Human-like systematic generalization through a meta-learning neural network

The four primitive words are direct mappings from one input word to one output symbol (for example, ‘dax’ is RED, ‘wif’ is GREEN, ‘lug’ is BLUE). Function 1 (‘fep’ in Fig. 2) takes the preceding primitive as an argument and repeats its output three times (‘dax fep’ is RED RED RED). Function 2 (‘blicket’) takes both the preceding primitive and following primitive as arguments, producing their outputs in a specific alternating sequence (‘wif blicket dax’ is GREEN RED GREEN). Last, function 3 (‘kiki’) takes both the preceding and following strings as input, processes them and concatenates their outputs in reverse order (‘dax kiki lug’ is BLUE RED). We also tested function 3 in cases in which its arguments were generated by the other functions, exploring function composition (‘wif blicket dax kiki lug’ is BLUE GREEN RED GREEN). During the study phase (see description below), participants saw examples that disambiguated the order of function application for the tested compositions (function 3 takes scope over the other functions).

Controversies arose from early on in symbolic AI, both within the field—e.g., between logicists (the pro-logic “neats”) and non-logicists (the anti-logic “scruffies”)—and between those who embraced AI but rejected symbolic approaches—primarily connectionists—and those outside the field. Critiques from outside of the field were primarily from philosophers, on intellectual grounds, but also from funding agencies, especially during the two AI winters. Marvin Minsky first proposed frames as a way of interpreting common visual situations, such as an office, and Roger Schank extended this idea to scripts for common routines, such as dining out.

The Frame Problem: knowledge representation challenges for first-order logic

Cyc has attempted to capture useful common-sense knowledge and has “micro-theories” to handle particular kinds of domain-specific reasoning. The logic clauses that describe programs are directly interpreted to run the programs specified. No explicit series of actions is required, as is the case with imperative programming languages.

• He gave a talk at an AI workshop at Stanford comparing symbols to aether, one of science’s greatest mistakes.
• Extensions to first-order logic include temporal logic, to handle time; epistemic logic, to reason about agent knowledge; modal logic, to handle possibility and necessity; and probabilistic logics to handle logic and probability together.
• A successful model must learn and use words in systematic ways from just a few examples, and prefer hypotheses that capture structured input/output relationships.
• For vision problems, an image classifier or generator could similarly receive specialized meta-training (through current prompt-based procedures57) to learn how to systematically combine object features or multiple objects with relations.

By 2015, his hostility toward all things symbols had fully crystallized. He gave a talk at an AI workshop at Stanford comparing symbols to aether, one of science’s greatest mistakes. In contrast, a multi-agent system consists of multiple agents that communicate amongst themselves with some inter-agent communication language such as Knowledge Query and Manipulation Language (KQML).

A simple guide to gradient descent in machine learning

You create a rule-based program that takes new images as inputs, compares the pixels to the original cat image, and responds by saying whether your cat is in those images. Using OOP, you can create extensive and complex symbolic AI programs that perform various tasks. While this may be unnerving to some, it must be remembered that symbolic AI still only works with numbers, just in a different way. By creating a more human-like thinking machine, organizations will be able to democratize the technology across the workforce so it can be applied to the real-world situations we face every day.

Critically, this model was trained only on the copy task of identifying which study example is the same as the query example, and then reproducing that study example’s output sequence (see specification in Extended Data Fig. 4; set 1 was used for both study and query examples). It was not trained to handle novel queries that generalize beyond the study set. Thus, the model was trained on the same study examples as MLC, using the same architecture and procedure, but it was not explicitly optimized for compositional generalization.

And their symbolic play evolves as they work with some sort of plan, assign roles, and act out sequenced steps. The early pioneers of AI believed that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” Therefore, symbolic AI took center stage and became the focus of research projects. Using symbolic AI, everything is visible, understandable and explainable, leading to what is called a “transparent box,” as opposed to the “black box” created by machine learning. This page includes some recent, notable research that attempts to combine deep learning with symbolic learning to answer those questions.

Advantages of multi-agent systems include the ability to divide work among the agents and to increase fault tolerance when agents are lost. Research problems include how agents reach consensus, distributed problem solving, multi-agent learning, multi-agent planning, and distributed constraint optimization. Natural language processing focuses on treating language as data to perform tasks such as identifying topics without necessarily understanding the intended meaning.

