacy protection but its semantics, especially in the presence of correlated data and in the adversarial interactive setting, is still not broadly understood among data science practitioners. In this paper, we first look at how DP originated from research on database-reconstruction attacks More On the Semantics of…
Cyber security and data security are closely related concepts that operate at different levels and provide different safeguards. Cyber security is primarily about controlling access to systems and data through different security protection mechanisms, from the physical network layer all the way to the application layer. These security mechanisms come primarily in the form of More Data Security vs…
Modern competitive analysis started in the 1930 s with the introduction of the Structural-Conduct-Performance (SCP) paradigm (shown in Figure 1 below) that was used to analyse the causal dynamics of Industrial Organisation, the field of study concerned to a large extent with how public policy can limit monopoly power. The SCP paradigm has since evolved into More Competitive Strategies: Structures…
A colleague and I have just released on arXiv a paper titled Enabling Secure and Ephemeral AI Workloads in Data Mesh Environments . The key innovation is in pushing the now well-established idea of minimal immutable data structures up and down the software infrastructure stack a bit further than what others have done, resulting in a More Secure and Ephemeral AI Workloads in Data Mesh Environments
Of the many questions we wish to answer using LLMs, it can be useful to distinguish between puzzles and mysteries. As Gregory Teverton explained in his many articles, a puzzle is a problem that has a definite and verifiable answer, but a mystery is one that poses a question that has no definitive answer because More Puzzles and Mysteries in Generative AI
I recently released on arXiv a new version of the paper The Problem of Social Cost in Multi-Agent General Reinforcement Learning: Survey and Synthesis, which can be found at https://arxiv.org/abs/2412.02091 The new version has These are all non-trivial extensions of the paper that build on recent new results in different fields and they are worth More Update on Social Cost of Multi-Agent…
I wrote the following comments very quickly on LinkedIn about 10 days ago in the middle of the DeepSeek frenzy, and the post turned out to be my second most-read post ever and people continues to look at it. I think the comments are holding up well so I am sharing them here as well, More DeepSeek and All That
There s been a lot written about Fourier Transforms over the years. From a physics perspective, I would recommend the Harmonics chapter in the Feynman Lectures on Physics. I also found Elan Ness-Cohn s visual explanation of Fourier Transforms really satisfying. In the notes linked below, I have tried to give, in 5 short pages, the intuition More Algebraic Intuitions behind Fourier Transforms
Fairness in Machine Learning is a topic that I have been wanting to better understand for a little while now, and this blog post summarises what I learned from reading the Fairness and Machine Learning (FML) book by Solon Barocas, Moritz Hardt and Arvind Narayanan available at https://fairmlbook.org these past couple of days. (The book More A Simplistic Guide to Using Fairness Criteria in Machine…
While I have worked on aspects of AI safety for quite a few years now, in particular privacy technologies and confidential computing, I am a late convert on the importance of Artificial General Intelligence (AGI) safety research and did not take the problem seriously until about 1 year ago. My mindset has now changed completely More The Problem of Social Cost in Multi-Agent Universal Reinforcement…
With NVIDIA seemingly steaming ahead in their latest quarterly result, Apple Intelligence receiving a lukewarm response from users, Wall Street increasingly worried about the return-on-investment from the hyperscalers massive capital investments, stories that CIOs are struggling to find ROI for AI, and news in the last two days that Intel and Samsung are both struggling More Winners and Losers in…
As is well-known by now, the universal AI agent AIXI is made up of two key components: Solomonoff Induction for universal sequential prediction, and expectimax search for planning. There are several proposed and reasonably effective approximations of the Solomonoff Induction component using the factored, binarised Context Tree Weighting algorithm [WST95, VNHUS09] and its generalisation to More…
Have you ever wondered why the natural exponential function shows up so frequently in mathematical inequalities? Here s a graph of the natural exponential function. The constant e has a special place in mathematics, which is beautifully chronicled in Eli Maor s book [M94]. The definition of e that is most useful and intuitive for our purpose More Natural Exponential Functions in Inequalities
A key claim of differential privacy in [DR14] is that it provides automatic neutralization of linkage attacks, including all those attempted with all past, present, and future datasets and other forms and sources of auxiliary information . This is an important and often repeated claim see e.g. [N17, Section E] and [PR23] but the More Dealing with Linkage Attacks using Differential Privacy
This blog post Sizing up the New Axis by Noah Smith that compares the combined economic strength of China and Russia against that of US and its key allies provides interesting insights into the robustness of the industrial base and manufacturing supply chains of the two blocs. On energy supply, which includes fossil fuels and More A Short Look at Geoeconomics
