Free NCA-GENM flashcards

In a latent diffusion U-Net, which mechanism directly uses CLIP text embeddings to steer the spatial features of the latent during denoising?

Cross-attention layers in the U-Net. The spatial feature map provides queries while the CLIP text token embeddings are linearly projected to keys and values, so every spatial position can attend to the most relevant text tokens. This gives a direct, token-level conditioning pathway that injects the text representation into the spatial features.

Self-attention inside the U-Net residual blocks injects the CLIP text embedding into the spatial features.

Self-attention relates spatial positions to each other within the feature map, not to an external conditioning signal such as a text embedding, so it cannot directly inject the CLIP representation. Cross-attention is the mechanism that brings token-level text into the spatial features.

Why is pixel-value normalisation applied to image tensors before neural-network training? A pipeline divides 16-bit greyscale pixels by 65535 to map them to 0 to 1.

Scaling inputs to a small bounded range keeps weight gradients at a workable magnitude, preventing the exploding gradients and overshooting weight updates that large activation values cause. Well-conditioned gradients throughout back-propagation improve training convergence, which is the canonical motivation for input normalisation.

Dividing pixel values by the maximum gives zero-mean, unit-variance standardised inputs for each channel.

Dividing by the maximum value scales to the unit interval but does not produce zero-mean output. Standardisation requires subtracting the dataset mean and dividing by its standard deviation, which is a different operation from scaling to 0 to 1.

In an ASR pipeline where the acoustic model already captures phonemes but rare technical terms are still misrecognised, which component should you target to reduce errors on those terms?

The language model component. It assigns probabilities to word sequences and can be updated with a custom vocabulary or n-gram model. Injecting domain vocabulary and phrase probabilities directly reduces substitution errors on rare terms that the acoustic model already produces as phoneme sequences.

The vocoder controls pronunciation fidelity of technical terms in ASR, so target it to fix recognition errors.

Vocoders are a TTS component that converts mel spectrograms into audio waveforms; they have no role in ASR decoding or vocabulary coverage. The language model is what ranks word-sequence hypotheses during recognition.

Why does the NIM microservice architecture of an AI Blueprint make swapping the embedding model practical?

AI Blueprints are reference architectures that wire together discrete NIM microservices as modular units. Because each service, such as an embedding model, is independently addressable, you can replace or reconfigure the embedding NIM with another of the same role while the rest of the pipeline remains intact.

An AI Blueprint packages all AI logic as a single compiled binary, so you swap the embedding model by recompiling with a different library flag.

Blueprints are not compiled monoliths; that misrepresents the composable microservice architecture and confuses it with traditional software packaging. They wire together discrete NIM microservices, so the embedding NIM is replaced as an independent unit, not via recompilation.