Interactive showcaseGITHUB ↗
AIT301 / GROUP RESEARCH / CIFAR-10
Rethinking theresidual network.
Five experimental mechanisms. One lightweight residual architecture. Explore how Res_UCN combines adaptive convolutions, channel-aware normalization and the Quinsh activation to investigate efficient image classification.
✦ A visual research exhibit, not a claim of separately validated component improvements.
FIG 01 / RESIDUAL SIGNAL FLOW VISUAL MODEL
Architecture visualization • conceptual
02 / ARCHITECTURE
Inside the network.
Four residual stages combining different convolution mechanisms, weighted normalization and Quinsh. Select a stage or component to see its purpose.
RES_UCN / FEATURE PIPELINECIFAR-10 · 10 CLASSES
Input32 × 32 × 3
↓
↓
GLOBAL AVERAGE POOLINGDropout 0.5 → Linear(10)Ten CIFAR-10 logits
Stage-channel counts follow the project README. The diagram summarizes the network; it is not a layer-by-layer export of trained weights.
STAGE INSPECTOR↗
STAGE 01 / 04
Early feature extraction
Processes the first image features using adaptive-scale convolution, channel weighting, weighted normalization and Quinsh activation.
CONCEPTUAL RESIDUAL BLOCK
Input x
↓ADSGC / CIAC
↓Weighted BatchNorm
↓Quinsh + shortcut
FEATURE CHANNELS64
MODULE FAMILYResidual block
Understand all five modules ↗
03 / YOUR CONTRIBUTION
Meet Quinsh.
A learnable gated activation with an explicit residual linear path. Change its parameters and observe how the output and gradient behave.
THE QUINSH FUNCTION
LEARNABLE / PARAMETRICQ(x) = γ · (x − δ) · σ(w · (x − δ)) + αx
Quinsh (editable) Swish ReLU
Illustrative function curves, mathematically generated in your browser. These are not measured training results or learned checkpoint values.
PARAMETER CONTROLS
Adjust the activation shape. These values are for exploration, not fitted Res_UCN parameters.
How strongly the gated Swish-like branch contributes.
Moves the nonlinear transition left or right.
Preserves a direct linear signal and gradient contribution.
Controls the slope of the sigmoid gate.
EXPLORE PRESETS
Beyond a fixed activation
The learnable shift and slope let Quinsh reshape its nonlinear response during training, rather than relying on one fixed curve.
A direct gradient path
The αx term supplies a residual linear contribution. It can help preserve signal flow, although it does not guarantee the absence of dead units.
Research origins
Inspired by Swish-style gating, parametric activations and residual learning. The formulation is described on slides 12–14 of the group presentation.
04 / COMPONENT EXPLORER
Five ideas. One system.
Explore the research rationale behind each module. Authorship is shown explicitly because the architecture was built as a team.
MODULE 01 · ACTIVATIONCONCEPT VIEW
YOUR CONTRIBUTION
Quinsh activation
Balances a Swish-like gated pathway with a residual linear branch to adapt the shape of the activation.
HOW IT WORKS
Shift the input by δ, adjust the sigmoid gate using w, scale the nonlinear branch by γ, then add αx.
RESPONSIBLE CONTRIBUTORWilbert Andrew Yonathan
05 / REPORTED EXPERIMENTS
Efficiency with context.
The original group's CIFAR-10 comparison is reproduced from slide 22. No component-specific ablation was reported.
RES_UCN VS BASELINECLASSIFICATION
METRICRESNET-18RES_UCN
Accuracy89.90%91.45% ↗
Parameters~11.2M~3.4M ↘
Training time24m 00s17m 17s ↘
Reported best epoch—18
TEST ACCURACY Percentage points
+1.55 percentage points compared with the reported baseline.
WHAT THE NUMBERS SUPPORT≠ ABLATION
+1.55 ppReported accuracy difference
~69.6%Approximate parameter reduction
~28%Approximate reduction in reported training time
Research boundary
These figures describe the combined group architecture under the reported experiment. They do not isolate Quinsh or demonstrate performance on datasets beyond CIFAR-10.
06 / INTERACTIVE DATASET LAB
The data behind the research.
Explore the ten CIFAR-10 class labels and inspect how a 32 × 32 image is prepared for a classifier. No model file, simulated prediction, or server is involved.
TRAIN SET50,000Standard CIFAR-10 specification
TEST SET10,000Standard CIFAR-10 specification
INPUT32 × 32Three RGB color channels
CATEGORIES106 animals · 4 vehicles
CLASS LABEL EXPLORER10 CLASSES
Icons are illustrative category symbols, not photos, not actual images from CIFAR-10.
SELECTED LABELDATASET REFERENCE
CLASS 03 / 10
Cat
One of the six animal categories in the CIFAR-10 image classification dataset.
Category AnimalStandard train/class 5,000Standard test/class 1,000
IMAGE PREPROCESSING PLAYGROUNDVISUAL EXPLANATION
From image to a 32 × 32 input
Use a category illustration or an image from your device. Flip it, adjust brightness, and inspect the actual resized pixels. Everything stays in your browser.
INPUT ILLUSTRATION / PHOTOCat · concept illustration
RESIZED RGB IMAGE32 × 32 RGB
These are educational image transformations; they do not reproduce the original training pipeline's exact stochastic settings or produce model predictions.
PIXEL & NORMALIZATION INSPECTORSLIDE 21
ACTUAL PIXEL MATH · NO INFERENCE
Understand the input.
Select a pixel in the 32 × 32 output image to inspect its raw RGB channels and normalized tensor values, using the means and standard deviations reported in the presentation.
SELECTED PIXEL (16, 16)
R
Red—
—G
Green—
—B
Blue—
—NORMALIZATION RULEnormalized = (channel / 255 − mean) / std
MEAN(0.49, 0.48, 0.44)
STD(0.20, 0.19, 0.20)
Research context
Slide 21 describes random cropping, flipping, and color jitter for training and tensor conversion with normalization at test time. This visual lab demonstrates the concepts; it does not run the Res_UCN network.
Source separation: the group's slide confirms CIFAR-10 and its preprocessing choices; dataset size, class balance, and labels come from the standard CIFAR-10 dataset specification. All gallery artwork is illustrative.
07 / PROJECT SCOPE
Built together. Credited clearly.
Res_UCN originated as an AIT301 group project. This website is a visual portfolio companion highlighting one contributor's specific research work.
INDIVIDUAL CONTRIBUTION
Q(x)
Quinsh activation function
Wilbert Andrew Yonathan designed and presented Quinsh, a parametric activation combining learnable shifting, gating and a linear residual path. This interactive explorer emphasizes that contribution.
Open Quinsh Studio ↗GROUP CONTRIBUTIONSAIT301 · 2025
ADSGCDerol Chong Chun Kit
CIACEdward Ooi Wee Xue
Weighted BatchNormWang XinYi
QuinshWilbert Andrew Yonathan
DE-StoppingChua Kian Hoong
Preprocessing & integrationSiah Jie Sheng
Contributions follow slide 2 of the original project presentation. The results belong to the overall group architecture.
END OF RESEARCH EXHIBIT