It took a little while to write, but hopefully the following survey paper by Yang Li, Thilina Ranbaduge and yours truly can help demystify financial intelligence and privacy technologies for practitioners and technologists alike. The focus is on anti-money laundering and counter-terrorism financing, but the opportunity set is much broader. https://arxiv.org/abs/2408.09935 Here s the abstract of…
Differential Privacy (DP) is, by now, the most widely adopted formal model of privacy protection used in industry [L23] and government [ABS22] but my sense is that its semantics , especially in the presence of correlated data and in the adversarial interactive setting, is still not broadly understood in the community, especially among practitioners. In the More What Can Differential Privacy…
We have just released another paper on arXiv: https://arxiv.org/abs/2406.17649 Here s the abstract: We consider the problem of privacy protection in Reinforcement Learning (RL) algorithms that operate over population processes, a practical but understudied setting that includes, for example, the control of epidemics in large populations of dynamically interacting individuals. In this setting, the…
The Bretagnolle-Huber Inequality provides a bound on the total variation distance between two probability distributions in terms of their Kullback-Leibler divergence, and it is better than Pinsker s Inequality when the KL divergence is larger than two, and it is never vacuous, as shown in this figure from [Canonne2023]. The following is an elementary proof, which More Notes on the…
I have been incredibly frustrated with the lack of quality and content in many responsible AI (RAI) conversations. Almost all the (non-academic) RAI meetings I attended these past 12 months involve the speakers repeating words like fairness, accountability, and transparency basically for the entire duration of the meeting, with everyone nodding furiously in agreement about More Improving the…
I have been thinking about data security issues, in particular database-reconstruction attacks. To quote Wikipedia, a reconstruction attack is any method for partially reconstructing a private database from public aggregate information. The question I am specifically interested in is this: Can an attacker with general interactive query access to a dataset recover a piece of More How To Deal with…
Regular users of arXiv.org may have noticed that on every paper s page, under the Related Papers tab, one can now find the paper s Influence Flower, which is a nice way to visualise citation influences among academic entities, including papers, authors, institutions, and research topics. The following, for example, are the author-centric and venue-centric influence flowers More Influence Flower
My phd student just got a new paper accepted at the upcoming AAAI Conference on Artificial Intelligence. Here s the abstract of the paper: Prior approximations of AIXI, a Bayesian optimality notion for general reinforcement learning, can only approximate AIXI’s Bayesian environment model using an a-priori defined set of models. This is a fundamental source of More Dynamic Knowledge Injection for…
My phd student has just released a paper on 3D Object-Centric Learning on arXiv. I am pretty proud of the work, although I really only understand around 40% of it. Here s the abstract: We tackle the task of scalable unsupervised object-centric representation learning on 3D scenes. Existing approaches to object-centric representation learning show limitations in More Variational Inference for…
There are many different definitions of Artificial Intelligence in the literature, all are suggestive and insightful. However, at the end of the day, I think there is really one simple enough to be understood and formalised rigorously. This is John McCarthy s original definition of AI from 1955: the science and engineering of making intelligent machines . More A Simple Definition of Artificial…
About 5 years ago, Tania Churchill and I assembled a team of researchers and engineers across AUSTRAC and ANU to work on privacy technologies for detecting criminal activities across the financial system, funded by the Fintel Alliance Expansion budget measure, the Investigative Analytics NPP (led by CSIRO s Data61), and an ANU Translational Fellowship. The overall More FinTracer and Friends
My colleagues Mike Purcell, Kelvin Yang Li and I have a new paper on differentially private set intersection cardinality algorithm accepted at this year’s Uncertainty in Artificial Intelligence conference. Here is the abstract:We describe a simple two-party protocol in which each party contributes a set as input. The output of the protocol is an estimate More Split Count and Share: A…
Here is my attempt to map out the major classes of algorithms in Machine Learning, organised around the associated induction principles and learning theory. The usual caveats apply around this being biased towards my own experience. At the highest level, we can distinguish between the Passive and Active learning settings. In the passive case, the More A Map of Machine Learning Principles and…
I realised recently that I don’t have a good working definition of what Data Products are or should be. Sure, a quick googling will surface many generic definitions of Data Products from respected sources like Forbes and McKinsey. They are all usually variations of DJ Patil’s definition: “A data product is a product that facilitates More What Are Data Products?
