Intermediate everything in AI & ML, page 8
Feature detector vs feature descriptor.
A detector finds where interesting points are, a descriptor encodes the local appearance around each so points can be matched.
How is an HDR radiance map constructed from exposures?
Recover the inverse camera response function from corresponding pixels, linearize each exposure to radiance, then merge with confidence weights into a floating-point radiance map.
CPU vs GPU vs Edge TPU for inference.
CPU is flexible but slow, GPU offers massive parallelism at high power, Edge TPU gives efficient low-power int8 inference but is constrained; choose by latency, power, cost, and model fit.
How do you train and evaluate on imbalanced defect data?
Resampling, class weighting, focal loss, and anomaly framing for training; evaluate with precision, recall, PR-AUC, and F-beta, not accuracy.
How do you speed up a slow detection model?
Quantization, pruning, distillation, lighter backbones, and resolution or batching tweaks, each trading some accuracy or effort for speed.
Compare Gray World and White Patch white balance.
Gray World assumes average scene color is gray, White Patch assumes the brightest pixel is white, both fail on dominant colors or clipping; learning predicts illuminant from data.
How does smartphone Portrait Mode produce bokeh?
Estimate per-pixel depth via dual-pixel or stereo or learning, segment the subject, then apply depth-dependent blur.
Prototypical Networks for few-shot classification
An encoder embeds support examples, each class prototype is the mean embedding of its support examples, and a query is classified by nearest prototype using a distance like Euclidean via softmax.
Contrastive self-supervised learning with SimCLR
Two augmentations of one image form a positive pair, other images in the batch are negatives; an encoder plus projection head and the NT-Xent loss pull positives together and push negatives apart.
Camera intrinsics, extrinsics, and the essential matrix
Intrinsics map camera coords to pixels, extrinsics are camera pose in the world; the essential matrix relates normalized points across two views, encoding relative rotation and translation up to scale…
Unpaired image translation with CycleGAN
CycleGAN uses two generators and two discriminators with a cycle-consistency loss that forces translating to the other domain and back to reconstruct the input, removing the need for paired data.
How text prompts guide Stable Diffusion
A frozen text encoder turns the prompt into token embeddings, which feed the U-Net via cross-attention at each denoising step so the prompt steers generation; classifier-free guidance amplifies the…
Why U-Net skip connections matter for denoising
Skips carry high-resolution spatial detail from encoder to decoder, preserving fine structure lost in downsampling and easing gradient flow, which lets the model restore detail while removing…
Evaluating generative models with FID versus IS
FID compares Inception feature distributions of real and fake images via Frechet distance between two Gaussians; it uses real data as reference and detects diversity issues, unlike IS which uses no real…
Mode collapse in GAN training
Generator produces few outputs ignoring data diversity, caused by chasing whatever fools the current discriminator; mitigate with minibatch discrimination, unrolled GANs, or Wasserstein loss.
Why ViTs need positional embeddings
Self-attention is permutation invariant so patch order is lost; positional embeddings restore spatial location. CNNs encode position implicitly via the fixed convolution grid.
Cross-attention for visual question answering
In cross-attention queries come from one modality and keys/values from the other, e.g. text queries attend over image features so the question selects relevant regions.
How Swin Transformer achieves linear attention
Swin computes attention within local non-overlapping windows of fixed size, making cost linear in patches, then shifts windows between layers so information crosses boundaries.
Inductive biases of ViT versus CNN
CNNs bake in locality and translation equivariance; a plain ViT has almost none beyond patch structure, so it must learn spatial relations from data, needing large datasets or strong pretraining.
Re-identification in multi-object tracking
Re-ID matches an object to its prior id using appearance embeddings, not just position; store track features and match re-entering detections by embedding similarity.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles