One evaluation boundary
All six approaches face the same four-class, 48×48 held-out test set, so the comparison reflects performance rather than inconsistent measurement.
Independent build · Evaluated machine learning
Facial-expression classification
The strongest model was selected from the error pattern—not the prestige of its architecture.
6approaches compared
76.56%held-out test accuracy
4expression classes
The constraint
A headline accuracy score can hide the class-level failures that determine whether a model is operationally useful.
Non-obvious diagnosis
Six approaches needed the same held-out boundary, followed by confusion-matrix analysis and a deployment recommendation bounded to the actual experiment.
The build starts here. The feature set follows from this diagnosis.What the system makes possible
Model selection becomes useful when the evaluation boundary and the operating failure are both explicit.
An MIT Professional Education capstone comparing six deep-learning approaches on four low-resolution grayscale expression classes, including a purpose-built five-block convolutional network.
Architecture comparison
Custom convolutional designs and transfer-learning baselines are tested side by side instead of assuming the most prestigious architecture will fit the task.
Class-level diagnosis
Confusion matrices and precision, recall, and F1 expose which expressions fail and in which direction.
Bounded deployment policy
Low-confidence and ambiguous cases remain human-reviewed rather than extending the model beyond what the experiment established.
See it working
Inspect the mechanism, then inspect the evidence.
The custom CNN reached 76.56% held-out test accuracy and outperformed the frozen ImageNet baselines for this narrow dataset. Sad/neutral ambiguity remained explicitly human-reviewed.
What the headline hides76.56%held-out test accuracy
The choice was not made on accuracy alone. The diagnostic work isolated the consequential error—sad expressions collapsing into neutral—and converted that failure pattern into a human-review boundary.
- CompareSix architectures
- InspectPer-class error
- DeployWith explicit limits
Final deck · model verdict
The common evaluation boundary produced a clear decision: the best-fit custom CNN beat the strongest transfer model by 15 percentage points and made roughly 40% fewer errors.
Final deck · failure diagnosis
The error pattern—not the headline score—exposed neutral as the failure sink and supported a 90.6% auto-route-or-review decision rule.
Final deck · deployment decision
The recommendation converts measured error into an operating policy: automate high-confidence cases, route ambiguity to review, and gate expansion on accuracy, recall, and fairness.
Final deck · actual dataset
The source material and held-out boundary are visible before the result: 20,214 grayscale faces across four classes, evaluated with class-balanced metrics.
How the parts compound
A feature matters when it improves the next decision.
- 01Normalize
Fix the task, classes, image boundary, split, and preprocessing.
- 02Compare
Train six approaches against the same evaluation contract.
- 03Diagnose
Inspect confusion patterns and per-class performance, not accuracy alone.
- 04Select
Choose the model whose measured behavior best fits the narrow task.
- 05Bound
Define human-review conditions and keep the claim inside the experiment.
The decisions that mattered
Architecture follows the failure mode.
The implementation is the visible surface. These choices determined whether it could solve the underlying problem.
- 01
Hold the boundary constant
Compare every approach on the same four-class 48×48 evaluation set.
- 02
Inspect asymmetric error
Use confusion matrices and class-level metrics to see which categories collapse into one another.
- 03
Bound the claim
Recommend human review for low-confidence cases rather than generalizing beyond the measured task.
Evidence classThird-party-visible scored work
Source
Final capstone report and verified Great Learning ePortfolio for the MIT Professional Education program.
What it establishes
Six models share one evaluation boundary; the custom CNN reached 76.56%; class-level diagnostics shaped the deployment recommendation.
Boundary
Specific to this four-class 48×48 grayscale dataset. This is facial-expression classification, not identity recognition.