Known corrections to the slide content. Each entry links to the affected slide.
L1 norm formula (sum of absolute values) labeled as "L2-Norm" in the blue box. Directly confuses students learning L1 vs L2 regularization.
Laplacian edge detection kernel labeled as "Edge Detection (Sobel)". Commentary repeats the error.
Query matrix diagram labeled "XΦ_k" (key subscript) instead of "XΦ_q" (query subscript). The slide's own formula and bullet text use the correct subscript.
Word2Vec labeled as "masked language model" on the slide. That's BERT. The slide even correctly labels BERT two lines below, contradicting its own Word2Vec label.
T5 listed as encoder-only alongside BERT. T5 is an encoder-decoder model. Commentary repeats the error.
Commentary says OpenAI models produce 768-dimensional embeddings. 768 is BERT-base (Google). OpenAI's embedding models use 1536 dimensions.
Google Gemini Ultra listed as a reasoning model alongside o1/o3-mini and DeepSeek R1. Gemini Ultra is the original large Gemini 1.0 model, not a reasoning-focused model.
States 20^10 = 10^12. Actual value is ~1.024 × 10^13, off by an order of magnitude. Both slide and commentary have the error.
Slide says ~168 GB VRAM for a 70B model at 16-bit, then immediately shows the math: 70B × 2 bytes = 140 GB. The two numbers contradict each other on the same slide.
States "3 customized amenity objects" — almost certainly should be 30, consistent with all surrounding slides. Commentary repeats the error.
24 consecutive slides have headings and commentary shifted one position ahead of the slide images. Every commentary from slide 13 onward describes the next slide's content.
Commentary shifted ~2 positions. Slides 1–3 describe Keras code that was removed from the web page; slides 4–6 commentary narrates earlier slides' content.
Slide says "Slope (w₁) and intercept (w₂)" but the model defined on Slide 25 is y = w₁ + w₂x, making w₁ the intercept and w₂ the slope. Labels are reversed.
Commentary says the green line is the noisy SGD (lr=0.1, mini-batch=1) and the blue line is smoother (lr=0.05, mini-batch=5). The slide legend shows the opposite: blue is noisy, green is smooth.
Slide lists "Privacy" as a soft objective. Commentary replaces it with "Fairness" — a different concept entirely.
Both Linear and ReLU output sections say "Appropriate for general regression problems." The ReLU description should specify non-negative regression. The commentary gets this right; the slide doesn't.
Commentary says "normalizing so the variance falls between -1 and 1." Variance becomes ~1 (a scalar); it's the data values that fall in [-1, 1].
Slide says "reduce [learning rate] by an order of magnitude" then gives examples (0.005, 0.0025, 0.001) that are reductions by factors of 2–2.5, not 10.
Commentary describes Q as K × M but the slide defines Q = V_k, which is m × k. The transposed dimensions could confuse students reconciling formulas.
Both slide and commentary describe F matrix entries as "how much item i increases proximity to concept d" — that describes Y (items × concepts), not F (users × items).
LLM training phases described as "three" (slides 2, 19), "four" (slide 11), and "two" (slide 12 commentary) without reconciliation.
ResNet slide cites arxiv link 1512.00567, which is the Inception v3 paper. The correct ResNet link is 1512.03385.
Commentary dates Word2Vec popularity to 2011. The Word2Vec paper was published in 2013.
Softmax chart plots 3 outputs (y1, y2, y3) but the slide's own text gives a 4-class example (Red, Green, Blue, Yellow). Internal contradiction.
Commentary describes manual checkpointing ("save weights, roll back if you overshoot") but the slide shows keras.callbacks.EarlyStopping — a different strategy.
Slide text says "selected 40" classes but the image labels on the same slide say "30 Classes."