Conducted on: 17/08/2025
MobileNetV2, U-Net, and EfficientNet
- Analyzed the shift from standard residuals to Inverted Residuals in MobileNetV2 and why connecting thin bottleneck layers is more memory-efficient.
- Discussed the intuition behind Linear Bottlenecks, specifically how non-linear activations like ReLU can destroy information in low-dimensional spaces.
- Compared the concatenation method in U-Net skip connections to the summation used in ResNet, noting how it preserves spatial texture for precise localization.
- Examined the Symmetric path of U-Net and how the decoder effectively reconstructs the image from the low-resolution context provided by the encoder.
- Critiqued traditional manual scaling methods (just adding layers vs. just adding channels) and why they eventually lead to accuracy saturation.
- Evaluated the Compound Scaling rule in EfficientNet as a method to balance depth, width, and resolution simultaneously for better FLOPs efficiency.
- Explored the role of Neural Architecture Search (NAS) in creating the EfficientNet-B0 baseline and how it optimizes for both accuracy and latency.
- Briefly touched upon the integration of Squeeze-and-Excitation blocks within the MBConv layers to help the model focus on the most important feature channels.
- RCNN
Anab Farooq
Fourth Year Attendees: Karaka Prasanth Naidu Sir, Manav Jain Sir.
Third Year Attendees: Green Kedia Sir, Mukil M Sir, Harshvardhan Saini Sir, Daksh Mor Sir, Mohd. Ashaz Khan Sir, Priyam Pritam Panda Sir, Abhinav Jha Sir, Dilshad Sir
Second Year Attendees: Anab, Arnav, Ritesh, Rajat, Arjav, Abhishek, Ayushman, Sreenandan, Anukul
Second Year: None