With the increasing reliance on collaborative and cloud-based systems, there is a drastic increase in attack surfaces and code vulnerabilities. Automation is key for fielding and defending software systems at scale. Researchers in Symbolic AI have had considerable success in finding flaws in human-created code. Also, run-time testing methods such as fuzzing do uncover numerous bugs. However, the major deficiency of both approaches is the inability of the methods to fix the discovered errors. They also do not scale and defy automation. Static analysis methods also suffer from the false positive problem – an overwhelming number of reported flaws are not real bugs. This brings up an interesting conundrum: Symbolic approaches actually have a detrimental impact on programmer productivity, and therefore do not necessarily con-tribute to improved code quality. What is needed is a combination of automation of code generation using large language models (LLMs), with scalable defect elimination methods using symbolic AI, to create an environment for the automated generation of defect-free code.

Combining Foundation Models with Symbolic AI for Automated Detection and Mitigation of Code Vulnerabilities

We introduce a simple initial working system in which relations (such as part-whole) are directly represented via an architecture with operating and learning rules fundamentally distinct from standard artificial neural network methods. Arbitrary data are straightforwardly encoded as graphs whose edges correspond to codes from a small fixed primitive set of elemental pairwise relations, such that simple relational encoding is not an add-on, but occurs intrinsically within the most basic components of the system. A novel graph-Hamiltonian operator calculates energies among these encodings, with ground states denoting simultaneous satisfaction of all relation constraints among graph vertices. The method solely uses radically low-precision arithmetic; computational cost is correspondingly low, and scales linearly with the number of edges in the data. The resulting unconventional architecture can process standard ANN examples, but also produces representations that exhibit characteristics of symbolic computation. Specifically, the method identifies simple logical relational structures in these data (part-of; next-to), building hierarchical representations that enable abductive inferential steps generating relational position-based encodings, rather than solely statistical representations. Notably, an equivalent set of ANN operations are derived, identifying a special case of embedded vector encodings that may constitute a useful approach to current work in higher-level semantic representation. The very simple current state of the implemented system invites additional tools and improvements.  


A logical re-conception of neural networks: Hamiltonian bitwise part-whole architecture 

This paper summarizes the current vision and results of our research since 2016, the beginning of our Jazz platform project.
We explain the design of Lifelong Formal Modeling Agents by starting with describing and defining basic ideas and discussing
their importance.

We present a complete architecture built upon these ideas, summarize our experimental results where available and discuss
how they represent minimum requirements towards building reliable human understandable AI agency. Finally, we briefly touch on how
these agents could produce more natural human-computer interfaces.

It should be noted that the definitions presented in this paper of intelligence, understanding, concept, object, symbol and other
terms are intended for practical implementation and will not completely align with other academic definitions of the same terms.

Lifelong Formal Modeling Agents -- Definitions and Implementation Strategies

Bongard problems, a collection of 100 image puzzles first introduced by Mikhail Bongard, have posed a long-standing challenge for traditional machine learning architectures. Despite their apparent simplicity, these puzzles demand explainability, reasoning and pattern recognition skills. Our work revives the approach to solving these problems proposed by Bongard's original research group, which has not been re-implemented or validated since. We demonstrate that this approach is able to successfully solve 100% of the problems in the Maksimov-Bongard problems (MBP) dataset which is explicitly constrained to contain only geometric puzzles. We extend the vocabulary of operators to demonstrate scalability of the architecture, yielding solutions to a considerable portion (44%) of the more widely known Bongard problems (BP) dataset. Finally, we argue for the value of Bongard-style architectures in AI applications that demand complex human-machine communication, transparency and compatibility with human cognitive processes.

Bongard Architecture: Towards Scalable and Transparent Machine Reasoning

Recent developments in generative AI have highlighted how the field is moving to a state where models start developing a large set of skills and can solve a multitude of tasks without having actively been optimized for them. Benchmarks have been a key component driving model development in the past, but as models’ capabilities become more complex, it becomes harder to create benchmarks which can meaningfully capture the skillsets of algorithms to inform future model developments and training pipelines. While in the domain of language we have seen the development of a wide-ranging array of benchmarks, these are currently missing for generative image algorithms. Here we introduce the Generative Image Model Benchmark for Reasoning and Representation (GIMBRR) which is an open-source software package to assess generative image algorithms on 11 cognitive tasks using manual and automated evaluation pipelines. GIMBRR is built with customizability in mind, so that it can easily be updated with new tasks and assessment routines. This way it can be adapted to suit the needs of research teams with specific goals in image generation and to update task difficulty as generative image algorithms progress in general. We used GIMBRR to measure performance of three popular generative image models (DALL-E 2, Midjourney, Stable Diffusion), demonstrating that reasoning and representation tasks pose a considerable challenge to all of them. We have also demonstrated how cognitive theory can be used to perform a systematic analysis of generative and representational capabilities of these models. 

Generative Image Model Benchmark for Reasoning and Representation (GIMBRR)

Building next-generation generalist systems requires machine learning to capture data distribution and constraint reasoning to ensure structure validity. 
Nevertheless, effective approaches are lacking in bridging constraint satisfaction and machine learning. We propose COnstraint REasoning embedded structured learning (CORE), a scalable constraint reasoning and machine learning integrated approach for learning over structured domains. We propose to embed the reasoning module as a layer in the sequential neural networks for structured prediction and content generation. We evaluate CORE on several applications: vehicle dispatching service planning, if-then program synthesis, text2SQL generation, and constrained image generation. The proposed CORE module demonstrates superior performance over state-of-the-art approaches in all the applications. The structures generated with CORE satisfy 100% of the constraints, when using exact decision diagrams.


Constraint Reasoning for Structured Prediction and Image Generation

There is a concerted effort to build domain-general artificial intelligence in the form of universal neural network models with sufficient computational flexibility to solve a wide variety of cognitive tasks but without requiring fine-tuning on individual problem spaces and domains.  To do this, models need appropriate priors and inductive biases, such that trained models can generalise to out-of-distribution examples and new problem sets. Here we provide an overview of the hallmarks endowing biological neural networks with the functionality needed for flexible cognition, in order to establish which features might also be important to achieve similar functionality in artificial systems. We specifically discuss the role of system-level distribution of network communication and recurrence, in addition to the role of short-term topological changes for efficient local computation. As machine learning models become more complex, these principles may provide valuable directions in an otherwise vast space of possible architectures. In addition, testing these inductive biases within artificial systems may help us to understand the biological principles underlying domain-general cognition.