Deep Reinforcement Learning's Struggle with Visuospatial Reasoning: Insights from the Same-Different Task
Bibliographic record
Abstract
Deep learning has dramatically changed the landscape of computational visual systems. One such prominent example is deep reinforcement learning, which is a type of machine learning solution that has seen numerous applications for a wide range of problems spanning game playing to finance, health care, natural language processing and embodied agents. We are interested in embodied agents that are free to visually examine their 3D environment, i.e., are active observers. We will show that deep reinforcement learning struggles to learn the fundamental visuospatial capability that is effortless for humans and birds, rodents and even insects. In order to collect data for our argument, we created a 3D physical version of the classic Same-Different task: Are two stimuli the same? The task was found to be easily solvable by human subjects with high accuracy from the first trial. Using human performance as the baseline, we sought to determine whether reinforcement learning could also solve the task. We have explored several reinforcement learning frameworks, including SAC, PPO, Imitation Learning and Curriculum Learning. Curriculum learning emerged as the only viable approach, but only when the task is simplified significantly to the point that it has only distant relevance to the original human task. Even with curriculum learning, the learned strategies differed significantly from human behaviour. Models exhibited a strong preference for a very limited set of viewpoints, often fixating on the same location repeatedly, lacking the flexibility and efficiency of human visuospatial problem-solving. Conversely, the outcomes of the human experiment were instrumental in developing a curriculum lesson plan that improved learning performance. Our human subjects seemed to develop correct strategies from the first trial and then, over additional trials, became more efficient, not more accurate. Reinforcement learning methods do not seem to have the foundation to match such human abilities.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.008 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".