I was recently asked to speak about my leadership brand, a topic I hadn’t seriously thought about ever, until now. So what is my brand. Here are a few things that I know: So I think my leadership brand is the rational leader who tries to understand how things work, and who seeks to inspire More My Leadership Brand
In this post on the last day of the year, I thought I will share a map of mathematical structures that are useful for thinking about knowledge representation and reasoning (KRR) issues in Artificial Intelligence and Machine Learning. It is built on top of the diagram shown in Max Tegmark s paper Is the theory of More A Map of Mathematical Structures for AI
Artificial Intelligence as a well-defined mathematical problem was solved a number of years ago through the formulation of the AIXI agent by Prof Marcus Hutter — see https://theconversation.com/to-create-a-super-intelligent-machine-start-with-an-equation-20756 for a quick introduction — but a key fundamental issue with the AIXI theory has always been the incomputability of the general solution. In…
A few of us have been working with homomorphic encryption for a number of years now, but we never found a paper / book that covers all the foundational mathematical material in one place. So we decided to write one — well my postdoc Kelvin Yang Li decided to write one and Mike Purcell and I assisted More A Tutorial Introduction to Lattice-based Cryptography and Homomorphic Encryption
A few colleagues and I have just completed a new research paper titled Factored Conditional Filtering: Tracking States and Estimating Parameters in High-Dimensional Spaces. The research took over 3 years and I am really excited about the underlying theory and its possible applications. In particular, the paper shows how we can lift Bayesian filtering to More Bayesian Filtering on Structured…
Cyclotomic polynomials are widely used in the construction of lattice cryptography and homomorphic encryption schemes based on the Ring Learning With Error problem. In this short note, Kelvin Li and I attempt a self-contained introduction to the cyclotomic polynomials and the Galois groups of cyclotomic extensions.
Data matching and entity resolution is a common first step in data preparation and there is a thousand academic papers written on the subject in the literature. In practice, for large datasets – anything more than a million records will do as a definition of large here because most data-matching algorithms can’t handle that because More A Note on Large Scale Data Matching and Entity Resolution
I read the book Divining a Digital Future: Mess and Mythology in Ubiquitous Computing by Genevieve Bell and Paul Dourish almost cover to cover last year and I have to say that finally gave me some understanding of the research agenda being pursued at the 3A Institute at the ANU. I hope I am not More Book Review: Divining a Digital Future
A few colleagues and I have just put on arXiv a new survey paper on Private Graph Data Release, which took us nearly 9 months to write. Here s the abstract: The application of graph analytics to various domains have yielded tremendous societal and economical benefits in recent years. However, the increasingly widespread adoption of graph More Private Graph Data Release using Differential Privacy
One of my PhD students has just released a paper titled Spatially Invariant Unsupervised 3D Object Segmentation Using Graph Neural Networks. Here s the abstract: In this paper, we tackle the problem of unsupervised 3D object segmentation from a point cloud without RGB information. In particular, we propose a framework, SPAIR3D, to model a point cloud More Unsupervised 3D Object Segmentation