Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates.
Shaping Sequence Attractor Schema in Recurrent Neural Networks
Authors: Zhikun Chu, Bo Ho, xiaolong zou, Yuanyuan Mi
NeurIPS 2025 | Venue PDF | LLM Run Details | Input Tokens: 18,152 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 4,142 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We train recurrent neural networks on an odor-sequence task using a shaping protocol inspired by well-established paradigms in experimental neuroscience. Our model provides the first systematic reproduction of key features of schema learning observed in the orbitofrontal cortex, including rapid adaptation to novel tasks, structured neural representation geometry, and progressive dimensionality compression during learning. |
| Researcher Affiliation | Academia | Zhikun Chu1 EMAIL Bo Hong3 EMAIL Xiaolong Zou3,B EMAIL Yuanyuan Mi2,B EMAIL 1, Bioengineering College,School of Medicine, Chongqing University. 2, Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing, China. 3, Biomedical Engineering,School of Medicine, Tsinghua University. |
| Pseudocode | No | The paper describes the model structure and task design with mathematical equations (e.g., Eq. 1, 2, 3), but it does not contain explicit pseudocode or algorithm blocks. |
| Open Source Code | No | Code is not publicly available at submission, but Section 3 and technical appendices provide detailed instructions to reproduce the main results. |
| Open Datasets | Yes | Applying our method to a keyword spotting task shows that shaping facilitates the rapid development of sequence attractor schemas, leading to enhanced learning efficiency. The task involves four spoken words water, wash, year, and had each composed of basic phonemes forming a hierarchical structure (Fig. 6a). These words are randomly selected from the TIMIT dataset [19]. |
| Dataset Splits | No | The paper mentions using the TIMIT dataset and describes how the odor-sequence tasks are constructed, but it does not specify explicit training, validation, or test dataset splits (e.g., percentages or sample counts) for any of the tasks. |
| Hardware Specification | Yes | Experiments ran on an Intel i9-14900k CPU with 32GB RAM, no GPU used. Each training took 30 minutes, with 200 runs for tuning and validation. |
| Software Dependencies | No | The paper mentions using the "Adam optimizer" but does not specify any software libraries or their version numbers (e.g., Python, PyTorch, TensorFlow, etc.) that would be needed to replicate the experiments. |
| Experiment Setup | Yes | We employ an RNN model to demonstrate schema learning via shaping on the odor-sequence task. The RNN comprises three components: an input layer with M units, a recurrent layer with N units, and an output layer with K units, as illustrated in Fig. 2a. At time t, given an input stimulus vector Im(t), the total input xn(t) to recurrent unit n evolves according to the following dynamics: dt = xn(t) + j=1 W rec nj rj(t) + m=1 W in nm Im(t), (1) here, W in and W rec represent the input-to-recurrent and recurrent-to-recurrent weight matrices, respectively. The activation of the unit n is computed as rn(t) = tanh(xn(t)), and τ denotes the time constant. The output yk(t) of unit k is given by: n=1 W out kn rn(t). (2) ... Each odor stimulus is represented as a one-hot vector of dimension M = 96, with the non-zero entry set to 20 (see Appendix Sec.A for details). Each stimulus presentation lasts Ts = 5 steps. ... Finally, the loss function used during task primitive learning is defined as a regression loss: y0:2(t) ˆyreward(t) 2 + β y2:K(t) ˆyclass(t) 2 . (3) The total loss is balanced between reward prediction and odor identity classification by a weighting factor β, which we set to 0.5 here. ... Across all three stages, we use the Adam optimizer. See Appendix Sec.A for more training details. |