TildAlice – Tech Blog for Developers
Python, AI, Deep Learning, Data Analysis, Computer Vision, Reinforcement Learning โ tips and tutorials
Latest Posts
-
Speculative Decoding: How Medusa and EAGLE Speed Up LLMs
Medusa and EAGLE promise 2-3x LLM speedup via speculative decoding. Test results on LLaMA 2: acceptance rates, memory cost, and when it fails.
-
ViT Review: 16×16 Patches That Replaced Convolutions
ViT-B/16 hits 77.9% top-1 accuracy on ImageNet with 16x16 patches and no convolutions. We review the architecture, training tricks, and where it breaks.
-
dataclasses vs Pydantic vs attrs: Python Data Model Guide
dataclasses vs Pydantic vs attrs: real benchmarks, validation tradeoffs, and which to pick for production. Spoiler: speed isn't the deciding factor.
-
MAE Review: Self-Supervised Vision via Masked Reconstruction
MAE masks 75% of image patches and reconstructs pixelsโsimple, brutal, effective. Why this extreme masking ratio works and the asymmetric trick that made it fast.
-
Topological Sort: From Task Scheduling to Build Systems
Topological sort powers build systems and package managers. Kahn's vs DFS algorithms compared with real dependency graphs and failure modes.
-
NeRF vs Gaussian Splatting for Real-Time 3D Detection
NeRF vs Gaussian Splatting for real-time 3D object detection: KITTI benchmark results, where each fails, and which to use for dynamic scenes.
-
ViT vs DeiT Review: Pure Attention Meets Distillation
ViT vs DeiT: how distillation and data augmentation let DeiT match ViT accuracy without 300M pretraining images. Training logs and ablation results included.
-
Self-Supervised Contrastive Learning for Fault Diagnosis
SimCLR contrastive pretraining boosted fault diagnosis from 73% to 89% using 8,000 unlabeled vibration signals. Here's the exact PHM workflow.