File size: 2,006 Bytes
077605d | 1 2 3 4 5 6 7 8 9 10 11 12 13 | # LoRA: method, efficiency and experimental evidence
Using the supplied LoRA v2 paper, create an eight-slide English PPTX (16:9) with matching 1280×720 PNGs for a graduate research meeting whose audience knows basic deep learning. Organize the problem, method, experimental evidence and applicability limits yourself. No template or full-paper translation is required.
Explain frozen W0, trainable low-rank A/B and addition of the two branches. Explain matrix dimensions, rank and scaling with at least one original method diagram. Explain why merging weights can avoid an additional inference branch and the resulting limitation on mixed-task batches. Do not imply that fewer trainable parameters eliminate the base model at deployment.
Use GPT-2 Medium and Large from Table 3, comparing full fine-tuning and LoRA separately for each. Show trainable parameter counts, BLEU and ROUGE-L, retaining model scale, E2E NLG task and metric names. Calculate any reduction factors from this table and label them; do not reuse GPT-3's 10,000× figure. Error ranges may be omitted, but do not interpret ± results as certain upper bounds or claim new significance tests.
Also explain one efficiency benefit from §4.2 with its experimental conditions, and at least one limitation. If mentioning no additional inference latency, state the weight-merging prerequisite. Later work such as QLoRA is outside scope. Distinguish paper-reported values from your calculations.
Academic preferences: projection-readable formulas/charts, restrained visuals and clear argument structure. Each slide needs a speaking point; avoid screenshots of long paragraphs. Use editable native tables for experiments and editable vector objects for the method diagram. Local excerpts of paper figures are allowed, but whole-page screenshots must not substitute for slides.
Deliver artifact/slides.pptx, 01.png–08.png and sources.md (page/table/version/calculation locations). Inputs are read-only; put intermediate files in scratch. |