The technology actually dates back to the 1950s, says expert.ai’s Luca Scagliarini, but was considered old-fashioned by the 1990s when demand for procedural knowledge of sensory and motor processes was all the rage. Now that AI is tasked with higher-order systems and data management, the capability to engage in logical thinking and knowledge representation is cool again. If machine learning can appear as a revolutionary approach at first, its lack of transparency and a large amount of data that is required in order for the system to learn are its two main flaws. Companies now realize how important it is to have a transparent AI, not only for ethical reasons but also for operational ones, and the deterministic (or symbolic) approach is now becoming popular again.

Instead, for each vocabulary word that takes a permuted meaning in an episode, the meta-training procedure chooses one arbitrary study example that also uses that word, providing the network an opportunity to infer its meaning. Any remaining study examples needed to reach a total of 8 are sampled arbitrarily from the training corpus. For organizations looking forward to the day they can interact with AI just like a person, symbolic AI is how it will happen, says tech journalist Surya Maddula. After all, we humans developed reason by first learning the rules of how things interrelate, then applying those rules to other situations – pretty much the way symbolic AI is trained.

Extended Data Fig. 1 Few-shot instruction learning task with full set of queries.

As a credit to Fodor and Pylyshyn’s prescience, the systematicity debate has endured. Systematicity continues to challenge models11,12,13,14,15,16,17,18 and motivates new frameworks34,35,36,37,38,39,40,41. Preliminary experiments reported in Supplementary Information 3 suggest that systematicity is still a challenge, or at the very least an open question, even for recent large language models such as GPT-4. To resolve the debate, and to understand whether neural networks can capture human-like compositional skills, we must compare humans and machines side-by-side, as in this Article and other recent work7,42,43. In our experiments, we found that the most common human responses were algebraic and systematic in exactly the ways that Fodor and Pylyshyn1 discuss. However, people also relied on inductive biases that sometimes support the algebraic solution and sometimes deviate from it; indeed, people are not purely algebraic machines3,6,7.

The expert system processes the rules to make deductions and to determine what additional information it needs, i.e. what questions to ask, using human-readable symbols. For example, OPS5, CLIPS and their successors Jess and Drools operate in this fashion. During the COGS test (an example episode is shown in Extended Data Fig. 8), MLC is evaluated on each query in the test corpus. For each query, eight study examples are sampled from the training corpus, using the same procedure as above for picking study examples that facilitate word overlap (note that picking study examples from the generalization corpus would inadvertently leak test information). Neither the study nor query examples are remapped to probe how models infer the original meanings.

The Neuro-Symbolic Concept Learner

For scoring a particular human response y1, …, y7 by log-likelihood, MLC uses the same factorization as in equation (1). Performance was averaged over 200 passes through the dataset, each episode with different random query orderings as well as word and colour assignments. Researchers note that a child who follows playing (stirring milk and then feeding the doll) will also be able to manage syntax in language (“I need paper and crayons”). Bruner views symbolic representation as crucial for cognitive development, and since language is our primary means of symbolizing the world, he attaches great importance to language in determining cognitive development.

• The Symbolic AI paradigm led to seminal ideas in search, symbolic programming languages, agents, multi-agent systems, the semantic web, and the strengths and limitations of formal knowledge and reasoning systems.
• MLC was evaluated on this task in several ways; in each case, MLC responded to this novel task through learned memory-based strategies, as its weights were frozen and not updated further.
• Our use of MLC for behavioural modelling relates to other approaches for reverse engineering human inductive biases.
• The first is that human compositional skills, although important, may not be as systematic and rule-like as Fodor and Pylyshyn indicated3,6,7.

When you provide it with a new image, it will return the probability that it contains a cat. Henry Kautz,[17] Francesca Rossi,[80] and Bart Selman[81] have also argued for a synthesis. Their arguments are based on a need to address the two kinds of thinking discussed in Daniel Kahneman’s book, Thinking, Fast and Slow.

This kind of meta-level reasoning is used in Soar and in the BB1 blackboard architecture. Forward chaining inference engines are the most common, and are seen in CLIPS and OPS5. Backward chaining occurs in Prolog, where a more limited logical representation is used, Horn Clauses. TPSR uncovering the governing symbolic mathematics of data, providing enhanced extrapolation capabilities.

Read more about https://www.metadialog.com/ here.