AS2 blends visual learning with soft logic, but perfect Sudoku results need a discrete decoder

Connecting perception and rules

In the arXiv preprint AS2 — Attention-Based Soft Answer Sets, Wael Abd-Almageed of Clemson University proposes an architecture combining neural perception, attention and continuous approximations to logical operations. The goal is to train perception and reasoning together without invoking an external symbolic solver during training or prediction.

Many neuro-symbolic systems connect a neural network to a separate reasoning engine. AS2 instead keeps probability distributions over possible symbols within its neural computation. Logical constraints can then contribute a training signal that flows back into the visual encoder. The reported experiments concern handwritten-digit puzzles, synthetic object scenes and digit addition—not unrestricted reasoning.

How soft constraints work

For Visual Sudoku, a convolutional network processes cell images, and a Transformer exchanges information across cells. Instead of conventional sequence-position embeddings, the model adds learned representations of each cell’s row, column and box membership. This describes the puzzle’s constraint structure rather than an arbitrary ordering of its cells.

The central training mechanism uses a probabilistic version of the immediate-consequence operator from logic programming. For Sudoku’s mutual-exclusivity constraints, a cell’s probability of taking a digit is multiplied by factors representing other group members not taking that digit. Competing assignments therefore reduce one another’s values.

Training penalizes the squared difference between the original probabilities and the operator’s output: a fixed-point residual. Under the stated normalization and group-size conditions, zero residual corresponds to a valid one-hot assignment. Unlike a simple penalty that merely checks whether probabilities sum correctly, this loss remains positive for a uniformly uncertain prediction.

That mathematical target does not guarantee that training reaches it, or that choosing each cell’s most probable digit produces a valid board. The model also receives direct supervision, including correct labels for blank cells; it is not learning Sudoku from constraints alone.

What the Sudoku results show

The Visual Sudoku experiment uses 9,000 training boards, 1,000 validation boards and 1,000 test boards, with approximately 45 clues per board. Raw predictions achieve 99.89% cell accuracy, but only 95.60% of boards are entirely correct. The reported constraint-satisfaction rate is also 95.60%.

Ten rounds of continuous constraint refinement raise board accuracy and constraint satisfaction to 98.70%. A separate confidence-ordered greedy decoder reaches 100% board accuracy and constraint satisfaction on these test boards.

This decoder locks perceived clues, commits uncertain cells to digits and removes those digits from neighboring candidates. It calls no external solver, but makes discrete choices and cannot backtrack. The perfect result therefore belongs to the model-plus-decoder combination, not the unprocessed neural output or a wholly differentiable final prediction procedure. The paper acknowledges that greedy choices can reach dead ends on other problems.

Comparisons with earlier Sudoku systems require caution. The reported methods differ in clue counts, access to digit labels, visual versus symbolic inputs and use of external solvers. Sudoku results are reported for a fixed test set without multi-seed variability.

Scene classification reveals a tradeoff

CLEVR-Hans tests classification of synthetic scenes containing several objects. Some training attributes correlate misleadingly with classes, while those shortcuts disappear at test time.

AS2 compares fixed, ground-truth logical rules with rule-initialized weights that can change during training. Here the explicit operator combines attribute probabilities into object-template matches and class scores in a single forward pass. It does not use Sudoku’s fixed-point residual loss or iterative refinement.

The crop-based explicit variant reaches 99.94% test accuracy on CLEVR-Hans3 and 87.77% on CLEVR-Hans7, with standard deviations of 0.04 and 1.50 percentage points across three seeds. The learned crop variant scores 78.19% and 82.38%, respectively. These results favor retaining explicit rules in these settings, but do not establish universal robustness to misleading correlations.

On the seven-class task, the published automated NeSy-CL baseline scores 90.97%, above AS2. Training setups differ: AS2 receives per-object attribute labels and ground-truth object coordinates, while its explicit variant also receives the correct rules. These are important sources of supplied information.

Scope and remaining limits

Digit-addition experiments extend the approach to arithmetic constraints, but do not show a consistent advantage. With eight digits, AS2 reports 99.01% sum accuracy, below its perception-only and pipeline baselines. It receives individual digit labels as well as sums, unlike several published comparisons, and selects addition checkpoints using test accuracy rather than a held-out validation set.

The supported logic remains restricted: default negation, choice rules and optimization statements are outside the demonstrated soft operator’s coverage. As a preprint, the work presents reported evidence rather than independently established performance. Its contribution is a way to integrate specified constraint patterns with neural learning—not proof of a general replacement for symbolic solvers.