NVIDIA-Certified Associate: Generative AI Multimodal (NCA-GENM) practice questions

A pipeline ingests 16-bit greyscale medical images and feeds them into a convolutional network. Before training, each pixel value is divided by 65535 so the resulting tensor contains values in the range 0 to 1. What is the primary reason for this operation?

• ATo normalise input values so that large pixel magnitudes do not cause unstable gradients during back-propagation, improving training convergence.check_circle Correct
• BTo reduce the spatial resolution of the image so that fewer parameters are needed in the first convolutional layer.
• CTo convert the integer pixel values into a floating-point representation that is compatible with the network's activation functions and loss computation.
• DTo apply zero-mean standardisation so that each channel has a mean of zero and a standard deviation of one before entering the network.

Why A is correct: Scaling inputs to a small, bounded range keeps weight gradients in a workable magnitude, preventing exploding gradients and accelerating convergence - the canonical motivation for input normalisation.

Why B is wrong: Dividing pixel values by a constant does not alter the spatial dimensions of the tensor; downsampling requires pooling or strided convolutions, not scalar division.

Why C is wrong: Casting to float is a necessary step, but the division itself is not required for the cast; the primary motivation is gradient stability from bounded inputs, not merely the numeric type.

Why D is wrong: 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.

An audio classification model must process raw waveform recordings. A data engineer proposes converting each recording to a mel-frequency spectrogram before passing it to a 2-D convolutional network. What does this conversion accomplish in terms of the tensor representation fed to the network?

• AIt encodes the waveform as a 1-D sequence of amplitude samples that the convolutional layers treat as a single-channel time-series.
• BIt transforms the time-domain signal into a 2-D frequency-versus-time tensor whose values represent energy in perceptually scaled frequency bands, suitable for image-like convolution.check_circle Correct
• CIt compresses the audio into a fixed-length dense vector using a pre-trained acoustic encoder, removing the need for further spatial feature extraction.
• DIt converts the waveform into a sequence of discrete tokens by quantising amplitude values, matching the input format expected by transformer-based language models.

Why A is wrong: A raw waveform is already a 1-D amplitude sequence; converting to a spectrogram produces a 2-D frequency-time representation, not a 1-D sequence.

Why B is correct: A mel spectrogram applies the Short-Time Fourier Transform followed by mel filterbank projection, yielding a 2-D matrix of log-energy values indexed by mel-scaled frequency and time. This format exposes both temporal and spectral structure and can be processed by 2-D convolutional filters the same way an image is.

Why C is wrong: A spectrogram is computed by a deterministic signal-processing transform, not by a learned encoder. It produces a variable-length 2-D array rather than a fixed-length embedding vector.

Why D is wrong: Amplitude quantisation into discrete tokens is used in codebook-based audio codecs such as EnCodec, not in mel-spectrogram extraction. Mel spectrograms are continuous-valued, not token sequences.

When preparing text for a transformer-based language model, a developer must choose between passing raw one-hot character vectors and using a learned dense embedding layer. Which statement best describes the practical advantage of learned dense embeddings over raw one-hot vectors for this purpose?

• ADense embeddings guarantee that semantically related tokens are always placed at equal distances from each other in the embedding space before any training begins.
• BDense embeddings eliminate the need for tokenisation because they operate directly on raw character byte values without any vocabulary mapping.
• CDense embeddings reduce the input dimensionality from the vocabulary size to a much smaller continuous space, and the learned geometry encodes semantic relationships between tokens.check_circle Correct
• DDense embeddings convert token indices into fixed sinusoidal signals that provide the model with absolute position information across the input sequence.

Why A is wrong: Embeddings are randomly initialised before training; semantic relationships emerge through gradient updates on task data, they are not guaranteed at initialisation.

Why B is wrong: An embedding layer still requires a tokenisation step to map input text to integer indices; it replaces the one-hot lookup but does not remove the tokenisation or vocabulary mapping requirement.

Why C is correct: One-hot vectors have dimensionality equal to the vocabulary size and are completely orthogonal, encoding no similarity between tokens. A learned embedding projects each token into a low-dimensional continuous space where the training objective shapes the geometry so that related tokens cluster together, enabling the network to generalise across synonyms and related concepts.

Why D is wrong: Sinusoidal positional encodings are a separate addition to token embeddings, not a property of the token embedding layer itself. The two components serve distinct purposes and are added together rather than merged into one.