AI0-001 AI Infrastructure and Technologies Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Post-training quantization of weights to 8-bit integers

  Why this is correct

  Post-training quantization converts 32-bit floating-point weights and activations to 8-bit integers, cutting model size by roughly four times and enabling faster integer arithmetic on constrained hardware. It requires no retraining, so the team can apply it to an existing model quickly. For a small set of object classes, the accuracy loss is usually small enough to remain acceptable on battery-powered cameras.

• ✓

  Pruning near-zero weight connections followed by fine-tuning

  Why this is correct

  Pruning removes weight connections that contribute little to the output, reducing parameter count and compute. Fine-tuning after pruning lets the model recover most of the accuracy lost to the removed connections. Combined with an appropriate sparsity format, this lowers both memory footprint and inference cost, which matters for cameras with limited RAM and battery budgets.

• ✗

  Switching the training optimizer from SGD to AdamW

  Why it's wrong here

  The optimizer affects how weights are updated during training and can influence convergence speed and final accuracy, but it does not change the deployed model's size, memory footprint, or inference cost. Once training finishes, the optimizer state is discarded. Changing optimizers therefore does nothing to help the model fit on battery-powered cameras with limited RAM.

• ✗

  Training with a larger batch size on the same dataset

  Why it's wrong here

  Batch size influences gradient noise and training throughput, not the architecture or numerical precision of the exported model. A model trained with a larger batch has essentially the same parameter count and inference cost as one trained with a smaller batch. It does not reduce memory footprint or energy use on the cameras and is irrelevant to the deployment constraint.

• ✗

  Increasing the number of convolutional filters in each layer

  Why it's wrong here

  Adding filters increases the parameter count and the number of multiply-accumulate operations, which raises memory use and inference energy rather than reducing them. This direction improves representational capacity, which might raise accuracy, but it directly conflicts with the goal of shrinking the model for low-power cameras. It is the opposite of what constrained deployment requires.