Navigating Simply, Aligning Deeply: Winning Solutions for Mouse vs. AI 2025

AI summary

Overview

• Research area: Reinforcement learning for visual navigation, combined with computational neuroscience (brain-model alignment); computer vision.
• Technical level: Intermediate. Familiarity with convolutional networks, PPO-style reinforcement learning, and basic representation metrics helps, though the paper explains its architectural choices in accessible terms.
• Scope: A competition report describing the winning solutions from Team HCMUS_TheFangs for both tracks of the NeurIPS 2025 Mouse vs. AI: Robust Visual Foraging Competition — one for visual robustness under perturbed conditions, one for predicting mouse visual cortex neural activity.

What This Paper Is About

Artificial agents that navigate using vision tend to break down when the visual world changes — fog, different lighting, or other shifts away from their training conditions — while mice handle such changes with ease. This paper describes the two architectures the authors built to compete in a benchmark that scores both behavioral robustness (Track 1) and how well an agent's internal features predict real neural recordings from mouse visual cortex (Track 2). Their goal was to find what kinds of model designs actually achieve each objective, and to document which popular designs fail.

Key Contributions

1. A simplicity-first solution for visual robustness. A two-layer CNN with Gated Linear Units (GLU) and observation normalization scored 95.4% final score on Track 1, beating deeper baselines that have many times more parameters.
2. A deep, capacity-heavy solution for neural alignment. A 16-convolutional-layer ResNet-style architecture with GLU-based gating and 17.8 million parameters reached top-1 neural prediction performance on Track 2.
3. A systematic training-duration study. Ten checkpoints trained between 60K and 1.14M steps were evaluated, revealing that the relationship between training length and performance is non-monotonic, with strong results around 200K steps.
4. Documentation of failed approaches. InceptionNet, a 24-layer IMPALA ResNet, LSTM and SRU-LSTM variants, and a naive data-augmentation pipeline were all evaluated and reported as underperforming, along with an analysis of why.

Main Findings

• Simple architecture wins on robustness: The final Track 1 model (SimpleCNN + GLU + observation normalization) reached 95.4% final score, with an Average Success Rate (ASR) of 96.80% and Modified Success Rate (MSR) of 94.00% — a gap of only 2.8 percentage points between standard and perturbed conditions.
• Depth hurts under perturbation: A 24-layer IMPALA ResNet scored 65.98% final (ASR 80.96%, MSR 51.00%), a 30 percentage point gap that the authors attribute to overfitting, while a 4-layer IMPALA ResNet scored 87.70% final (ASR 91.40%, MSR 84.00%).
• Augmentation backfired: Adding a comprehensive augmentation pipeline (weather effects, color jitter, random masking) to the ResNet baseline dropped its final score from 87.7% to 59.8%, a 27.9 percentage point decrease.
• InceptionNet never converged: Over 500,000 training steps, the multi-branch model showed high-variance returns and no consistent improvement trend.
• Normalization was the single biggest robustness gain: Removing it cut the final score from 95.40% to 91.80%, a 3.8 percentage point drop.
• GLU contributed modestly but consistently: Removing GLU changed the final score from 91.80% to 91.60% (0.2 percentage points). Removing both normalization and GLU gave 91.60%, matching the removal of GLU alone in the reported table, which the authors describe as approximately additive contributions.
• Capacity drives neural alignment: The 17.8M-parameter deep model achieved competition ranks 1 through 6 across checkpoints, while the 1.4M-parameter Track 1 SimpleCNN reached rank 13 with a score of 0.1451. The 12.8x parameter difference translated to roughly a 4.6% improvement in neural alignment score.
• Training length is non-monotonic: Model 368923 trained for 1,139,998 steps held rank 1 with a score of 0.1517, but model 368751 trained for 199,996 steps took rank 2 at 0.1507 — only a 0.66% gap with 5.7x fewer steps. Model 368628 at 79,998 steps (rank 5) outperformed model 369015 at 519,976 steps (rank 6).
• Only the first GLU layer learns significantly in the deep model: The authors report subsequent GLU layers showed minimal parameter updates.
• Behavioral and neural objectives dissociate: The lightweight SimpleCNN achieved stronger behavioral robustness (95.4%) than most Track 2 checkpoints (89–91%) while trailing badly on neural metrics.

Methodology in Plain English

The task is a first-person foraging problem in a 3D Unity environment: the agent starts at a random position, sees the scene through an egocentric camera as an 86 × 155 grayscale image, and must steer toward a target sphere using three continuous controls (forward-backward velocity, strafing velocity, and rotation rate). Episodes end on reaching the target or timing out.

For Track 1, the authors deliberately built a small network: two convolutional layers with large kernels and aggressive striding (8 × 8 with stride 4, then 4 × 4 with stride 2) producing 16 then 32 channels, flattened and projected to 256 dimensions. On top of that they added a GLU module — two parallel paths, one passed through a Swish activation and one through a sigmoid gate, multiplied element-wise — which learns to suppress features that are unreliable under visual perturbation. They also normalized observations using running mean and standard deviation statistics, which removes global brightness and contrast shifts.

