How to Learn and Represent Abstractions: An Investigation using Symbolic Alchemy

Badr, AlKhamissi,; Srinivasan, Akshay; Nelson, Zeb-Kurth; Ritter, Sam

doi:10.48550/arxiv.2112.08360

Cited by 1 publication

(3 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To test whether the general pattern of results described above hold for other kinds of neural network architectures, we repeated our experiments for three more architectures (Figs 4 and S7): Episodic Planning Networks [43,44], transformers [45], and Compositional Relational Networks [46]. Below is a description of each architecture.…”

Section: Performance Of Other Neural Network Architecturesmentioning

confidence: 99%

“…Episodic Planning Networks augment recurrent-based reinforcement learning agents [26] with an external episodic memory implemented through a self-attention mechanism [44]. Our implementation was closely inspired by the implementation of AlKhammasi et al [43]. It is an augmentation of the LSTM-based meta-learner [26] used in our previous experiment with one additional input.…”

Section: Performance Of Other Neural Network Architecturesmentioning

confidence: 99%

“…Specifically, for each timestep i in the sequence, a single element of the sequence is a tuple of the previous timestep's observation o(i-1), the action taken at the timestep a(i), and the resulting observation after taking the action o(i). This tuple is meant to represent a single timestep's transition of the underlying MDP of the task (see AlKhammasi et al 2022) [43]. The sequence of these n tuples is then fed into a single transformer block (selfattention followed by an MLP, see AlKhammasi et al 2022 and Vaswani et al 2017) [45] and the output of the transformer block is then fed into the main LSTM of the agent.…”

Section: Plos Computational Biologymentioning

confidence: 99%

See 2 more Smart Citations

Disentangling Abstraction from Statistical Pattern Matching in Human and Machine Learning

Kumar,

Dasgupta,

Daw

et al. 2023

PLoS Comput Biol

View full text Add to dashboard Cite

The ability to acquire abstract knowledge is a hallmark of human intelligence and is believed by many to be one of the core differences between humans and neural network models. Agents can be endowed with an inductive bias towards abstraction through meta-learning, where they are trained on a distribution of tasks that share some abstract structure that can be learned and applied. However, because neural networks are hard to interpret, it can be difficult to tell whether agents have learned the underlying abstraction, or alternatively statistical patterns that are characteristic of that abstraction. In this work, we compare the performance of humans and agents in a meta-reinforcement learning paradigm in which tasks are generated from abstract rules. We define a novel methodology for building “task metamers” that closely match the statistics of the abstract tasks but use a different underlying generative process, and evaluate performance on both abstract and metamer tasks. We find that humans perform better at abstract tasks than metamer tasks whereas common neural network architectures typically perform worse on the abstract tasks than the matched metamers. This work provides a foundation for characterizing differences between humans and machine learning that can be used in future work towards developing machines with more human-like behavior.

show abstract

Section: Performance Of Other Neural Network Architecturesmentioning

confidence: 99%