← Back to Paper Reading

Build a foundation for reading methods papers through classic vision architectures, training choices, and evaluation evidence.

Reader question

Which design choices and evidence from a classic model still transfer to today’s systems?

Reading library

Read this topic in depth

Every paper reading currently filed under this research topic.

  • ResNet: Residuals Make Depth Trainable, but ImageNet 2015 Is Not a Ready-Made Detection or ViT Contract

    Intermediate Build the foundations first
    ResNet deep reading: Part 1 , Note: Aug 28, 2026 , Paper: 2016 , CV

    A source-grounded reading of He et al., CVPR 2016 / arXiv:1512.03385: identity shortcuts let stacked layers learn residual F(x)+x and fix plain-net degradation. ResNet-152 reaches 4.49% top-5 validation error on ImageNet; this is 2015 classification evidence, not a YOLO, ViT, or modern ConvNet leaderboard contract.

    Understand it in 90 seconds
    Problem
    After batch normalization and good initialization made tens of layers trainable, stacking more plain conv layers still triggers degradation—deeper models show higher training error (Figure 1, Figure 4 left), which is not ordinary overfitting.
    Core insight
    Recast the target mapping. Instead of asking stacked nonlinear layers to fit $\mathcal{H}(\mathbf{x})$ directly, let them fit $\mathcal{F}(\mathbf{x}):=\mathcal{H}(\mathbf{x})-\mathbf{x}$ and output $\mathbf{y}=\mathcal{F}(\mathbf{x})+\mathbf{x}$ through identity shortcuts (Equation 1). The control point is whether the solver must fit $H(x)$ from scratch or learn a correction $F(x)$ on top of identity.
    Strongest evidence
    On ImageNet with matched parameter counts, plain-34 top-1 error is 28.54% versus plain-18 27.94%, while ResNet-34 is 25.03% and beats ResNet-18 27.88% (Table 2, 10-crop validation). On CIFAR-10, plain-56 training error exceeds 60% and is omitted from Figure 6 left, while ResNet depth scans down to ResNet-110 6.43% (Table 6, Figure 6). ResNet-152 single-model top-5 validation error is 4.49%; a six-model ensemble reaches 3.57% top-5 on test (Tables 4–5).
    Main boundary
    The headline contract is 2012 ImageNet classification plus CIFAR-10 depth diagnostics; PASCAL/COCO detection is a Faster R-CNN backbone transfer table (Tables 7–8), not a YOLO contract, not ViT, and not a modern ConvNet leaderboard.
    Read the full deep dive
  • YOLO: Detect From One Full-Image Pass, but VOC 2016 Does Not Represent the Later YOLO Family

    Intermediate Build the foundations first
    Original YOLO deep reading: Part 1 , Note: Aug 28, 2026 , Paper: 2016 , CV

    A source-grounded reading of Redmon et al., CVPR 2016 / arXiv:1506.02640: object detection as a single forward-pass regression—an S×S grid, B boxes, and C class probabilities in one shot. On VOC 2007, YOLO reaches 63.4% mAP at 45 FPS; this is 2016 unified-detection evidence, not YOLOv3 COCO or Ultralytics product numbers.

    Understand it in 90 seconds
    Problem
    Pre-2016 detectors (DPM, R-CNN, Fast/Faster R-CNN) split proposals, features, scoring, and post-processing into separate stages that are hard to optimize end-to-end and slow at test time (for example R-CNN over 40 seconds per image; Fast R-CNN about 0.5 FPS).
    Core insight
    Recast detection as single regression: one CNN maps the full image directly to spatially separated bounding boxes and class probabilities. The control point is one-shot global reasoning versus two-stage propose-then-classify; the whole pipeline is one network trained end-to-end on detection loss (Sections 1-2, Figures 1-2).
    Strongest evidence
    PASCAL VOC 2007 (Table 1, train 2007+2012): YOLO 63.4% mAP at 45 FPS (Titan X, no batching); Fast YOLO 52.7% mAP at 155 FPS. Same table: Fast R-CNN 70.0% mAP at 0.5 FPS; Faster R-CNN VGG-16 73.2% mAP at 7 FPS. Figure 4: YOLO's top error bucket is localization at 19.0%; background false positives are 4.75% versus Fast R-CNN 13.6%.
    Main boundary
    Coarse grid (two boxes and one class set per cell), VOC's 20 classes, not instance segmentation; VOC 2012 test 57.9% mAP trails leaderboard leaders. YOLOv2/v3/v8, COCO 2017, and Ultralytics product mAP are not in this PDF; ResNet-152 ImageNet 4.49% is not a detection contract either.
    Read the full deep dive
  • AlexNet Part 2: Turning the Training Recipe into Testable Design Choices

    Intermediate Build the foundations first
    AlexNet Deep Dive: Part 2 , Note: Mar 19, 2026 , Paper: 2012 , CV

    A source-grounded reading of ReLU, multi-GPU splitting, overlapping pooling, augmentation, and dropout in Figure 1–3 and Sections 3–6.

    Understand it in 90 seconds
    Problem
    a 60M-parameter CNN can overfit even with 1.2M images, and its training recipe must be separated from the competition result.
    Core insight
    random crop/flip, RGB PCA lighting jitter, and dropout change or regularize the effective training distribution; SGD, momentum, weight decay, and a learning-rate schedule make Part 1's architecture converge.
    Strongest evidence
    color augmentation reduces top-1 error by over 1%, overlapping pooling by 0.4/0.3 points, and the full system reaches 37.5/17.0 on ILSVRC-2010 and 15.3 top-5 in 2012 (Sections 4–6; Table 1).
    Main boundary
    these ablations belong to the era's architecture, data, and compute; they do not show every modern vision model needs ten-crop, LRN, or the same schedule.
    Read the full deep dive
  • AlexNet Part 1: Reading the Evidence Behind an ImageNet Turning Point

    Introductory Build the foundations first
    AlexNet Deep Dive: Part 1 , Note: Mar 18, 2026 , Paper: 2012 , CV

    A source-grounded rereading of AlexNet’s problem, evaluation, historical result, and evidence boundary.

    Understand it in 90 seconds
    Problem
    in 2012, training a deep CNN on millions of high-resolution images was constrained by optimization speed, GPU memory, and overfitting.
    Core insight
    AlexNet is not one “big network” trick: convolutional locality, ReLU, a constrained two-GPU split, and an eight-layer architecture form a trainable system.
    Strongest evidence
    ILSVRC-2010 top-1/top-5 error is 37.5%/17.0%; the 2012 competition top-5 error is 15.3% versus 26.2% for second place (Section 6; Table 1).
    Main boundary
    LRN, the two-GPU split, and some kernel choices are hardware-era tradeoffs, not claims of modern optimality.
    Read the full deep dive

For speaking invitations, internal engineering sessions, or architecture exchange, see the topics and public work I can bring into the conversation.

Speaking & contact