Search research articles

ABOUT JoVE

Overview Leadership Blog JoVE Help Center

AUTHORS

Publishing Process Editorial Board Scope & Policies Peer Review FAQ Submit

LIBRARIANS

Testimonials Subscriptions Access Resources Library Advisory Board FAQ

RESEARCH

JoVE Journal Methods Collections JoVE Encyclopedia of Experiments Archive

EDUCATION

JoVE Core JoVE Business JoVE Science Education JoVE Lab Manual Faculty Resource Center Faculty Site

Terms & Conditions of Use

Related Concept Videos

Sequence Networks of Rotating Machines

Sequence Networks of Rotating Machines

A Y-connected synchronous generator, grounded through a neutral impedance, is designed to produce balanced internal phase voltages with only positive-sequence components. The generator's sequence networks include a source voltage that is exclusively in the positive-sequence network. The sequence components of line-to-ground voltages at the generator terminals illustrate this configuration.
Zero-sequence current induces a voltage drop across the generator's neutral impedance and other...

Signal Sequences and Sorting Receptors

Signal Sequences and Sorting Receptors

Signal sequences are short amino acid sequences that guide newly synthesized proteins to their proper location within the cell. Classical signal sequences are fifteen to sixty amino acids long and present at the N-terminus of a polypeptide chain. Each signal sequence has a conserved segment of basic residues towards their N terminus, a hydrophobic core, and a C-terminus rich in polar residues. The C-terminus also contains a signal cleavage site and features a -3 -1 sequence motif. The -3-1...

Classification of Systems-I

Classification of Systems-I

Linearity is a system property characterized by a direct input-output relationship, combining homogeneity and additivity.
Homogeneity dictates that if an input x(t) is multiplied by a constant c, the output y(t) is multiplied by the same constant. Mathematically, this is expressed as:

Neural Circuits

Neural Circuits

Neural circuits and neuronal pools are two of the main structures found in the nervous system. Neural circuits are networks of neurons that work together to carry out a specific task or process. They consist of interconnected neurons and glial cells, which provide structural and metabolic support.
Neuronal pools are collections of nerve cells with similar functions and interact through chemical and electrical signals. These pools include both interneurons (the central neural circuit nodes that...

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by

Same author

Geometry and dynamics of annealed optimization in the coherent Ising machine with hidden and planted solutions.

Physical review. E·2026

Same author

Intrinsic space-time couplings governing multi-scale cortical dynamics.

bioRxiv : the preprint server for biology·2026

Same author

Tracking the Fidelity of Internal Neural Representations with Error-In-Variables Regression.

bioRxiv : the preprint server for biology·2026

Same author

Causal Interpretation of Neural Network Computations with Contribution Decomposition.

ArXiv·2026

Same author

David Sussillo.

Neuron·2026

Same author

Gaussian Process Inference Reveals Non-Separability of Position and Velocity Tuning in Grid Cells.

Hippocampus·2026

Search research articles

Related Experiment Video

Updated: Dec 12, 2025

Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention

Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention

Published on: December 15, 2023

Reverse engineering recurrent networks for sentiment classification reveals line attractor dynamics.

Niru Maheswaranathan¹, Alex H Williams², Matthew D Golub²

¹Google Brain, Google Inc., Mountain View, CA.

Advances in Neural Information Processing Systems

|August 13, 2020

Summary

Researchers reverse-engineered recurrent neural networks (RNNs) for sentiment analysis. They discovered a universal, interpretable line attractor mechanism explaining how RNNs process language data across various architectures.

Related Experiment Videos

Last Updated: Dec 12, 2025

Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention

Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention

Published on: December 15, 2023

Area of Science:

Artificial Intelligence
Computational Neuroscience
Natural Language Processing

Background:

Recurrent Neural Networks (RNNs) are powerful tools for sequential data but often function as black boxes.
Understanding the internal workings of trained RNNs for specific tasks remains a significant challenge.

Purpose of the Study:

To reverse-engineer trained RNNs and provide a quantitative, interpretable description of their task-solving mechanisms.
To investigate how RNNs perform sentiment classification using dynamical systems analysis.

Main Methods:

Applied dynamical systems analysis to trained RNNs performing sentiment classification.
Identified fixed points of recurrent dynamics and linearized the system around them.
Analyzed the topological structure of fixed points and linearized dynamics.

Main Results:

Trained RNNs converge to highly interpretable, low-dimensional representations, contrary to their theoretical capacity for complex computations.
An approximate line attractor was identified within the RNNs, quantitatively explaining sentiment analysis.
This line attractor mechanism was observed consistently across different RNN architectures (LSTMs, GRUs, vanilla RNNs) and datasets.

Conclusions:

Dynamical systems analysis provides a powerful lens for interpreting RNN computations.
Interpretable, low-dimensional mechanisms like line attractors are surprisingly universal in trained RNNs.
These findings offer a path toward understanding and potentially designing more transparent neural network models.