An official website of the United States government
Autonomous maze navigation is appealing yet challenging in soft robotics for exploring priori unknown unstructured environments, as it often requires human-like brain that integrates onboard power, sensors, and control for computational intelligence. Here, we report harnessing both geometric and materials intelligence in liquid crystal elastomer–based self-rolling robots for autonomous escaping from complex multichannel mazes without the need for human-like brain. The soft robot powered by environmental thermal energy has asymmetric geometry with hybrid twisted and helical shapes on two ends. Such geometric asymmetry enables built-in active and sustained self-turning capabilities, unlike its symmetric counterparts in either twisted or helical shapes that only demonstrate transient self-turning through untwisting. Combining self-snapping for motion reflection, it shows unique curved zigzag paths to avoid entrapment in its counterparts, which allows for successful self-escaping from various challenging mazes, including mazes on granular terrains, mazes with narrow gaps, and even mazes with in situ changing layouts.
more »
« less
Linearly Structured World Representations in Maze-Solving Transformers
The emergence of seemingly similar representations across tasks and neural architectures suggests that convergent properties may underlie sophisticated behavior. One form of representation that seems particularly fundamental to reasoning in many artificial (and perhaps natural) networks is the formation of world models, which decompose observed task structures into re-usable perceptual primitives and task-relevant relations. In this work, we show that auto-regressive transformers tasked with solving mazes learn to linearly represent the structure of mazes, and that the formation of these representations coincides with a sharp increase in generalization performance. Furthermore, we find preliminary evidence for Adjacency Heads which may play a role in computing valid paths through mazes.
more »
« less
- Award ID(s):
- 2110745
- PAR ID:
- 10636462
- Publisher / Repository:
- Proceedings of UniReps: the First Workshop on Unifying Representations in Neural Models, PMLR
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Although engineers can control the internal geometry of materials down to the micro-scale, it is unclear what configuration is ideal for a given transport process. We explore the use of mazes as abstract representations of two-phase systems. Mazes can be easily generated using many different algorithms and then represented as graphs for analysis. The three, dimensionless graph parameters of effective tortuous resistance, average tortuosity, and minimum-cut-size were derived and then correlated to the maze’s effective transport property (e.g., permeability), average residence time, and robustness, respectively. It was shown that by tuning the settings of the maze algorithm, one can obtain desired maze performance. Finally, a composite maze was constructed and shown to mimic the geometry and permeability of a real commercial membrane. In principle, a surrogate maze geometry can be optimized/tuned for a given transport process and then used to guide the rational design of the engineered system it represents.more » « less
-
Weitzenfeld, A (Ed.)In the last decade, studies have demonstrated that hippocampal place cells influence rats’ navigational learning ability. Moreover, researchers have observed that place cell sequences associated with routes leading to a reward are reactivated during rest periods. This phenomenon is known as Hippocampal Replay, which is thought to aid navigational learning and memory consolidation. These findings in neuroscience have inspired new robot navigation models that emulate the learning process of mammals. This study presents a novel model that encodes path information using place cell connections formed during online navigation. Our model employs these connections to generate sequences of state-action pairs to train our actor-critic reinforcement learning model offline. Our results indicate that our method can accelerate the learning process of solving an open-world navigational task. Specifically, we demonstrate that our approach can learn optimal paths through open-field mazes with obstacles.more » « less
-
Understanding neural representations will help open the black box of neural networks and advance our scientific understanding of modern AI systems. However, how complex, structured, and transferable representations emerge in modern neural networks has remained a mystery. Building on previous results, we propose the Canonical Representation Hypothesis (CRH), which posits a set of six alignment relations to universally govern the formation of representations in most hidden layers of a neural network. Under the CRH, the latent representations (R), weights (W), and neuron gradients (G) become mutually aligned during training. This alignment implies that neural networks naturally learn compact representations, where neurons and weights are invariant to task-irrelevant transformations. We then show that the breaking of CRH leads to the emergence of reciprocal power-law relations between R, W, and G, which we refer to as the Polynomial Alignment Hypothesis (PAH). We present a minimal-assumption theory demonstrating that the balance between gradient noise and regularization is crucial for the emergence the canonical representation. The CRH and PAH lead to an exciting possibility of unifying major key deep learning phenomena, including neural collapse and the neural feature ansatz, in a single framework.more » « less
-
Much of the progress in contemporary NLP has come from learning representations, such as masked language model (MLM) contextual embeddings, that turn challenging problems into simple classification tasks. But how do we quantify and explain this effect? We adapt general tools from computational learning theory to fit the specific characteristics of text datasets and present a method to evaluate the compatibility between representations and tasks. Even though many tasks can be easily solved with simple bag-of-words (BOW) representations, BOW does poorly on hard natural language inference tasks. For one such task we find that BOW cannot distinguish between real and randomized labelings, while pre-trained MLM representations show 72x greater distinction between real and random labelings than BOW. This method provides a calibrated, quantitative measure of the difficulty of a classification-based NLP task, enabling comparisons between representations without requiring empirical evaluations that may be sensitive to initializations and hyperparameters. The method provides a fresh perspective on the patterns in a dataset and the alignment of those patterns with specific labels.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Workshop Report:
- The DOI is not currently available.