For Track 2, they scaled up deliberately: 16 convolutional layers in residual blocks, channel widths growing 64 → 128 → 256 → 512, with GLU gating for adaptive feature routing, ending in separate policy and value heads. The model has 17.8 million parameters versus 1.4 million for the Track 1 model.

Both models were trained with Proximal Policy Optimization (PPO) using the ML-Agents toolkit on NVIDIA GPUs. Track 1 used a two-phase strategy: 1,400,000 steps on the convolutional backbone, then 350,000 additional steps with the GLU module added from the best Phase 1 checkpoint. Track 2 involved training runs from 60,000 to 1,140,000 steps with checkpoints saved every 20,000 steps, and the authors submitted and compared multiple checkpoints rather than only the final one. Training took 6–8 hours for Track 1 and 12–24 hours for Track 2 depending on step count.

Neural alignment was measured with a linear readout R²: a ridge regression was fitted to predict neural responses from the agent's features, with cross-validation for regularization strength. A secondary metric compared the geometry of agent and neural representational spaces via correlation of representational dissimilarity matrices. The neural data came from 19,000+ neurons in primary visual cortex (V1) and higher visual areas, recorded with two-photon calcium imaging while mice performed the same foraging task.

Why This Matters

Impact on research. The paper pushes back on the assumption that harder visual tasks demand deeper networks. It documents a concrete case where a 1.4M-parameter model beats much larger ones on robustness, and where naive data augmentation actively harms performance. It also shows that behavioral competence and biological plausibility can pull architectures in opposite directions — a useful result for anyone building brain-aligned models. The non-monotonic training-duration finding is a practical warning against submitting converged checkpoints by default.

Real-world applications:

• Robots and drones operating outdoors, where fog, glare, and day-night lighting changes are the dominant failure mode.
• Autonomous inspection or agricultural robots trained in controlled conditions that must work in variable real-world illumination.
• Search-and-rescue or delivery agents that need reliable target detection under degraded visibility.
• Neurotechnology and brain-computer interfaces, where models whose internal representations match cortical activity can serve as testable hypotheses about biological visual processing.

Industry relevance. Practitioner teams spend heavily on scaling architectures and augmentation pipelines. This paper's evidence that observation normalization alone yields 3.8 percentage points of robustness, while complex augmentation costs 27.9 points in their setup, argues for cheaper, more targeted engineering. The checkpoint-selection finding — getting near-top performance at 200K steps instead of 1.14M — has direct cost implications for architecture search and hyperparameter tuning.

Future Directions

• Test generality beyond this task. The authors note their results may not transfer to richer visual scenes, longer planning horizons, or more diverse perturbation types, and call for evaluation on other visual control tasks.
• Move beyond linear readout metrics. Current alignment scoring captures only certain aspects of biological vision; the authors suggest metrics incorporating dynamics, attention, or causal structure.
• Explore transformer-based architectures. Self-attention could combine the robustness benefits of attention with architectural simplicity, and the authors flag Vision Transformers for robust visual RL as an open direction.
• Add training techniques and multimodal sensing. Curiosity-driven exploration, hindsight experience replay, and meta-learning remain untried, and mice use vision together with proprioception and vestibular information, whereas these agents receive only egocentric images.

Target Audience

Reinforcement learning and computer vision researchers working on visual robustness or sim-to-real generalization will find the failed-approach documentation and ablation tables directly actionable. Computational neuroscientists interested in brain-model alignment will value the Track 2 capacity analysis and the checkpoint study. Competition participants and applied ML engineers looking for practical architecture and training-budget guidance are a secondary audience, though they should note that at least one figure in Table 3 (removing GLU alone and removing both components yield identical scores) is reported as-is without further explanation in the paper.

Authors’ abstract

Visual robustness and neural alignment remain critical challenges in developing artificial agents that can match biological vision systems. We present the winning approaches from Team HCMUS_TheFangs for both tracks of the NeurIPS 2025 Mouse vs. AI: Robust Visual Foraging Competition. For Track 1 (Visual Robustness), we demonstrate that architectural simplicity combined with targeted components yields superior generalization, achieving 95.4% final score with a lightweight two-layer CNN enhanced by Gated Linear Units and observation normalization. For Track 2 (Neural Alignment), we develop a deep ResNet-like architecture with 16 convolutional layers and GLU-based gating that achieves top-1 neural prediction performance with 17.8 million parameters. Our systematic analysis of ten model checkpoints trained between 60K to 1.14M steps reveals that training duration exhibits a non-monotonic relationship with performance, with optimal results achieved around 200K steps. Through comprehensive ablation studies and failure case analysis, we provide insights into why simpler architectures excel at visual robustness while deeper models with increased capacity achieve better neural alignment. Our results challenge conventional assumptions about model complexity in visuomotor learning and offer practical guidance for developing robust, biologically-inspired visual agents.